Import ggml to SC

ggml/CMakeLists.txt

@@ -0,0 +1,185 @@
+cmake_minimum_required (VERSION 3.3)
+project(ggml VERSION 0.1.0)
+
+set(CMAKE_EXPORT_COMPILE_COMMANDS "on")
+set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin)
+set(CMAKE_INSTALL_RPATH "${CMAKE_INSTALL_PREFIX}/lib")
+
+if(CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR)
+    set(GGML_STANDALONE ON)
+    include(cmake/GitVars.cmake)
+    include(cmake/BuildTypes.cmake)
+else()
+    set(GGML_STANDALONE OFF)
+endif()
+
+# options
+
+option(GGML_ALL_WARNINGS            "ggml: enable all compiler warnings"                   ON)
+option(GGML_ALL_WARNINGS_3RD_PARTY  "ggml: enable all compiler warnings in 3rd party libs" OFF)
+
+option(GGML_SANITIZE_THREAD         "ggml: enable thread sanitizer"    OFF)
+option(GGML_SANITIZE_ADDRESS        "ggml: enable address sanitizer"   OFF)
+option(GGML_SANITIZE_UNDEFINED      "ggml: enable undefined sanitizer" OFF)
+
+option(GGML_BUILD_TESTS             "ggml: build tests"    ${GGML_STANDALONE})
+option(GGML_BUILD_EXAMPLES          "ggml: build examples" ${GGML_STANDALONE})
+
+option(GGML_TEST_COVERAGE           "ggml: enable test coverage" OFF)
+
+option(GGML_PERF                    "ggml: enable perf timings"          OFF)
+option(GGML_NO_ACCELERATE           "ggml: disable Accelerate framework" OFF)
+option(GGML_OPENBLAS                "ggml: use OpenBLAS"                 OFF)
+option(GGML_CLBLAST                 "ggml: use clBLAST"                  OFF)
+option(GGML_CUBLAS                  "ggml: use cuBLAS"                   OFF)
+option(GGML_METAL                   "ggml: use Metal"                    OFF)
+
+# sanitizers
+
+if (GGML_SANITIZE_THREAD)
+    set(CMAKE_C_FLAGS   "${CMAKE_C_FLAGS}   -fsanitize=thread")
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsanitize=thread")
+endif()
+
+if (GGML_SANITIZE_ADDRESS)
+    set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS}     -fsanitize=address -fno-omit-frame-pointer")
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsanitize=address -fno-omit-frame-pointer")
+endif()
+
+if (GGML_SANITIZE_UNDEFINED)
+    set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS}     -fsanitize=undefined")
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsanitize=undefined")
+endif()
+
+# instruction set specific
+option(GGML_AVX                     "ggml: enable AVX"                                     ON)
+option(GGML_AVX2                    "ggml: enable AVX2"                                    ON)
+option(GGML_AVX512                  "ggml: enable AVX512"                                  OFF)
+option(GGML_AVX512_VBMI             "ggml: enable AVX512-VBMI"                             OFF)
+option(GGML_AVX512_VNNI             "ggml: enable AVX512-VNNI"                             OFF)
+option(GGML_FMA                     "ggml: enable FMA"                                     ON)
+# in MSVC F16C is implied with AVX2/AVX512
+if (NOT MSVC)
+    option(GGML_F16C                "ggml: enable F16C"                                    ON)
+endif()
+
+#set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -ffast-math")
+#set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -march=native")
+#set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -mcpu=native")
+
+# warning flags
+
+if (GGML_ALL_WARNINGS)
+    if (NOT MSVC)
+        set(c_flags   -Wall -Wpedantic -Wformat=2 -Wno-unused -Wstrict-prototypes)
+        set(cxx_flags -Wall -Wpedantic -Wformat=2)
+    else()
+        # todo : windows
+    endif()
+
+    add_compile_options(
+        "$<$<COMPILE_LANGUAGE:C>:${c_flags}>"
+        "$<$<COMPILE_LANGUAGE:CXX>:${cxx_flags}>"
+    )
+endif()
+
+if (NOT MSVC)
+    add_compile_options(
+        "$<$<COMPILE_LANGUAGE:C>:-Werror=vla>"
+        "$<$<COMPILE_LANGUAGE:CXX>:-Werror=vla>"
+        "$<$<COMPILE_LANGUAGE:CUDA>:-Xcompiler;-Werror=vla>"
+    )
+endif()
+
+#
+# POSIX conformance
+#
+
+# clock_gettime came in POSIX.1b (1993)
+# CLOCK_MONOTONIC came in POSIX.1-2001 / SUSv3 as optional
+# posix_memalign came in POSIX.1-2001 / SUSv3
+# M_PI is an XSI extension since POSIX.1-2001 / SUSv3, came in XPG1 (1985)
+add_compile_definitions(_XOPEN_SOURCE=600)
+
+# Somehow in OpenBSD whenever POSIX conformance is specified
+# some string functions rely on locale_t availability,
+# which was introduced in POSIX.1-2008, forcing us to go higher
+if (CMAKE_SYSTEM_NAME MATCHES "OpenBSD")
+    remove_definitions(-D_XOPEN_SOURCE=600)
+    add_compile_definitions(_XOPEN_SOURCE=700)
+endif()
+
+# Data types, macros and functions related to controlling CPU affinity
+# are available on Linux through GNU extensions in libc
+if (CMAKE_SYSTEM_NAME MATCHES "Linux")
+    add_compile_definitions(_GNU_SOURCE)
+endif()
+
+# RLIMIT_MEMLOCK came in BSD, is not specified in POSIX.1,
+# and on macOS its availability depends on enabling Darwin extensions
+# similarly on DragonFly, enabling BSD extensions is necessary
+if (CMAKE_SYSTEM_NAME MATCHES "Darwin")
+    add_compile_definitions(_DARWIN_C_SOURCE)
+endif()
+if (CMAKE_SYSTEM_NAME MATCHES "DragonFly")
+    add_compile_definitions(_DARWIN_C_SOURCE)
+endif()
+
+# alloca is a non-standard interface that is not visible on BSDs when
+# POSIX conformance is specified, but not all of them provide a clean way
+# to enable it in such cases
+if (CMAKE_SYSTEM_NAME MATCHES "FreeBSD")
+    add_compile_definitions(__BSD_VISIBLE)
+endif()
+if (CMAKE_SYSTEM_NAME MATCHES "NetBSD")
+    add_compile_definitions(_NETBSD_SOURCE)
+endif()
+if (CMAKE_SYSTEM_NAME MATCHES "OpenBSD")
+    add_compile_definitions(_BSD_SOURCE)
+endif()
+
+if (WHISPER_PERF)
+    set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DGGML_PERF)
+endif()
+
+# dependencies
+
+set(CMAKE_C_STANDARD   11)
+set(CMAKE_CXX_STANDARD 11)
+
+find_package(Threads REQUIRED)
+
+# main
+
+if (NOT CMAKE_BUILD_TYPE AND NOT CMAKE_CONFIGURATION_TYPES)
+    set(CMAKE_BUILD_TYPE Release CACHE STRING "Build type" FORCE)
+    set_property(CACHE CMAKE_BUILD_TYPE PROPERTY STRINGS "Debug" "Release" "RelWithDebInfo")
+endif ()
+
+if (GGML_BUILD_TESTS)
+    if (GGML_TEST_COVERAGE)
+        if (CMAKE_C_COMPILER_ID MATCHES "Clang")
+            set(CMAKE_C_FLAGS   "${CMAKE_C_FLAGS}   -fprofile-instr-generate -fcoverage-mapping")
+            set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fprofile-instr-generate -fcoverage-mapping")
+        else()
+            message(WARNING "Test coverage is only supported for Clang")
+        endif()
+    endif()
+endif()
+
+add_subdirectory(src)
+
+if (GGML_BUILD_TESTS)
+    enable_testing()
+    add_subdirectory(tests)
+endif ()
+
+if (GGML_BUILD_EXAMPLES)
+    add_subdirectory(examples)
+endif ()
+
+configure_file(${CMAKE_CURRENT_SOURCE_DIR}/ggml.pc.in
+               ${CMAKE_CURRENT_BINARY_DIR}/ggml.pc
+               @ONLY)
+install(FILES ${CMAKE_CURRENT_BINARY_DIR}/ggml.pc
+        DESTINATION share/pkgconfig)

ggml/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2022 Georgi Gerganov
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

ggml/README.md

@@ -0,0 +1,126 @@
+# ggml
+
+[Roadmap](https://github.com/users/ggerganov/projects/7) / [Manifesto](https://github.com/ggerganov/llama.cpp/discussions/205)
+
+Tensor library for machine learning
+
+***Note that this project is under active development. \
+Some of the development is currently happening in the [llama.cpp](https://github.com/ggerganov/llama.cpp) and [whisper.cpp](https://github.com/ggerganov/whisper.cpp) repos***
+
+## Features
+
+- Written in C
+- 16-bit float support
+- Integer quantization support (4-bit, 5-bit, 8-bit, etc.)
+- Automatic differentiation
+- ADAM and L-BFGS optimizers
+- Optimized for Apple Silicon
+- On x86 architectures utilizes AVX / AVX2 intrinsics
+- On ppc64 architectures utilizes VSX intrinsics
+- No third-party dependencies
+- Zero memory allocations during runtime
+
+## Updates
+
+- [X] Example of GPT-2 inference [examples/gpt-2](https://github.com/ggerganov/ggml/tree/master/examples/gpt-2)
+- [X] Example of GPT-J inference [examples/gpt-j](https://github.com/ggerganov/ggml/tree/master/examples/gpt-j)
+- [X] Example of Whisper inference [examples/whisper](https://github.com/ggerganov/ggml/tree/master/examples/whisper)
+- [X] Support 4-bit integer quantization https://github.com/ggerganov/ggml/pull/27
+- [X] Example of Cerebras-GPT inference [examples/gpt-2](https://github.com/ggerganov/ggml/tree/master/examples/gpt-2)
+- [ ] Example of FLAN-T5 inference https://github.com/ggerganov/ggml/pull/12
+- [X] Example of LLaMA inference [ggerganov/llama.cpp](https://github.com/ggerganov/llama.cpp)
+- [X] Example of LLaMA training [ggerganov/llama.cpp/examples/baby-llama](https://github.com/ggerganov/llama.cpp/tree/master/examples/baby-llama)
+- [X] Example of Falcon inference [cmp-nct/ggllm.cpp](https://github.com/cmp-nct/ggllm.cpp)
+- [X] Example of BLOOM inference [NouamaneTazi/bloomz.cpp](https://github.com/NouamaneTazi/bloomz.cpp)
+- [X] Example of RWKV inference [saharNooby/rwkv.cpp](https://github.com/saharNooby/rwkv.cpp)
+- [X] Example of SAM inference [examples/sam](https://github.com/ggerganov/ggml/tree/master/examples/sam)
+- [X] Idea for GPU support: https://github.com/ggerganov/llama.cpp/discussions/915
+- [X] Example of StableLM (GPT-NeoX) inference [examples/gpt-neox](https://github.com/ggerganov/ggml/tree/master/examples/gpt-neox)
+- [X] Example of BERT inference [skeskinen/bert.cpp](https://github.com/skeskinen/bert.cpp)
+- [X] Example of 💫 StarCoder inference [examples/starcoder](https://github.com/ggerganov/ggml/tree/master/examples/starcoder)
+- [X] Example of MPT inference [examples/mpt](https://github.com/ggerganov/ggml/tree/master/examples/mpt)
+- [X] Example of Replit inference [examples/replit](https://github.com/ggerganov/ggml/tree/master/examples/replit)
+- [X] Example of BioGPT inference [PABannier/biogpt.cpp](https://github.com/PABannier/biogpt.cpp)
+- [X] Example of Encodec inference [PABannier/encodec.cpp](https://github.com/PABannier/encodec.cpp)
+- [X] Example of CLIP inference [monatis/clip.cpp](https://github.com/monatis/clip.cpp)
+- [X] Example of MiniGPT4 inference [Maknee/minigpt4.cpp](https://github.com/Maknee/minigpt4.cpp)
+- [X] Example of ChatGLM inference [li-plus/chatglm.cpp](https://github.com/li-plus/chatglm.cpp)
+- [X] Example of Stable Diffusion inference [leejet/stable-diffusion.cpp](https://github.com/leejet/stable-diffusion.cpp)
+
+## Whisper inference (example)
+
+With ggml you can efficiently run [Whisper](examples/whisper) inference on the CPU.
+
+Memory requirements:
+
+| Model  | Disk   | Mem     |
+| ---    | ---    | ---     |
+| tiny   |  75 MB | ~280 MB |
+| base   | 142 MB | ~430 MB |
+| small  | 466 MB | ~1.0 GB |
+| medium | 1.5 GB | ~2.6 GB |
+| large  | 2.9 GB | ~4.7 GB |
+
+## GPT inference (example)
+
+With ggml you can efficiently run [GPT-2](examples/gpt-2) and [GPT-J](examples/gpt-j) inference on the CPU.
+
+Here is how to run the example programs:
+
+```bash
+# Build ggml + examples
+git clone https://github.com/ggerganov/ggml
+cd ggml
+mkdir build && cd build
+cmake ..
+make -j4 gpt-2 gpt-j
+
+# Run the GPT-2 small 117M model
+../examples/gpt-2/download-ggml-model.sh 117M
+./bin/gpt-2 -m models/gpt-2-117M/ggml-model.bin -p "This is an example"
+
+# Run the GPT-J 6B model (requires 12GB disk space and 16GB CPU RAM)
+../examples/gpt-j/download-ggml-model.sh 6B
+./bin/gpt-j -m models/gpt-j-6B/ggml-model.bin -p "This is an example"
+
+# Install Python dependencies
+python3 -m pip install -r ../requirements.txt
+
+# Run the Cerebras-GPT 111M model
+# Download from: https://huggingface.co/cerebras
+python3 ../examples/gpt-2/convert-cerebras-to-ggml.py /path/to/Cerebras-GPT-111M/
+./bin/gpt-2 -m /path/to/Cerebras-GPT-111M/ggml-model-f16.bin -p "This is an example"
+```
+
+The inference speeds that I get for the different models on my 32GB MacBook M1 Pro are as follows:
+
+| Model | Size  | Time / Token |
+| ---   | ---   | ---    |
+| GPT-2 |  117M |   5 ms |
+| GPT-2 |  345M |  12 ms |
+| GPT-2 |  774M |  23 ms |
+| GPT-2 | 1558M |  42 ms |
+| ---   | ---   | ---    |
+| GPT-J |    6B | 125 ms |
+
+For more information, checkout the corresponding programs in the [examples](examples) folder.
+
+## Using cuBLAS
+
+```bash
+# fix the path to point to your CUDA compiler
+cmake -DGGML_CUBLAS=ON -DCMAKE_CUDA_COMPILER=/usr/local/cuda-12.1/bin/nvcc ..
+```
+
+## Using clBLAST
+
+```bash
+cmake -DGGML_CLBLAST=ON ..
+```
+
+## Resources
+
+- [GGML - Large Language Models for Everyone](https://github.com/rustformers/llm/blob/main/crates/ggml/README.md): a description of the GGML format provided by the maintainers of the `llm` Rust crate, which provides Rust bindings for GGML
+- [marella/ctransformers](https://github.com/marella/ctransformers): Python bindings for GGML models.
+- [go-skynet/go-ggml-transformers.cpp](https://github.com/go-skynet/go-ggml-transformers.cpp): Golang bindings for GGML models
+- [smspillaz/ggml-gobject](https://github.com/smspillaz/ggml-gobject): GObject-introspectable wrapper for use of GGML on the GNOME platform.

ggml/build.zig

@@ -0,0 +1,158 @@
+const std = @import("std");
+const builtin = @import("builtin");
+
+// Zig Version: 0.11.0
+// Zig Build Command: zig build
+// Zig Run Command: zig build -h
+//     zig build run_dolly-v2
+//     zig build run_gpt-2
+//     zig build run_gpt-j
+//     zig build run_gpt-neox
+//     zig build run_mnist
+//     zig build run_mpt
+//     zig build run_replit
+//     zig build run_starcoder
+//     zig build run_test-grad0
+//     zig build run_test-mul-mat0
+//     zig build run_test-mul-mat2
+//     zig build run_test-opt
+//     zig build run_test-vec1
+//     zig build run_test0
+//     zig build run_test1
+//     zig build run_test2
+//     zig build run_test3
+//     zig build run_zig_test0
+//     zig build run_zig_test1
+//     zig build run_zig_test2
+//     zig build run_zig_test3
+pub fn build(b: *std.build.Builder) void {
+    const target = b.standardTargetOptions(.{});
+    const optimize = b.standardOptimizeOption(.{});
+    const lib = b.addStaticLibrary(.{
+        .name = "ggml",
+        .target = target,
+        .optimize = optimize,
+    });
+    lib.addIncludePath(.{ .path = "./include" });
+    lib.addIncludePath(.{ .path = "./include/ggml" });
+    lib.addCSourceFiles(&.{
+        "src/ggml.c",
+    }, &.{"-std=c11"});
+    lib.linkLibC();
+    lib.linkLibCpp();
+    b.installArtifact(lib);
+
+    // examples
+    const examples = .{
+        "dolly-v2",
+        "gpt-2",
+        "gpt-j",
+        "gpt-neox",
+        "mnist",
+        "mpt",
+        "replit",
+        "starcoder",
+        // "whisper",
+    };
+    inline for (examples) |name| {
+        const exe = b.addExecutable(.{
+            .name = name,
+            .target = target,
+            .optimize = optimize,
+        });
+        exe.addIncludePath(.{ .path = "./include" });
+        exe.addIncludePath(.{ .path = "./include/ggml" });
+        exe.addIncludePath(.{ .path = "./examples" });
+        // exe.addIncludePath("./examples/whisper");
+        exe.addCSourceFiles(&.{
+            std.fmt.comptimePrint("examples/{s}/main.cpp", .{name}),
+            "examples/common.cpp",
+            "examples/common-ggml.cpp",
+            // "examples/whisper/whisper.cpp",
+        }, &.{"-std=c++11"});
+        exe.linkLibrary(lib);
+        b.installArtifact(exe);
+        const run_cmd = b.addRunArtifact(exe);
+        run_cmd.step.dependOn(b.getInstallStep());
+        if (b.args) |args| run_cmd.addArgs(args);
+        const run_step = b.step("run_" ++ name, "Run examples");
+        run_step.dependOn(&run_cmd.step);
+    }
+
+    // tests
+    const tests = if (builtin.target.cpu.arch == .x86_64) .{
+        // "test-blas0",
+        // "test-grad0",
+        "test-mul-mat0",
+        // "test-mul-mat1",
+        "test-mul-mat2",
+        // "test-opt",
+        // "test-svd0",
+        // "test-vec0",
+        "test-vec1",
+        // "test-vec2",
+        "test0",
+        "test1",
+        "test2",
+        "test3",
+    } else .{
+        // "test-blas0",
+        // "test-grad0",
+        "test-mul-mat0",
+        // "test-mul-mat1",
+        "test-mul-mat2",
+        // "test-opt",
+        // "test-svd0",
+        // "test-vec0",
+        // "test-vec1",
+        // "test-vec2",
+        "test0",
+        "test1",
+        "test2",
+        "test3",
+    };
+    inline for (tests) |name| {
+        const exe = b.addExecutable(.{
+            .name = name,
+            .target = target,
+            .optimize = optimize,
+        });
+        exe.addIncludePath(.{ .path = "./include" });
+        exe.addIncludePath(.{ .path = "./include/ggml" });
+        exe.addCSourceFiles(&.{
+            std.fmt.comptimePrint("tests/{s}.c", .{name}),
+        }, &.{"-std=c11"});
+        exe.linkLibrary(lib);
+        b.installArtifact(exe);
+        const run_cmd = b.addRunArtifact(exe);
+        run_cmd.step.dependOn(b.getInstallStep());
+        if (b.args) |args| run_cmd.addArgs(args);
+        const run_step = b.step("run_" ++ name, "Run tests");
+        run_step.dependOn(&run_cmd.step);
+    }
+
+    // zig_tests
+    const zig_tests = .{
+        "test0",
+        "test1",
+        "test2",
+        "test3",
+    };
+    inline for (zig_tests) |name| {
+        const exe = b.addExecutable(.{
+            .name = name,
+            .root_source_file = .{ .path = std.fmt.comptimePrint("tests/{s}.zig", .{name}) },
+            .target = target,
+            .optimize = optimize,
+        });
+        exe.addIncludePath(.{ .path = "./include" });
+        exe.addIncludePath(.{ .path = "./include/ggml" });
+        exe.linkLibrary(lib);
+        b.installArtifact(exe);
+        const run_cmd = b.addRunArtifact(exe);
+        run_cmd.step.dependOn(b.getInstallStep());
+        if (b.args) |args| run_cmd.addArgs(args);
+        const run_step = b.step("run_zig_" ++ name, "Run zig_tests");
+        run_step.dependOn(&run_cmd.step);
+    }
+}

ggml/ci/run.sh

@@ -0,0 +1,260 @@
+#/bin/bash
+#
+# sample usage:
+#
+# mkdir tmp
+#
+# # CPU-only build
+# bash ./ci/run.sh ./tmp/results ./tmp/mnt
+#
+# # with CUDA support
+# GG_BUILD_CUDA=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
+#
+
+if [ -z "$2" ]; then
+    echo "usage: $0 <output-dir> <mnt-dir>"
+    exit 1
+fi
+
+mkdir -p "$1"
+mkdir -p "$2"
+
+OUT=$(realpath "$1")
+MNT=$(realpath "$2")
+
+rm -v $OUT/*.log
+rm -v $OUT/*.exit
+rm -v $OUT/*.md
+
+sd=`dirname $0`
+cd $sd/../
+SRC=`pwd`
+
+## helpers
+
+# download a file if it does not exist or if it is outdated
+function gg_wget {
+    local out=$1
+    local url=$2
+
+    local cwd=`pwd`
+
+    mkdir -p $out
+    cd $out
+
+    # should not re-download if file is the same
+    wget -nv -N $url
+
+    cd $cwd
+}
+
+function gg_printf {
+    printf -- "$@" >> $OUT/README.md
+}
+
+function gg_run {
+    ci=$1
+
+    set -o pipefail
+    set -x
+
+    gg_run_$ci | tee $OUT/$ci.log
+    cur=$?
+    echo "$cur" > $OUT/$ci.exit
+
+    set +x
+    set +o pipefail
+
+    gg_sum_$ci
+
+    ret=$((ret | cur))
+}
+
+## ci
+
+# ctest_debug
+
+function gg_run_ctest_debug {
+    cd ${SRC}
+
+    rm -rf build-ci-debug && mkdir build-ci-debug && cd build-ci-debug
+
+    set -e
+
+    (time cmake -DCMAKE_BUILD_TYPE=Debug ..     ) 2>&1 | tee -a $OUT/${ci}-cmake.log
+    (time make -j                               ) 2>&1 | tee -a $OUT/${ci}-make.log
+
+    (time ctest --output-on-failure -E test-opt ) 2>&1 | tee -a $OUT/${ci}-ctest.log
+
+    set +e
+}
+
+function gg_sum_ctest_debug {
+    gg_printf '### %s\n\n' "${ci}"
+
+    gg_printf 'Runs ctest in debug mode\n'
+    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
+    gg_printf '```\n'
+    gg_printf '%s\n' "$(cat $OUT/${ci}-ctest.log)"
+    gg_printf '```\n'
+    gg_printf '\n'
+}
+
+# ctest_release
+
+function gg_run_ctest_release {
+    cd ${SRC}
+
+    rm -rf build-ci-release && mkdir build-ci-release && cd build-ci-release
+
+    set -e
+
+    (time cmake -DCMAKE_BUILD_TYPE=Release ..   ) 2>&1 | tee -a $OUT/${ci}-cmake.log
+    (time make -j                               ) 2>&1 | tee -a $OUT/${ci}-make.log
+
+    if [ -z $GG_BUILD_LOW_PERF ]; then
+        (time ctest --output-on-failure ) 2>&1 | tee -a $OUT/${ci}-ctest.log
+    else
+        (time ctest --output-on-failure -E test-opt ) 2>&1 | tee -a $OUT/${ci}-ctest.log
+    fi
+
+    set +e
+}
+
+function gg_sum_ctest_release {
+    gg_printf '### %s\n\n' "${ci}"
+
+    gg_printf 'Runs ctest in release mode\n'
+    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
+    gg_printf '```\n'
+    gg_printf '%s\n' "$(cat $OUT/${ci}-ctest.log)"
+    gg_printf '```\n'
+}
+
+# gpt_2
+
+function gg_run_gpt_2 {
+    cd ${SRC}
+
+    gg_wget models-mnt/gpt-2 https://huggingface.co/ggerganov/ggml/resolve/main/ggml-model-gpt-2-117M.bin
+
+    cd build-ci-release
+
+    set -e
+
+    model="../models-mnt/gpt-2/ggml-model-gpt-2-117M.bin"
+    prompts="../examples/prompts/gpt-2.txt"
+
+    (time ./bin/gpt-2 --model ${model} -s 1234 -n 64 -tt ${prompts}                       ) 2>&1 | tee -a $OUT/${ci}-tg.log
+    (time ./bin/gpt-2 --model ${model} -s 1234 -n 64 -p "I believe the meaning of life is") 2>&1 | tee -a $OUT/${ci}-tg.log
+
+    set +e
+}
+
+function gg_sum_gpt_2 {
+    gg_printf '### %s\n\n' "${ci}"
+
+    gg_printf 'Runs short GPT-2 text generation\n'
+    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
+    gg_printf '```\n'
+    gg_printf '%s\n' "$(cat $OUT/${ci}-tg.log)"
+    gg_printf '```\n'
+}
+
+# mpt
+
+function gg_run_mpt {
+    cd ${SRC}
+
+    gg_wget models-mnt/mpt/7B/ https://huggingface.co/mosaicml/mpt-7b/raw/main/config.json
+    gg_wget models-mnt/mpt/7B/ https://huggingface.co/mosaicml/mpt-7b/raw/main/tokenizer.json
+    gg_wget models-mnt/mpt/7B/ https://huggingface.co/mosaicml/mpt-7b/raw/main/tokenizer_config.json
+    gg_wget models-mnt/mpt/7B/ https://huggingface.co/mosaicml/mpt-7b/raw/main/pytorch_model.bin.index.json
+    gg_wget models-mnt/mpt/7B/ https://huggingface.co/mosaicml/mpt-7b/raw/main/configuration_mpt.py
+    gg_wget models-mnt/mpt/7B/ https://huggingface.co/mosaicml/mpt-7b/resolve/main/pytorch_model-00001-of-00002.bin
+    gg_wget models-mnt/mpt/7B/ https://huggingface.co/mosaicml/mpt-7b/resolve/main/pytorch_model-00002-of-00002.bin
+
+    cd build-ci-release
+
+    set -e
+
+    path_models="../models-mnt/mpt/7B"
+    model_f16="${path_models}/ggml-model-f16.bin"
+    model_q4_0="${path_models}/ggml-model-q4_0.bin"
+
+    python3 ../examples/mpt/convert-h5-to-ggml.py ${path_models} 1
+    ./bin/mpt-quantize ${model_f16} ${model_q4_0} q4_0
+
+    (time ./bin/mpt --model ${model_f16}  -s 1234 -n 64 -p "I believe the meaning of life is") 2>&1 | tee -a $OUT/${ci}-tg.log
+    (time ./bin/mpt --model ${model_q4_0} -s 1234 -n 64 -p "I believe the meaning of life is") 2>&1 | tee -a $OUT/${ci}-tg.log
+
+    set +e
+}
+
+function gg_sum_mpt {
+    gg_printf '### %s\n\n' "${ci}"
+
+    gg_printf 'Runs short MPT text generation\n'
+    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
+    gg_printf '```\n'
+    gg_printf '%s\n' "$(cat $OUT/${ci}-tg.log)"
+    gg_printf '```\n'
+}
+
+# mnist
+
+function gg_run_mnist {
+    cd ${SRC}
+
+    cd build-ci-release
+
+    set -e
+
+    mkdir -p models/mnist
+    python3 ../examples/mnist/convert-h5-to-ggml.py ../examples/mnist/models/mnist/mnist_model.state_dict
+
+    model_f32="./models/mnist/ggml-model-f32.bin"
+    samples="../examples/mnist/models/mnist/t10k-images.idx3-ubyte"
+
+    # first command runs and exports "mnist.ggml", the second command runs the exported model
+
+    (time ./bin/mnist     ${model_f32} ${samples} ) 2>&1 | tee -a $OUT/${ci}-mnist.log
+    (time ./bin/mnist-cpu ./mnist.ggml ${samples} ) 2>&1 | tee -a $OUT/${ci}-mnist.log
+
+    set +e
+}
+
+function gg_sum_mnist {
+    gg_printf '### %s\n\n' "${ci}"
+
+    gg_printf 'MNIST\n'
+    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
+    gg_printf '```\n'
+    gg_printf '%s\n' "$(cat $OUT/${ci}-mnist.log)"
+    gg_printf '```\n'
+}
+
+## main
+
+if [ -z $GG_BUILD_LOW_PERF ]; then
+    rm -rf ${SRC}/models-mnt
+
+    mnt_models=${MNT}/models
+    mkdir -p ${mnt_models}
+    ln -sfn ${mnt_models} ${SRC}/models-mnt
+fi
+
+python3 -m pip install -r ${SRC}/requirements.txt
+
+ret=0
+
+test $ret -eq 0 && gg_run ctest_debug
+test $ret -eq 0 && gg_run ctest_release
+test $ret -eq 0 && gg_run gpt_2
+test $ret -eq 0 && gg_run mnist
+
+if [ -z $GG_BUILD_LOW_PERF ]; then
+    test $ret -eq 0 && gg_run mpt
+fi
+
+exit $ret

ggml/cmake/BuildTypes.cmake

@@ -0,0 +1,54 @@
+# Add new build types
+
+# ReleaseGG - Release with enabled asserts
+
+SET(CMAKE_CXX_FLAGS_RELEASEGG
+    "-O3"
+    CACHE STRING "Flags used by the c++ compiler during release builds with enabled asserts."
+    FORCE )
+SET(CMAKE_C_FLAGS_RELEASEGG
+    "-O3"
+    CACHE STRING "Flags used by the compiler during release builds with enabled asserts."
+    FORCE )
+SET(CMAKE_EXE_LINKER_FLAGS_RELEASEGG
+    ""
+    CACHE STRING "Flags used for linking binaries during release builds with enabled asserts."
+    FORCE )
+SET(CMAKE_SHARED_LINKER_FLAGS_RELEASEGG
+    ""
+    CACHE STRING "Flags used by the shared libraries linker during release builds with enabled asserts."
+    FORCE )
+MARK_AS_ADVANCED(
+    CMAKE_CXX_FLAGS_RELEASEGG
+    CMAKE_C_FLAGS_RELEASEGG
+    CMAKE_EXE_LINKER_FLAGS_RELEASEGG
+    CMAKE_SHARED_LINKER_FLAGS_RELEASEGG )
+
+# RelWithDebInfoGG - RelWithDebInfo with enabled asserts
+
+SET(CMAKE_CXX_FLAGS_RELWITHDEBINFOGG
+    "-O2 -g"
+    CACHE STRING "Flags used by the c++ compiler during release builds with debug symbols and enabled asserts."
+    FORCE )
+SET(CMAKE_C_FLAGS_RELWITHDEBINFOGG
+    "-O2 -g"
+    CACHE STRING "Flags used by the compiler during release builds with debug symbols and enabled asserts."
+    FORCE )
+SET(CMAKE_EXE_LINKER_FLAGS_RELWITHDEBINFOGG
+    ""
+    CACHE STRING "Flags used for linking binaries during release builds with debug symbols and enabled asserts."
+    FORCE )
+SET(CMAKE_SHARED_LINKER_FLAGS_RELWITHDEBINFOGG
+    ""
+    CACHE STRING "Flags used by the shared libraries linker during release builds with debug symbols and enabled asserts."
+    FORCE )
+MARK_AS_ADVANCED(
+    CMAKE_CXX_FLAGS_RELWITHDEBINFOGG
+    CMAKE_C_FLAGS_RELWITHDEBINFOGG
+    CMAKE_EXE_LINKER_FLAGS_RELWITHDEBINFOGG
+    CMAKE_SHARED_LINKER_FLAGS_RELWITHDEBINFOGG )
+
+if (NOT XCODE AND NOT MSVC AND NOT CMAKE_BUILD_TYPE)
+    set(CMAKE_BUILD_TYPE Release CACHE STRING "Build type" FORCE)
+    set_property(CACHE CMAKE_BUILD_TYPE PROPERTY STRINGS "Debug" "Release" "MinSizeRel" "RelWithDebInfo" "ReleaseGG" "RelWithDebInfoGG")
+endif()

ggml/cmake/GitVars.cmake

@@ -0,0 +1,22 @@
+find_package(Git)
+
+# the commit's SHA1
+execute_process(COMMAND
+    "${GIT_EXECUTABLE}" describe --match=NeVeRmAtCh --always --abbrev=8
+    WORKING_DIRECTORY "${CMAKE_SOURCE_DIR}"
+    OUTPUT_VARIABLE GIT_SHA1
+    ERROR_QUIET OUTPUT_STRIP_TRAILING_WHITESPACE)
+
+# the date of the commit
+execute_process(COMMAND
+    "${GIT_EXECUTABLE}" log -1 --format=%ad --date=local
+    WORKING_DIRECTORY "${CMAKE_SOURCE_DIR}"
+    OUTPUT_VARIABLE GIT_DATE
+    ERROR_QUIET OUTPUT_STRIP_TRAILING_WHITESPACE)
+
+# the subject of the commit
+execute_process(COMMAND
+    "${GIT_EXECUTABLE}" log -1 --format=%s
+    WORKING_DIRECTORY "${CMAKE_SOURCE_DIR}"
+    OUTPUT_VARIABLE GIT_COMMIT_SUBJECT
+    ERROR_QUIET OUTPUT_STRIP_TRAILING_WHITESPACE)

ggml/examples/CMakeLists.txt

@@ -0,0 +1,30 @@
+if (GGML_ALL_WARNINGS)
+  if (NOT MSVC)
+      set(cxx_flags
+          # TODO(marella): Add other warnings.
+          -Wpedantic
+          -Wunused-variable
+          -Wno-unused-function
+          -Wno-multichar
+      )
+      add_compile_options("$<$<COMPILE_LANGUAGE:CXX>:${cxx_flags}>")
+  endif()
+endif()
+
+add_library(common STATIC common.cpp)
+target_include_directories(common PUBLIC ${CMAKE_CURRENT_SOURCE_DIR})
+
+add_library(common-ggml STATIC common-ggml.cpp)
+target_link_libraries(common-ggml PRIVATE ggml)
+target_include_directories(common-ggml PUBLIC ${CMAKE_CURRENT_SOURCE_DIR})
+
+add_subdirectory(gpt-2)
+add_subdirectory(gpt-j)
+add_subdirectory(whisper)
+add_subdirectory(mnist)
+add_subdirectory(gpt-neox)
+add_subdirectory(dolly-v2)
+add_subdirectory(replit)
+add_subdirectory(mpt)
+add_subdirectory(starcoder)
+add_subdirectory(sam)

ggml/examples/common-ggml.cpp

@@ -0,0 +1,246 @@
+#include "common-ggml.h"
+
+#include <regex>
+#include <map>
+
+static const std::map<std::string, enum ggml_ftype> GGML_FTYPE_MAP = {
+    {"q4_0", GGML_FTYPE_MOSTLY_Q4_0},
+    {"q4_1", GGML_FTYPE_MOSTLY_Q4_1},
+    {"q5_0", GGML_FTYPE_MOSTLY_Q5_0},
+    {"q5_1", GGML_FTYPE_MOSTLY_Q5_1},
+    {"q8_0", GGML_FTYPE_MOSTLY_Q8_0},
+};
+
+void ggml_print_ftypes(FILE * fp) {
+    for (auto it = GGML_FTYPE_MAP.begin(); it != GGML_FTYPE_MAP.end(); it++) {
+        fprintf(fp, "  type = \"%s\" or %d\n", it->first.c_str(), it->second);
+    }
+}
+
+enum ggml_ftype ggml_parse_ftype(const char * str) {
+    enum ggml_ftype ftype;
+    if (str[0] == 'q') {
+        const auto it = GGML_FTYPE_MAP.find(str);
+        if (it == GGML_FTYPE_MAP.end()) {
+            fprintf(stderr, "%s: unknown ftype '%s'\n", __func__, str);
+            return GGML_FTYPE_UNKNOWN;
+        }
+        ftype = it->second;
+    } else {
+        ftype = (enum ggml_ftype) atoi(str);
+    }
+
+    return ftype;
+}
+
+bool ggml_common_quantize_0(
+        std::ifstream & finp,
+        std::ofstream & fout,
+        const ggml_ftype ftype,
+        const std::vector<std::string> & to_quant,
+        const std::vector<std::string> & to_skip) {
+
+    ggml_type qtype = GGML_TYPE_F32;
+
+    switch (ftype) {
+        case GGML_FTYPE_MOSTLY_Q4_0: qtype = GGML_TYPE_Q4_0; break;
+        case GGML_FTYPE_MOSTLY_Q4_1: qtype = GGML_TYPE_Q4_1; break;
+        case GGML_FTYPE_MOSTLY_Q5_0: qtype = GGML_TYPE_Q5_0; break;
+        case GGML_FTYPE_MOSTLY_Q5_1: qtype = GGML_TYPE_Q5_1; break;
+        case GGML_FTYPE_MOSTLY_Q8_0: qtype = GGML_TYPE_Q8_0; break;
+        case GGML_FTYPE_UNKNOWN:
+        case GGML_FTYPE_ALL_F32:
+        case GGML_FTYPE_MOSTLY_F16:
+        case GGML_FTYPE_MOSTLY_Q4_1_SOME_F16:
+        case GGML_FTYPE_MOSTLY_Q2_K:
+        case GGML_FTYPE_MOSTLY_Q3_K:
+        case GGML_FTYPE_MOSTLY_Q4_K:
+        case GGML_FTYPE_MOSTLY_Q5_K:
+        case GGML_FTYPE_MOSTLY_Q6_K:
+                {
+                    fprintf(stderr, "%s: invalid model type %d\n", __func__, ftype);
+                    return false;
+                }
+    };
+
+    if (!ggml_is_quantized(qtype)) {
+        fprintf(stderr, "%s: invalid quantization type %d (%s)\n", __func__, qtype, ggml_type_name(qtype));
+        return false;
+    }
+
+    size_t total_size_org = 0;
+    size_t total_size_new = 0;
+
+    std::vector<float> work;
+
+    std::vector<uint8_t>     data_u8;
+    std::vector<ggml_fp16_t> data_f16;
+    std::vector<float>       data_f32;
+
+    std::vector<int64_t> hist_all(1 << 4, 0);
+
+    while (true) {
+        int32_t n_dims;
+        int32_t length;
+        int32_t ttype;
+
+        finp.read(reinterpret_cast<char *>(&n_dims), sizeof(n_dims));
+        finp.read(reinterpret_cast<char *>(&length), sizeof(length));
+        finp.read(reinterpret_cast<char *>(&ttype),  sizeof(ttype));
+
+        if (finp.eof()) {
+            break;
+        }
+
+        int32_t nelements = 1;
+        int32_t ne[4] = { 1, 1, 1, 1 };
+        for (int i = 0; i < n_dims; ++i) {
+            finp.read (reinterpret_cast<char *>(&ne[i]), sizeof(ne[i]));
+            nelements *= ne[i];
+        }
+
+        std::string name(length, 0);
+        finp.read (&name[0], length);
+
+        printf("%64s - [%5d, %5d, %5d], type = %6s ", name.data(), ne[0], ne[1], ne[2], ggml_type_name((ggml_type) ttype));
+
+        bool quantize = false;
+
+        // check if we should quantize this tensor
+        for (const auto & s : to_quant) {
+            if (std::regex_match(name, std::regex(s))) {
+                quantize = true;
+                break;
+            }
+        }
+
+        // check if we should skip this tensor
+        for (const auto & s : to_skip) {
+            if (std::regex_match(name, std::regex(s))) {
+                quantize = false;
+                break;
+            }
+        }
+
+        // quantize only 2D tensors
+        quantize &= (n_dims == 2);
+
+        if (quantize) {
+            if (ttype != GGML_TYPE_F32 && ttype != GGML_TYPE_F16) {
+                fprintf(stderr, "%s: unsupported ttype %d (%s) for integer quantization\n", __func__, ttype, ggml_type_name((ggml_type) ttype));
+                return false;
+            }
+
+            if (ttype == GGML_TYPE_F16) {
+                data_f16.resize(nelements);
+                finp.read(reinterpret_cast<char *>(data_f16.data()), nelements * sizeof(ggml_fp16_t));
+                data_f32.resize(nelements);
+                for (int i = 0; i < nelements; ++i) {
+                    data_f32[i] = ggml_fp16_to_fp32(data_f16[i]);
+                }
+            } else {
+                data_f32.resize(nelements);
+                finp.read(reinterpret_cast<char *>(data_f32.data()), nelements * sizeof(float));
+            }
+
+            ttype = qtype;
+        } else {
+            const int bpe = (ttype == 0) ? sizeof(float) : sizeof(uint16_t);
+
+            data_u8.resize(nelements*bpe);
+            finp.read(reinterpret_cast<char *>(data_u8.data()), nelements * bpe);
+        }
+
+        fout.write(reinterpret_cast<char *>(&n_dims), sizeof(n_dims));
+        fout.write(reinterpret_cast<char *>(&length), sizeof(length));
+        fout.write(reinterpret_cast<char *>(&ttype),  sizeof(ttype));
+        for (int i = 0; i < n_dims; ++i) {
+            fout.write(reinterpret_cast<char *>(&ne[i]), sizeof(ne[i]));
+        }
+        fout.write(&name[0], length);
+
+        if (quantize) {
+            work.resize(nelements); // for quantization
+
+            size_t cur_size = 0;
+            std::vector<int64_t> hist_cur(1 << 4, 0);
+
+            switch ((ggml_type) ttype) {
+                case GGML_TYPE_Q4_0:
+                    {
+                        cur_size = ggml_quantize_q4_0(data_f32.data(), work.data(), nelements, ne[0], hist_cur.data());
+                    } break;
+                case GGML_TYPE_Q4_1:
+                    {
+                        cur_size = ggml_quantize_q4_1(data_f32.data(), work.data(), nelements, ne[0], hist_cur.data());
+                    } break;
+                case GGML_TYPE_Q5_0:
+                    {
+                        cur_size = ggml_quantize_q5_0(data_f32.data(), work.data(), nelements, ne[0], hist_cur.data());
+                    } break;
+                case GGML_TYPE_Q5_1:
+                    {
+                        cur_size = ggml_quantize_q5_1(data_f32.data(), work.data(), nelements, ne[0], hist_cur.data());
+                    } break;
+                case GGML_TYPE_Q8_0:
+                    {
+                        cur_size = ggml_quantize_q8_0(data_f32.data(), work.data(), nelements, ne[0], hist_cur.data());
+                    } break;
+                case GGML_TYPE_F32:
+                case GGML_TYPE_F16:
+                case GGML_TYPE_I8:
+                case GGML_TYPE_I16:
+                case GGML_TYPE_I32:
+                case GGML_TYPE_Q8_1:
+                case GGML_TYPE_Q2_K:
+                case GGML_TYPE_Q3_K:
+                case GGML_TYPE_Q4_K:
+                case GGML_TYPE_Q5_K:
+                case GGML_TYPE_Q6_K:
+                case GGML_TYPE_Q8_K:
+                case GGML_TYPE_COUNT:
+                    {
+                        fprintf(stderr, "%s: unsupported quantization type %d (%s)\n", __func__, ttype, ggml_type_name((ggml_type) ttype));
+                        return false;
+                    }
+            }
+
+            fout.write(reinterpret_cast<char *>(work.data()), cur_size);
+            total_size_new += cur_size;
+
+            printf("size = %8.2f MB -> %8.2f MB | hist: ", nelements * sizeof(float)/1024.0/1024.0, cur_size/1024.0/1024.0);
+            for (int i = 0; i < (int) hist_cur.size(); ++i) {
+                hist_all[i] += hist_cur[i];
+            }
+
+            for (int i = 0; i < (int) hist_cur.size(); ++i) {
+                printf("%5.3f ", hist_cur[i] / (float)nelements);
+            }
+            printf("\n");
+        } else {
+            printf("size = %8.3f MB\n", data_u8.size()/1024.0/1024.0);
+            fout.write(reinterpret_cast<char *>(data_u8.data()), data_u8.size());
+            total_size_new += data_u8.size();
+        }
+
+        total_size_org += nelements * sizeof(float);
+    }
+
+    printf("%s: model size  = %8.2f MB\n", __func__, total_size_org/1024.0/1024.0);
+    printf("%s: quant size  = %8.2f MB | ftype = %d (%s)\n", __func__, total_size_new/1024.0/1024.0, ftype, ggml_type_name(qtype));
+
+    {
+        int64_t sum_all = 0;
+        for (int i = 0; i < (int) hist_all.size(); ++i) {
+            sum_all += hist_all[i];
+        }
+
+        printf("%s: hist: ", __func__);
+        for (int i = 0; i < (int) hist_all.size(); ++i) {
+            printf("%5.3f ", hist_all[i] / (float)sum_all);
+        }
+        printf("\n");
+    }
+
+    return true;
+}

ggml/examples/common-ggml.h

@@ -0,0 +1,18 @@
+#pragma once
+
+#include "ggml.h"
+
+#include <fstream>
+#include <vector>
+#include <string>
+
+enum ggml_ftype ggml_parse_ftype(const char * str);
+
+void ggml_print_ftypes(FILE * fp = stderr);
+
+bool ggml_common_quantize_0(
+        std::ifstream & finp,
+        std::ofstream & fout,
+        const ggml_ftype ftype,
+        const std::vector<std::string> & to_quant,
+        const std::vector<std::string> & to_skip);

ggml/examples/common.cpp

@@ -0,0 +1,809 @@
+#define _USE_MATH_DEFINES // for M_PI
+
+#include "common.h"
+
+// third-party utilities
+// use your favorite implementations
+#define DR_WAV_IMPLEMENTATION
+#include "dr_wav.h"
+
+#include <cmath>
+#include <cstring>
+#include <fstream>
+#include <regex>
+#include <locale>
+#include <codecvt>
+#include <sstream>
+
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
+// Function to check if the next argument exists
+std::string get_next_arg(int& i, int argc, char** argv, const std::string& flag, gpt_params& params) {
+    if (i + 1 < argc && argv[i + 1][0] != '-') {
+        return argv[++i];
+    } else {
+        fprintf(stderr, "error: %s requires one argument.\n", flag.c_str());
+        gpt_print_usage(argc, argv, params);
+        exit(0);
+    }
+}
+
+bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
+    for (int i = 1; i < argc; i++) {
+        std::string arg = argv[i];
+
+        if (arg == "-s" || arg == "--seed") {
+            params.seed = std::stoi(get_next_arg(i, argc, argv, arg, params));
+        } else if (arg == "-t" || arg == "--threads") {
+            params.n_threads = std::stoi(get_next_arg(i, argc, argv, arg, params));
+        } else if (arg == "-ngl" || arg == "--gpu-layers" || arg == "--n-gpu-layers") {
+            params.n_gpu_layers = std::stoi(get_next_arg(i, argc, argv, arg, params));
+        } else if (arg == "-p" || arg == "--prompt") {
+            params.prompt = get_next_arg(i, argc, argv, arg, params);
+        } else if (arg == "-n" || arg == "--n_predict") {
+            params.n_predict = std::stoi(get_next_arg(i, argc, argv, arg, params));
+        } else if (arg == "--top_k") {
+            params.top_k = std::stoi(get_next_arg(i, argc, argv, arg, params));
+        } else if (arg == "--top_p") {
+            params.top_p = std::stof(get_next_arg(i, argc, argv, arg, params));
+        } else if (arg == "--temp") {
+            params.temp = std::stof(get_next_arg(i, argc, argv, arg, params));
+        } else if (arg == "--repeat-last-n") {
+            params.repeat_last_n = std::stoi(get_next_arg(i, argc, argv, arg, params));
+        } else if (arg == "--repeat-penalty") {
+            params.repeat_penalty = std::stof(get_next_arg(i, argc, argv, arg, params));
+        } else if (arg == "-b" || arg == "--batch_size") {
+            params.n_batch= std::stoi(get_next_arg(i, argc, argv, arg, params));
+        } else if (arg == "-m" || arg == "--model") {
+            params.model = get_next_arg(i, argc, argv, arg, params);
+        } else if (arg == "-i" || arg == "--interactive") {
+            params.interactive = true;
+        } else if (arg == "-ip" || arg == "--interactive-port") {
+            params.interactive = true;
+            params.interactive_port = std::stoi(get_next_arg(i, argc, argv, arg, params));
+        } else if (arg == "-h" || arg == "--help") {
+            gpt_print_usage(argc, argv, params);
+            exit(0);
+        } else if (arg == "-f" || arg == "--file") {
+            get_next_arg(i, argc, argv, arg, params);
+            std::ifstream file(argv[i]);
+            if (!file) {
+                fprintf(stderr, "error: failed to open file '%s'\n", argv[i]);
+                break;
+            }
+            std::copy(std::istreambuf_iterator<char>(file), std::istreambuf_iterator<char>(), back_inserter(params.prompt));
+            if (params.prompt.back() == '\n') {
+                params.prompt.pop_back();
+            }
+        } else if (arg == "-tt" || arg == "--token_test") {
+            params.token_test = get_next_arg(i, argc, argv, arg, params);
+        }
+        else {
+            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
+            gpt_print_usage(argc, argv, params);
+            exit(0);
+        }
+    }
+
+    return true;
+}
+
+void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
+    fprintf(stderr, "usage: %s [options]\n", argv[0]);
+    fprintf(stderr, "\n");
+    fprintf(stderr, "options:\n");
+    fprintf(stderr, "  -h, --help            show this help message and exit\n");
+    fprintf(stderr, "  -s SEED, --seed SEED  RNG seed (default: -1)\n");
+    fprintf(stderr, "  -t N, --threads N     number of threads to use during computation (default: %d)\n", params.n_threads);
+    fprintf(stderr, "  -ngl N, --gpu-layers N  number of layers to offload to GPU on supported models (default: %d)\n", params.n_gpu_layers);
+    fprintf(stderr, "  -p PROMPT, --prompt PROMPT\n");
+    fprintf(stderr, "                        prompt to start generation with (default: random)\n");
+    fprintf(stderr, "  -f FNAME, --file FNAME\n");
+    fprintf(stderr, "                        load prompt from a file\n");
+    fprintf(stderr, "  -tt TOKEN_TEST, --token_test TOKEN_TEST\n");
+    fprintf(stderr, "                        test tokenization\n");
+    fprintf(stderr, "  -n N, --n_predict N   number of tokens to predict (default: %d)\n", params.n_predict);
+    fprintf(stderr, "  --top_k N             top-k sampling (default: %d)\n", params.top_k);
+    fprintf(stderr, "  --top_p N             top-p sampling (default: %.1f)\n", params.top_p);
+    fprintf(stderr, "  --temp N              temperature (default: %.1f)\n", params.temp);
+    fprintf(stderr, "  --repeat-last-n N     last n tokens to consider for penalize (default: %d, 0 = disabled)\n", params.repeat_last_n);
+    fprintf(stderr, "  --repeat-penalty N    penalize repeat sequence of tokens (default: %.2f, 1.0 = disabled)\n", (double)params.repeat_penalty);
+    fprintf(stderr, "  -b N, --batch_size N  batch size for prompt processing (default: %d)\n", params.n_batch);
+    fprintf(stderr, "  -m FNAME, --model FNAME\n");
+    fprintf(stderr, "                        model path (default: %s)\n", params.model.c_str());
+    fprintf(stderr, "\n");
+}
+
+std::string gpt_random_prompt(std::mt19937 & rng) {
+    const int r = rng() % 10;
+    switch (r) {
+        case 0: return "So";
+        case 1: return "Once upon a time";
+        case 2: return "When";
+        case 3: return "The";
+        case 4: return "After";
+        case 5: return "If";
+        case 6: return "import";
+        case 7: return "He";
+        case 8: return "She";
+        case 9: return "They";
+        default: return "To";
+    }
+
+    return "The";
+}
+
+std::string trim(const std::string & s) {
+    std::regex e("^\\s+|\\s+$");
+    return std::regex_replace(s, e, "");
+}
+
+std::string replace(const std::string & s, const std::string & from, const std::string & to) {
+    std::string result = s;
+    size_t pos = 0;
+    while ((pos = result.find(from, pos)) != std::string::npos) {
+        result.replace(pos, from.length(), to);
+        pos += to.length();
+    }
+    return result;
+}
+
+void gpt_vocab::add_special_token(const std::string & token) {
+    special_tokens.push_back(token);
+}
+
+std::map<std::string, int32_t> json_parse(const std::string & fname) {
+    std::map<std::string, int32_t> result;
+
+    // read file into string
+    std::string json;
+    {
+        std::ifstream ifs(fname);
+        if (!ifs) {
+            fprintf(stderr, "Failed to open %s\n", fname.c_str());
+            exit(1);
+        }
+
+        json = std::string((std::istreambuf_iterator<char>(ifs)),
+                (std::istreambuf_iterator<char>()));
+    }
+
+    if (json[0] != '{') {
+        return result;
+    }
+
+    // parse json
+    {
+        bool has_key  = false;
+        bool in_token = false;
+
+        std::string str_key = "";
+        std::string str_val = "";
+
+        int n = json.size();
+        for (int i = 1; i < n; ++i) {
+            if (!in_token) {
+                if (json[i] == ' ') continue;
+                if (json[i] == '"') {
+                    in_token = true;
+                    continue;
+                }
+            } else {
+                if (json[i] == '\\' && i+1 < n) {
+                    if (has_key == false) {
+                        str_key += json[i];
+                    } else {
+                        str_val += json[i];
+                    }
+                    ++i;
+                } else if (json[i] == '"') {
+                    if (has_key == false) {
+                        has_key = true;
+                        ++i;
+                        while (json[i] == ' ') ++i;
+                        ++i; // :
+                        while (json[i] == ' ') ++i;
+                        if (json[i] != '\"') {
+                            while (json[i] != ',' && json[i] != '}') {
+                                str_val += json[i++];
+                            }
+                            has_key = false;
+                        } else {
+                            in_token = true;
+                            continue;
+                        }
+                    } else {
+                        has_key = false;
+                    }
+
+                    str_key = ::replace(str_key, "\\u0120", " " ); // \u0120 -> space
+                    str_key = ::replace(str_key, "\\u010a", "\n"); // \u010a -> new line
+                    str_key = ::replace(str_key, "\\\"",    "\""); // \\\"   -> "
+
+                    try {
+                        result[str_key] = std::stoi(str_val);
+                    } catch (...) {
+                        //fprintf(stderr, "%s: ignoring key '%s' with value '%s'\n", fname.c_str(), str_key.c_str(), str_val.c_str());
+
+                    }
+                    str_key = "";
+                    str_val = "";
+                    in_token = false;
+                    continue;
+                }
+                if (has_key == false) {
+                    str_key += json[i];
+                } else {
+                    str_val += json[i];
+                }
+            }
+        }
+    }
+
+    return result;
+}
+
+std::string convert_to_utf8(const std::wstring & input) {
+    std::wstring_convert<std::codecvt_utf8<wchar_t>> converter;
+    return converter.to_bytes(input);
+}
+
+
+std::wstring convert_to_wstring(const std::string & input) {
+    std::wstring_convert<std::codecvt_utf8<wchar_t>> converter;
+    return converter.from_bytes(input);
+}
+
+void gpt_split_words(std::string str, std::vector<std::string>& words) {
+    const std::string pattern = R"('s|'t|'re|'ve|'m|'ll|'d| ?[[:alpha:]]+| ?[[:digit:]]+| ?[^\s[:alpha:][:digit:]]+|\s+(?!\S)|\s+)";
+    const std::regex re(pattern);
+    std::smatch m;
+
+    while (std::regex_search(str, m, re)) {
+        for (auto x : m) {
+            words.push_back(x);
+        }
+        str = m.suffix();
+    }
+}
+
+std::vector<gpt_vocab::id> gpt_tokenize(const gpt_vocab & vocab, const std::string & text) {
+    std::vector<std::string> words;
+
+    // first split the text into words
+    {
+        std::string str = text;
+
+        // Generate the subpattern from the special_tokens vector if it's not empty
+        if (!vocab.special_tokens.empty()) {
+            const std::regex escape(R"([\[\\\^\$\.\|\?\*\+\(\)\{\}])");
+            std::string special_tokens_subpattern;
+            for (const auto & token : vocab.special_tokens) {
+                if (!special_tokens_subpattern.empty()) {
+                    special_tokens_subpattern += "|";
+                }
+                special_tokens_subpattern += std::regex_replace(token, escape, R"(\$&)");
+            }
+
+            std::regex re(special_tokens_subpattern);
+            std::smatch m;
+            // Split the text by special tokens.
+            while (std::regex_search(str, m, re)) {
+                // Split the substrings in-between special tokens into words.
+                gpt_split_words(m.prefix(), words);
+                // Add matched special tokens as words.
+                for (auto x : m) {
+                    words.push_back(x);
+                }
+                str = m.suffix();
+            }
+            // Remaining text without special tokens will be handled below.
+        }
+
+        gpt_split_words(str, words);
+    }
+
+    // find the longest token that forms each word in words:
+    std::vector<gpt_vocab::id> tokens;
+    for (const auto & word : words) {
+        for (int i = 0; i < (int) word.size(); ){
+            for (int j = word.size() - 1; j >= i; j--){
+                auto cand = word.substr(i, j-i+1);
+                auto it = vocab.token_to_id.find(cand);
+                if (it != vocab.token_to_id.end()){ // word.substr(i, j-i+1) in vocab
+                    tokens.push_back(it->second);
+                    i = j + 1;
+                    break;
+                }
+                else if (j == i){ // word.substr(i, 1) has no matching
+                    fprintf(stderr, "%s: unknown token '%s'\n", __func__, word.substr(i, 1).data());
+                    i++;
+                }
+            }
+        }
+    }
+
+    return tokens;
+}
+
+std::vector<gpt_vocab::id> parse_tokens_from_string(const std::string& input, char delimiter) {
+    std::vector<gpt_vocab::id> output;
+    std::stringstream ss(input);
+    std::string token;
+
+    while (std::getline(ss, token, delimiter)) {
+        output.push_back(std::stoi(token));
+    }
+
+    return output;
+}
+
+std::map<std::string, std::vector<gpt_vocab::id>> extract_tests_from_file(const std::string & fpath_test){
+    if (fpath_test.empty()){
+        fprintf(stderr, "%s : No test file found.\n", __func__);
+        return std::map<std::string, std::vector<gpt_vocab::id>>();
+    }
+
+    std::map<std::string, std::vector<gpt_vocab::id>> tests;
+
+    auto fin = std::ifstream(fpath_test, std::ios_base::in);
+    const char * delimeter = " => ";
+    const char del_tok = ',';
+    std::string line;
+    while (std::getline(fin, line)) {
+        size_t delimiterPos = line.find(delimeter);
+        if (delimiterPos != std::string::npos) {
+            std::string text = line.substr(0, delimiterPos);
+            std::string s_tokens = line.substr(delimiterPos + std::strlen(delimeter));
+            tests[text] = parse_tokens_from_string(s_tokens, del_tok);
+        }
+    }
+    return tests;
+}
+
+void test_gpt_tokenizer(gpt_vocab & vocab, const std::string & fpath_test){
+    std::map<std::string, std::vector<gpt_vocab::id>> tests = extract_tests_from_file(fpath_test);
+
+    size_t n_fails = 0;
+
+    for (const auto & test : tests) {
+        std::vector<gpt_vocab::id> tokens = gpt_tokenize(vocab, test.first);
+
+        if (tokens != test.second){
+            n_fails++;
+
+            // print out failure cases
+            fprintf(stderr, "%s : failed test: '%s'\n", __func__, test.first.c_str());
+            fprintf(stderr, "%s : tokens in hf:   ", __func__);
+            for (const auto & t : test.second) {
+                fprintf(stderr, "%s(%d), ", vocab.id_to_token[t].c_str(), t);
+            }
+            fprintf(stderr, "\n");
+            fprintf(stderr, "%s : tokens in ggml: ", __func__);
+            for (const auto & t : tokens) {
+                fprintf(stderr, "%s(%d), ", vocab.id_to_token[t].c_str(), t);
+            }
+            fprintf(stderr, "\n");
+        }
+    }
+
+    fprintf(stderr, "%s : %zu tests failed out of %zu tests.\n", __func__, n_fails, tests.size());
+}
+
+bool gpt_vocab_init(const std::string & fname, gpt_vocab & vocab) {
+    printf("%s: loading vocab from '%s'\n", __func__, fname.c_str());
+
+    vocab.token_to_id = ::json_parse(fname);
+
+    for (const auto & kv : vocab.token_to_id) {
+        vocab.id_to_token[kv.second] = kv.first;
+    }
+
+    printf("%s: vocab size = %d\n", __func__, (int) vocab.token_to_id.size());
+
+    // print the vocabulary
+    //for (auto kv : vocab.token_to_id) {
+    //    printf("'%s' -> %d\n", kv.first.data(), kv.second);
+    //}
+
+    return true;
+}
+
+gpt_vocab::id gpt_sample_top_k_top_p(
+        const gpt_vocab & vocab,
+        const float * logits,
+        int    top_k,
+        double top_p,
+        double temp,
+        std::mt19937 & rng) {
+    int n_logits = vocab.id_to_token.size();
+
+    std::vector<std::pair<double, gpt_vocab::id>> logits_id;
+    logits_id.reserve(n_logits);
+
+    {
+        const double scale = 1.0/temp;
+        for (int i = 0; i < n_logits; ++i) {
+            logits_id.push_back(std::make_pair(logits[i]*scale, i));
+        }
+    }
+
+    // find the top K tokens
+    std::partial_sort(
+            logits_id.begin(),
+            logits_id.begin() + top_k, logits_id.end(),
+            [](const std::pair<double, gpt_vocab::id> & a, const std::pair<double, gpt_vocab::id> & b) {
+        return a.first > b.first;
+    });
+
+    logits_id.resize(top_k);
+
+    double maxl = -INFINITY;
+    for (const auto & kv : logits_id) {
+        maxl = std::max(maxl, kv.first);
+    }
+
+    // compute probs for the top K tokens
+    std::vector<double> probs;
+    probs.reserve(logits_id.size());
+
+    double sum = 0.0;
+    for (const auto & kv : logits_id) {
+        double p = exp(kv.first - maxl);
+        probs.push_back(p);
+        sum += p;
+    }
+
+    // normalize the probs
+    for (auto & p : probs) {
+        p /= sum;
+    }
+
+    if (top_p < 1.0f) {
+        double cumsum = 0.0f;
+        for (int i = 0; i < top_k; i++) {
+            cumsum += probs[i];
+            if (cumsum >= top_p) {
+                top_k = i + 1;
+                probs.resize(top_k);
+                logits_id.resize(top_k);
+                break;
+            }
+        }
+
+        cumsum = 1.0/cumsum;
+        for (int i = 0; i < (int) probs.size(); i++) {
+            probs[i] *= cumsum;
+        }
+    }
+
+    //printf("\n");
+    //for (int i = 0; i < (int) probs.size(); i++) {
+    //    printf("%d: '%s' %f\n", i, vocab.id_to_token.at(logits_id[i].second).c_str(), probs[i]);
+    //}
+    //exit(0);
+
+    std::discrete_distribution<> dist(probs.begin(), probs.end());
+    int idx = dist(rng);
+
+    return logits_id[idx].second;
+}
+
+gpt_vocab::id gpt_sample_top_k_top_p_repeat(
+        const gpt_vocab & vocab,
+        const float * logits,
+        const int32_t * last_n_tokens_data,
+        size_t last_n_tokens_data_size,
+        int    top_k,
+        double top_p,
+        double temp,
+        int repeat_last_n,
+        float repeat_penalty,
+        std::mt19937 & rng) {
+
+    int n_logits = vocab.id_to_token.size();
+
+    const auto * plogits = logits;
+
+    const auto last_n_tokens = std::vector<int32_t>(last_n_tokens_data, last_n_tokens_data + last_n_tokens_data_size);
+
+    if (temp <= 0) {
+        // select the token with the highest logit directly
+        float max_logit = plogits[0];
+        gpt_vocab::id max_id = 0;
+
+        for (int i = 1; i < n_logits; ++i) {
+            if (plogits[i] > max_logit) {
+                max_logit = plogits[i];
+                max_id = i;
+            }
+        }
+        return max_id;
+    }
+
+
+    std::vector<std::pair<double, gpt_vocab::id>> logits_id;
+    logits_id.reserve(n_logits);
+
+    {
+        const float scale = 1.0f/temp;
+        for (int i = 0; i < n_logits; ++i) {
+            // repetition penalty from ctrl paper (https://arxiv.org/abs/1909.05858)
+            // credit https://github.com/facebookresearch/llama/compare/main...shawwn:llama:main
+            if (repeat_last_n > 0 && std::find(last_n_tokens.end()-repeat_last_n, last_n_tokens.end(), i) != last_n_tokens.end()) {
+                // if score < 0 then repetition penalty has to multiplied to reduce the previous token probability
+                if (plogits[i] < 0.0f) {
+                    logits_id.push_back(std::make_pair(plogits[i]*scale*repeat_penalty, i));
+                } else {
+                    logits_id.push_back(std::make_pair(plogits[i]*scale/repeat_penalty, i));
+                }
+            } else {
+                logits_id.push_back(std::make_pair(plogits[i]*scale, i));
+            }
+        }
+    }
+
+    // find the top K tokens
+    std::partial_sort(
+            logits_id.begin(),
+            logits_id.begin() + top_k, logits_id.end(),
+            [](const std::pair<double, gpt_vocab::id> & a, const std::pair<double, gpt_vocab::id> & b) {
+        return a.first > b.first;
+    });
+
+    logits_id.resize(top_k);
+
+    double maxl = -INFINITY;
+    for (const auto & kv : logits_id) {
+        maxl = std::max(maxl, kv.first);
+    }
+
+    // compute probs for the top K tokens
+    std::vector<double> probs;
+    probs.reserve(logits_id.size());
+
+    double sum = 0.0;
+    for (const auto & kv : logits_id) {
+        double p = exp(kv.first - maxl);
+        probs.push_back(p);
+        sum += p;
+    }
+
+    // normalize the probs
+    for (auto & p : probs) {
+        p /= sum;
+    }
+
+    if (top_p < 1.0f) {
+        double cumsum = 0.0f;
+        for (int i = 0; i < top_k; i++) {
+            cumsum += probs[i];
+            if (cumsum >= top_p) {
+                top_k = i + 1;
+                probs.resize(top_k);
+                logits_id.resize(top_k);
+                break;
+            }
+        }
+
+        cumsum = 1.0/cumsum;
+        for (int i = 0; i < (int) probs.size(); i++) {
+            probs[i] *= cumsum;
+        }
+    }
+
+//    printf("\n");
+//    for (int i = 0; i < (int) probs.size(); i++) {
+//    for (int i = 0; i < 10; i++) {
+//        printf("%d: '%s' %f\n", i, vocab.id_to_token.at(logits_id[i].second).c_str(), probs[i]);
+//    }
+
+    std::discrete_distribution<> dist(probs.begin(), probs.end());
+    int idx = dist(rng);
+
+    return logits_id[idx].second;
+
+}
+
+bool read_wav(const std::string & fname, std::vector<float>& pcmf32, std::vector<std::vector<float>>& pcmf32s, bool stereo) {
+    drwav wav;
+    std::vector<uint8_t> wav_data; // used for pipe input from stdin
+
+    if (fname == "-") {
+        {
+            uint8_t buf[1024];
+            while (true)
+            {
+                const size_t n = fread(buf, 1, sizeof(buf), stdin);
+                if (n == 0) {
+                    break;
+                }
+                wav_data.insert(wav_data.end(), buf, buf + n);
+            }
+        }
+
+        if (drwav_init_memory(&wav, wav_data.data(), wav_data.size(), nullptr) == false) {
+            fprintf(stderr, "error: failed to open WAV file from stdin\n");
+            return false;
+        }
+
+        fprintf(stderr, "%s: read %zu bytes from stdin\n", __func__, wav_data.size());
+    }
+    else if (drwav_init_file(&wav, fname.c_str(), nullptr) == false) {
+        fprintf(stderr, "error: failed to open '%s' as WAV file\n", fname.c_str());
+        return false;
+    }
+
+    if (wav.channels != 1 && wav.channels != 2) {
+        fprintf(stderr, "%s: WAV file '%s' must be mono or stereo\n", __func__, fname.c_str());
+        return false;
+    }
+
+    if (stereo && wav.channels != 2) {
+        fprintf(stderr, "%s: WAV file '%s' must be stereo for diarization\n", __func__, fname.c_str());
+        return false;
+    }
+
+    if (wav.sampleRate != COMMON_SAMPLE_RATE) {
+        fprintf(stderr, "%s: WAV file '%s' must be %i kHz\n", __func__, fname.c_str(), COMMON_SAMPLE_RATE/1000);
+        return false;
+    }
+
+    if (wav.bitsPerSample != 16) {
+        fprintf(stderr, "%s: WAV file '%s' must be 16-bit\n", __func__, fname.c_str());
+        return false;
+    }
+
+    const uint64_t n = wav_data.empty() ? wav.totalPCMFrameCount : wav_data.size()/(wav.channels*wav.bitsPerSample/8);
+
+    std::vector<int16_t> pcm16;
+    pcm16.resize(n*wav.channels);
+    drwav_read_pcm_frames_s16(&wav, n, pcm16.data());
+    drwav_uninit(&wav);
+
+    // convert to mono, float
+    pcmf32.resize(n);
+    if (wav.channels == 1) {
+        for (uint64_t i = 0; i < n; i++) {
+            pcmf32[i] = float(pcm16[i])/32768.0f;
+        }
+    } else {
+        for (uint64_t i = 0; i < n; i++) {
+            pcmf32[i] = float(pcm16[2*i] + pcm16[2*i + 1])/65536.0f;
+        }
+    }
+
+    if (stereo) {
+        // convert to stereo, float
+        pcmf32s.resize(2);
+
+        pcmf32s[0].resize(n);
+        pcmf32s[1].resize(n);
+        for (uint64_t i = 0; i < n; i++) {
+            pcmf32s[0][i] = float(pcm16[2*i])/32768.0f;
+            pcmf32s[1][i] = float(pcm16[2*i + 1])/32768.0f;
+        }
+    }
+
+    return true;
+}
+
+void high_pass_filter(std::vector<float> & data, float cutoff, float sample_rate) {
+    const float rc = 1.0f / (2.0f * M_PI * cutoff);
+    const float dt = 1.0f / sample_rate;
+    const float alpha = dt / (rc + dt);
+
+    float y = data[0];
+
+    for (size_t i = 1; i < data.size(); i++) {
+        y = alpha * (y + data[i] - data[i - 1]);
+        data[i] = y;
+    }
+}
+
+bool vad_simple(std::vector<float> & pcmf32, int sample_rate, int last_ms, float vad_thold, float freq_thold, bool verbose) {
+    const int n_samples      = pcmf32.size();
+    const int n_samples_last = (sample_rate * last_ms) / 1000;
+
+    if (n_samples_last >= n_samples) {
+        // not enough samples - assume no speech
+        return false;
+    }
+
+    if (freq_thold > 0.0f) {
+        high_pass_filter(pcmf32, freq_thold, sample_rate);
+    }
+
+    float energy_all  = 0.0f;
+    float energy_last = 0.0f;
+
+    for (int i = 0; i < n_samples; i++) {
+        energy_all += fabsf(pcmf32[i]);
+
+        if (i >= n_samples - n_samples_last) {
+            energy_last += fabsf(pcmf32[i]);
+        }
+    }
+
+    energy_all  /= n_samples;
+    energy_last /= n_samples_last;
+
+    if (verbose) {
+        fprintf(stderr, "%s: energy_all: %f, energy_last: %f, vad_thold: %f, freq_thold: %f\n", __func__, energy_all, energy_last, vad_thold, freq_thold);
+    }
+
+    if (energy_last > vad_thold*energy_all) {
+        return false;
+    }
+
+    return true;
+}
+
+float similarity(const std::string & s0, const std::string & s1) {
+    const size_t len0 = s0.size() + 1;
+    const size_t len1 = s1.size() + 1;
+
+    std::vector<int> col(len1, 0);
+    std::vector<int> prevCol(len1, 0);
+
+    for (size_t i = 0; i < len1; i++) {
+        prevCol[i] = i;
+    }
+
+    for (size_t i = 0; i < len0; i++) {
+        col[0] = i;
+        for (size_t j = 1; j < len1; j++) {
+            col[j] = std::min(std::min(1 + col[j - 1], 1 + prevCol[j]), prevCol[j - 1] + (i > 0 && s0[i - 1] == s1[j - 1] ? 0 : 1));
+        }
+        col.swap(prevCol);
+    }
+
+    const float dist = prevCol[len1 - 1];
+
+    return 1.0f - (dist / std::max(s0.size(), s1.size()));
+}
+
+bool sam_params_parse(int argc, char ** argv, sam_params & params) {
+    for (int i = 1; i < argc; i++) {
+        std::string arg = argv[i];
+
+        if (arg == "-s" || arg == "--seed") {
+            params.seed = std::stoi(argv[++i]);
+        } else if (arg == "-t" || arg == "--threads") {
+            params.n_threads = std::stoi(argv[++i]);
+        } else if (arg == "-m" || arg == "--model") {
+            params.model = argv[++i];
+        } else if (arg == "-i" || arg == "--inp") {
+            params.fname_inp = argv[++i];
+        } else if (arg == "-o" || arg == "--out") {
+            params.fname_out = argv[++i];
+        } else if (arg == "-h" || arg == "--help") {
+            sam_print_usage(argc, argv, params);
+            exit(0);
+        } else {
+            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
+            sam_print_usage(argc, argv, params);
+            exit(0);
+        }
+    }
+
+    return true;
+}
+
+void sam_print_usage(int /*argc*/, char ** argv, const sam_params & params) {
+    fprintf(stderr, "usage: %s [options]\n", argv[0]);
+    fprintf(stderr, "\n");
+    fprintf(stderr, "options:\n");
+    fprintf(stderr, "  -h, --help            show this help message and exit\n");
+    fprintf(stderr, "  -s SEED, --seed SEED  RNG seed (default: -1)\n");
+    fprintf(stderr, "  -t N, --threads N     number of threads to use during computation (default: %d)\n", params.n_threads);
+    fprintf(stderr, "  -m FNAME, --model FNAME\n");
+    fprintf(stderr, "                        model path (default: %s)\n", params.model.c_str());
+    fprintf(stderr, "  -i FNAME, --inp FNAME\n");
+    fprintf(stderr, "                        input file (default: %s)\n", params.fname_inp.c_str());
+    fprintf(stderr, "  -o FNAME, --out FNAME\n");
+    fprintf(stderr, "                        output file (default: %s)\n", params.fname_out.c_str());
+    fprintf(stderr, "\n");
+}

ggml/examples/common.h

@@ -0,0 +1,176 @@
+// Various helper functions and utilities
+
+#pragma once
+
+#include <string>
+#include <map>
+#include <vector>
+#include <random>
+#include <thread>
+
+#define COMMON_SAMPLE_RATE 16000
+
+//
+// GPT CLI argument parsing
+//
+
+struct gpt_params {
+    int32_t seed      = -1;  // RNG seed
+    int32_t n_threads = std::min(4, (int32_t) std::thread::hardware_concurrency());
+    int32_t n_predict = 200; // new tokens to predict
+    int32_t n_batch   = 8;   // batch size for prompt processing
+
+    // sampling parameters
+    int32_t top_k          = 40;
+    float   top_p          = 0.9f;
+    float   temp           = 0.9f;
+    int32_t repeat_last_n  = 64;
+    float   repeat_penalty = 1.00f;
+
+    std::string model      = "models/gpt-2-117M/ggml-model.bin"; // model path
+    std::string prompt     = "";
+    std::string token_test = "";
+
+    bool    interactive      = false;
+    int32_t interactive_port = -1;
+
+    int32_t n_gpu_layers     = 0;
+};
+
+bool gpt_params_parse(int argc, char ** argv, gpt_params & params);
+
+void gpt_print_usage(int argc, char ** argv, const gpt_params & params);
+
+std::string gpt_random_prompt(std::mt19937 & rng);
+
+//
+// Vocab utils
+//
+
+std::string trim(const std::string & s);
+
+std::string replace(
+        const std::string & s,
+        const std::string & from,
+        const std::string & to);
+
+struct gpt_vocab {
+    using id    = int32_t;
+    using token = std::string;
+
+    std::map<token, id> token_to_id;
+    std::map<id, token> id_to_token;
+    std::vector<std::string> special_tokens;
+
+    void add_special_token(const std::string & token);
+};
+
+// poor-man's JSON parsing
+std::map<std::string, int32_t> json_parse(const std::string & fname);
+
+std::string convert_to_utf8(const std::wstring & input);
+
+std::wstring convert_to_wstring(const std::string & input);
+
+void gpt_split_words(std::string str, std::vector<std::string>& words);
+
+// split text into tokens
+//
+// ref: https://github.com/openai/gpt-2/blob/a74da5d99abaaba920de8131d64da2862a8f213b/src/encoder.py#L53
+//
+// Regex (Python):
+// r"""'s|'t|'re|'ve|'m|'ll|'d| ?\p{L}+| ?\p{N}+| ?[^\s\p{L}\p{N}]+|\s+(?!\S)|\s+"""
+//
+// Regex (C++):
+// R"('s|'t|'re|'ve|'m|'ll|'d| ?[[:alpha:]]+| ?[[:digit:]]+| ?[^\s[:alpha:][:digit:]]+|\s+(?!\S)|\s+)"
+//
+std::vector<gpt_vocab::id> gpt_tokenize(const gpt_vocab & vocab, const std::string & text);
+
+// test outputs of gpt_tokenize
+//
+//   - compare with tokens generated by the huggingface tokenizer
+//   - test cases are chosen based on the model's main language (under 'prompt' directory)
+//   - if all sentences are tokenized identically, print 'All tests passed.'
+//   - otherwise, print sentence, huggingface tokens, ggml tokens
+//
+void test_gpt_tokenizer(gpt_vocab & vocab, const std::string & fpath_test);
+
+// load the tokens from encoder.json
+bool gpt_vocab_init(const std::string & fname, gpt_vocab & vocab);
+
+// sample next token given probabilities for each embedding
+//
+//   - consider only the top K tokens
+//   - from them, consider only the top tokens with cumulative probability > P
+//
+// TODO: not sure if this implementation is correct
+// TODO: temperature is not implemented
+//
+gpt_vocab::id gpt_sample_top_k_top_p(
+        const gpt_vocab & vocab,
+        const float * logits,
+        int    top_k,
+        double top_p,
+        double temp,
+        std::mt19937 & rng);
+
+gpt_vocab::id gpt_sample_top_k_top_p_repeat(
+        const gpt_vocab & vocab,
+        const float * logits,
+        const int32_t * last_n_tokens_data,
+        size_t last_n_tokens_data_size,
+        int    top_k,
+        double top_p,
+        double temp,
+        int repeat_last_n,
+        float repeat_penalty,
+        std::mt19937 & rng);
+
+//
+// Audio utils
+//
+
+// Read WAV audio file and store the PCM data into pcmf32
+// The sample rate of the audio must be equal to COMMON_SAMPLE_RATE
+// If stereo flag is set and the audio has 2 channels, the pcmf32s will contain 2 channel PCM
+bool read_wav(
+        const std::string & fname,
+        std::vector<float> & pcmf32,
+        std::vector<std::vector<float>> & pcmf32s,
+        bool stereo);
+
+// Apply a high-pass frequency filter to PCM audio
+// Suppresses frequencies below cutoff Hz
+void high_pass_filter(
+        std::vector<float> & data,
+        float cutoff,
+        float sample_rate);
+
+// Basic voice activity detection (VAD) using audio energy adaptive threshold
+bool vad_simple(
+        std::vector<float> & pcmf32,
+        int   sample_rate,
+        int   last_ms,
+        float vad_thold,
+        float freq_thold,
+        bool  verbose);
+
+// compute similarity between two strings using Levenshtein distance
+float similarity(const std::string & s0, const std::string & s1);
+
+//
+// SAM argument parsing
+//
+
+struct sam_params {
+    int32_t seed      = -1; // RNG seed
+    int32_t n_threads = std::min(4, (int32_t) std::thread::hardware_concurrency());
+
+    std::string model     = "models/sam-vit-b/ggml-model-f16.bin"; // model path
+    std::string fname_inp = "img.jpg";
+    std::string fname_out = "img.out";
+};
+
+bool sam_params_parse(int argc, char ** argv, sam_params & params);
+
+void sam_print_usage(int argc, char ** argv, const sam_params & params);

ggml/examples/dolly-v2/CMakeLists.txt

@@ -0,0 +1,13 @@
+#
+# dollyv2
+
+set(TEST_TARGET dollyv2)
+add_executable(${TEST_TARGET} main.cpp)
+target_link_libraries(${TEST_TARGET} PRIVATE ggml common common-ggml)
+
+#
+# dollyv2-quantize
+
+set(TEST_TARGET dollyv2-quantize)
+add_executable(${TEST_TARGET} quantize.cpp)
+target_link_libraries(${TEST_TARGET} PRIVATE ggml common common-ggml)

ggml/examples/dolly-v2/README.md

@@ -0,0 +1,187 @@
+# Dolly-V2
+
+Transformer architecture: GPT-NeoX
+
+Modeled from examples/stablelm
+
+Ref: https://github.com/databrickslabs/dolly
+
+Ref: https://github.com/stability-AI/stableLM/#stablelm-alpha
+
+## Usage
+
+```bash
+# get the repo and build it
+git clone https://github.com/ggerganov/ggml
+cd ggml
+mkdir build && cd build
+cmake ..
+make -j
+
+# get the Dolly-V2 3B model
+git clone https://huggingface.co/databricks/dolly-v2-3b
+
+# install Python dependencies
+python3 -m pip install -r ../requirements.txt
+
+# convert model to FP16
+python3 ../examples/dolly-v2/convert-h5-to-ggml.py ./dolly-v2-3b/ 1
+
+# run inference using FP16 precision
+./bin/dollyv2 -m ./dolly-v2-3b/ggml-model-f16.bin -p "State the meaning of life." -t 6 -n 64
+
+main: seed = 1683218142
+dollyv2_model_load: loading model from './dolly-v2-3b/ggml-model-f16.bin' - please wait ...
+dollyv2_model_load: n_vocab = 50280
+dollyv2_model_load: n_ctx   = 2048
+dollyv2_model_load: n_embd  = 2560
+dollyv2_model_load: n_head  = 32
+dollyv2_model_load: n_layer = 32
+dollyv2_model_load: n_rot   = 20
+dollyv2_model_load: ftype   = 1
+dollyv2_model_load: ggml ctx size = 7374.91 MB
+dollyv2_model_load: memory_size =   640.00 MB, n_mem = 65536
+dollyv2_model_load: ................................................ done
+dollyv2_model_load: model size =  5295.10 MB / num tensors = 388
+main: number of tokens in prompt = 32
+main: token[0] =  30003, Below
+main: token[1] =    310,  is
+main: token[2] =    271,  an
+main: token[3] =   9775,  instruction
+main: token[4] =    326,  that
+main: token[5] =   8631,  describes
+main: token[6] =    247,  a
+main: token[7] =   4836,  task
+main: token[8] =    964, .
+main: token[9] =  19566,  Write
+main: token[10] =    247,  a
+main: token[11] =   2380,  response
+main: token[12] =    326,  that
+main: token[13] =  20420,  appropriately
+main: token[14] =  29141,  completes
+main: token[15] =    253,  the
+main: token[16] =   2748,  request
+main: token[17] =    964, .
+main: token[18] =    187, 
+
+main: token[19] =    187, 
+
+main: token[20] =  50278, ### Instruction:
+main: token[21] =    187, 
+
+main: token[22] =   5443, State
+main: token[23] =    253,  the
+main: token[24] =   4495,  meaning
+main: token[25] =    273,  of
+main: token[26] =   1495,  life
+main: token[27] =    964, .
+main: token[28] =    187, 
+
+main: token[29] =    187, 
+
+main: token[30] =  50279, ### Response:
+main: token[31] =    187, 
+
+
+Below is an instruction that describes a task. Write a response that appropriately completes the request.
+
+### Instruction:
+State the meaning of life.
+
+### Response:
+The meaning of life is to love and be loved.
+
+### End
+
+main: mem per token = 16136720 bytes
+main:     load time =  2202.58 ms
+main:   sample time =     2.57 ms
+main:  predict time =  1497.14 ms / 33.27 ms per token
+main:    total time =  6187.27 ms
+```
+
+## 5-bit integer quantization mode
+
+```bash
+# quantize the model to 5-bits using Q5_0 quantization
+./bin/dollyv2-quantize ./dolly-v2-3b/ggml-model-f16.bin ./dolly-v2-3b/ggml-model-q5_0.bin q5_0
+
+# run the quantized model
+./bin/dollyv2 -m ./dolly-v2-3b/ggml-model-q5_0.bin -p "State the meaning of life." -t 6 -n 64
+
+main: seed = 1683218518
+dollyv2_model_load: loading model from './dolly-v2-3b/ggml-model-q5_0.bin' - please wait ...
+dollyv2_model_load: n_vocab = 50280
+dollyv2_model_load: n_ctx   = 2048
+dollyv2_model_load: n_embd  = 2560
+dollyv2_model_load: n_head  = 32
+dollyv2_model_load: n_layer = 32
+dollyv2_model_load: n_rot   = 20
+dollyv2_model_load: ftype   = 8
+dollyv2_model_load: ggml ctx size = 3902.68 MB
+dollyv2_model_load: memory_size =   640.00 MB, n_mem = 65536
+dollyv2_model_load: ................................................ done
+dollyv2_model_load: model size =  1822.87 MB / num tensors = 388
+main: number of tokens in prompt = 32
+main: token[0] =  30003, Below
+main: token[1] =    310,  is
+main: token[2] =    271,  an
+main: token[3] =   9775,  instruction
+main: token[4] =    326,  that
+main: token[5] =   8631,  describes
+main: token[6] =    247,  a
+main: token[7] =   4836,  task
+main: token[8] =    964, .
+main: token[9] =  19566,  Write
+main: token[10] =    247,  a
+main: token[11] =   2380,  response
+main: token[12] =    326,  that
+main: token[13] =  20420,  appropriately
+main: token[14] =  29141,  completes
+main: token[15] =    253,  the
+main: token[16] =   2748,  request
+main: token[17] =    964, .
+main: token[18] =    187, 
+
+main: token[19] =    187, 
+
+main: token[20] =  50278, ### Instruction:
+main: token[21] =    187, 
+
+main: token[22] =   5443, State
+main: token[23] =    253,  the
+main: token[24] =   4495,  meaning
+main: token[25] =    273,  of
+main: token[26] =   1495,  life
+main: token[27] =    964, .
+main: token[28] =    187, 
+
+main: token[29] =    187, 
+
+main: token[30] =  50279, ### Response:
+main: token[31] =    187, 
+
+
+Below is an instruction that describes a task. Write a response that appropriately completes the request.
+
+### Instruction:
+State the meaning of life.
+
+### Response:
+The meaning of life is the discovery of the true self.
+
+### End
+
+main: mem per token = 16127760 bytes
+main:     load time =  1011.09 ms
+main:   sample time =     2.79 ms
+main:  predict time =  1271.62 ms / 27.64 ms per token
+main:    total time =  2802.51 ms
+```
+
+## Notes
+
+- No guarantees for correctness
+- The tokenizer is currently hacked - probably works only for English
+- Non-parallel residual is not supported
+- Contributions and improvements are welcome

ggml/examples/dolly-v2/convert-h5-to-ggml.py

@@ -0,0 +1,116 @@
+import sys
+import struct
+import json
+import numpy as np
+
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+if len(sys.argv) < 3:
+    print("Usage: convert-h5-to-ggml.py dir-model [use-f32]\n")
+    print("  ftype == 0 -> float32")
+    print("  ftype == 1 -> float16")
+    sys.exit(1)
+
+# output in the same directory as the model
+dir_model = sys.argv[1]
+fname_out = sys.argv[1] + "/ggml-model.bin"
+
+with open(dir_model + "/tokenizer.json", "r", encoding="utf-8") as f:
+    encoder = json.load(f)
+
+with open(dir_model + "/config.json", "r", encoding="utf-8") as f:
+    hparams = json.load(f)
+
+# possible data types
+#   ftype == 0 -> float32
+#   ftype == 1 -> float16
+#
+# map from ftype to string
+ftype_str = ["f32", "f16"]
+
+ftype = 1
+if len(sys.argv) > 2:
+    ftype = int(sys.argv[2])
+    if ftype < 0 or ftype > 1:
+        print("Invalid ftype: " + str(ftype))
+        sys.exit(1)
+    fname_out = sys.argv[1] + "/ggml-model-" + ftype_str[ftype] + ".bin"
+
+
+tokenizer = AutoTokenizer.from_pretrained(dir_model)
+model = AutoModelForCausalLM.from_pretrained(dir_model, low_cpu_mem_usage=True)
+#print (model)
+
+#print(tokenizer.encode('I believe the meaning of life is'))
+
+list_vars = model.state_dict()
+for name in list_vars.keys():
+    print(name, list_vars[name].shape, list_vars[name].dtype)
+
+fout = open(fname_out, "wb")
+
+print(hparams)
+
+fout.write(struct.pack("i", 0x67676d6c)) # magic: ggml in hex
+fout.write(struct.pack("i", hparams["vocab_size"]))
+fout.write(struct.pack("i", hparams["max_position_embeddings"]))
+fout.write(struct.pack("i", hparams["hidden_size"]))
+fout.write(struct.pack("i", hparams["num_attention_heads"]))
+fout.write(struct.pack("i", hparams["num_hidden_layers"]))
+fout.write(struct.pack("i", int(hparams["rotary_pct"]*(hparams["hidden_size"]//hparams["num_attention_heads"]))))
+fout.write(struct.pack("i", hparams["use_parallel_residual"]))
+fout.write(struct.pack("i", ftype))
+
+# TODO: temporary hack to not deal with implementing the tokenizer
+dot_token = tokenizer.encode('.')[0]
+for i in range(hparams["vocab_size"]):
+    text = tokenizer.decode([dot_token, i]).encode('utf-8')
+    # remove the first byte (it's always '.')
+    text = text[1:]
+    fout.write(struct.pack("i", len(text)))
+    fout.write(text)
+
+for name in list_vars.keys():
+    data = list_vars[name].squeeze().numpy()
+    print("Processing variable: " + name + " with shape: ", data.shape)
+
+    # we don't need these
+    if name.endswith(".attention.masked_bias") or     \
+       name.endswith(".attention.bias") or \
+       name.endswith(".attention.rotary_emb.inv_freq"):
+        print("  Skipping variable: " + name)
+        continue
+
+    n_dims = len(data.shape);
+
+    # ftype == 0 -> float32, ftype == 1 -> float16
+    ftype_cur = 0;
+    if ftype != 0:
+        if name[-7:] == ".weight" and n_dims == 2:
+            print("  Converting to float16")
+            data = data.astype(np.float16)
+            ftype_cur = 1
+        else:
+            print("  Converting to float32")
+            data = data.astype(np.float32)
+            ftype_cur = 0
+    else:
+        if data.dtype != np.float32:
+            print("  Converting to float32")
+            data = data.astype(np.float32)
+            ftype_cur = 0
+
+    # header
+    str = name.encode('utf-8')
+    fout.write(struct.pack("iii", n_dims, len(str), ftype_cur))
+    for i in range(n_dims):
+        fout.write(struct.pack("i", data.shape[n_dims - 1 - i]))
+    fout.write(str);
+
+    # data
+    data.tofile(fout)
+
+fout.close()
+
+print("Done. Output file: " + fname_out)
+print("")

ggml/examples/dolly-v2/main.cpp

@@ -0,0 +1,962 @@
+#include "ggml/ggml.h"
+
+#include "common.h"
+#include "common-ggml.h"
+
+#include <cassert>
+#include <cmath>
+#include <cstdio>
+#include <cstring>
+#include <cinttypes>
+#include <fstream>
+#include <iostream>
+#include <map>
+#include <string>
+#include <vector>
+
+#if !defined(_WIN32)
+#define DOLLY_INTERACTIVE_PORT
+#endif
+
+#if defined(DOLLY_INTERACTIVE_PORT)
+#include <arpa/inet.h>
+#include <netinet/in.h>
+#include <sys/socket.h>
+#include <unistd.h>
+#endif
+
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
+// default hparams (Dolly-V2 3B)
+struct dollyv2_hparams {
+    int32_t n_vocab = 50254; // tokenizer.vocab_size
+    int32_t n_ctx   = 2048;  // model.config.max_position_embeddings
+    int32_t n_embd  = 2560;  // model.config.hidden_size
+    int32_t n_head  = 32;    // model.config.num_attention_heads
+    int32_t n_layer = 32;    // model.config.num_hidden_layers
+    int32_t n_rot   = 20;    // rotary_pct[25%] * (n_embd / n_head)
+    int32_t par_res = 1; // 1 = true, 0 = false
+    int32_t ftype   = GGML_FTYPE_MOSTLY_F16;
+    float   eps     = 1e-5f;
+};
+
+const std::string INSTRUCTION_KEY = "### Instruction:";
+const std::string RESPONSE_KEY    = "### Response:";
+const std::string END_KEY         = "### End";
+const std::string INTRO_BLURB     = "Below is an instruction that describes a task. Write a response that appropriately completes the request.";
+
+// dollyv2 prompt format
+std::string prompt_for_generation(const std::string& instruction) {
+    return INTRO_BLURB + "\n\n" + INSTRUCTION_KEY + "\n" + instruction + "\n\n" + RESPONSE_KEY + "\n";
+}
+
+struct dollyv2_layer {
+    // pre normalization
+    struct ggml_tensor * ln_1_g;
+    struct ggml_tensor * ln_1_b;
+
+    // attention
+    struct ggml_tensor * c_attn_attn_w;
+    struct ggml_tensor * c_attn_attn_b;
+
+    struct ggml_tensor * c_attn_proj_w;
+    struct ggml_tensor * c_attn_proj_b;
+
+    // post normalization
+    struct ggml_tensor * ln_2_g;
+    struct ggml_tensor * ln_2_b;
+
+    // ff
+    struct ggml_tensor * c_mlp_fc_w;
+    struct ggml_tensor * c_mlp_fc_b;
+
+    struct ggml_tensor * c_mlp_proj_w;
+    struct ggml_tensor * c_mlp_proj_b;
+};
+
+struct dollyv2_model {
+    dollyv2_hparams hparams;
+
+    // normalization
+    struct ggml_tensor * ln_f_g;
+    struct ggml_tensor * ln_f_b;
+
+    struct ggml_tensor * wte; // position embedding
+
+    struct ggml_tensor * lmh_g; // language model head
+    //struct ggml_tensor * lmh_b; // language model bias
+
+    std::vector<dollyv2_layer> layers;
+
+    // key + value memory
+    struct ggml_tensor * memory_k;
+    struct ggml_tensor * memory_v;
+
+    //
+    struct ggml_context * ctx;
+    std::map<std::string, struct ggml_tensor *> tensors;
+};
+
+// load the model's weights from a file
+bool dollyv2_model_load(const std::string & fname, dollyv2_model & model, gpt_vocab & vocab) {
+    printf("%s: loading model from '%s' - please wait ...\n", __func__, fname.c_str());
+
+    auto fin = std::ifstream(fname, std::ios::binary);
+    if (!fin) {
+        fprintf(stderr, "%s: failed to open '%s'\n", __func__, fname.c_str());
+        return false;
+    }
+
+    // verify magic
+    {
+        uint32_t magic;
+        fin.read((char *) &magic, sizeof(magic));
+        if (magic != GGML_FILE_MAGIC) {
+            fprintf(stderr, "%s: invalid model file '%s' (bad magic)\n", __func__, fname.c_str());
+            return false;
+        }
+    }
+
+    // load hparams
+    {
+        auto & hparams = model.hparams;
+
+        fin.read((char *) &hparams.n_vocab, sizeof(hparams.n_vocab));
+        fin.read((char *) &hparams.n_ctx,   sizeof(hparams.n_ctx));
+        fin.read((char *) &hparams.n_embd,  sizeof(hparams.n_embd));
+        fin.read((char *) &hparams.n_head,  sizeof(hparams.n_head));
+        fin.read((char *) &hparams.n_layer, sizeof(hparams.n_layer));
+        fin.read((char *) &hparams.n_rot,   sizeof(hparams.n_rot));
+        fin.read((char *) &hparams.par_res, sizeof(hparams.par_res));
+        fin.read((char *) &hparams.ftype,   sizeof(hparams.ftype));
+
+        const int32_t qntvr = hparams.ftype / GGML_QNT_VERSION_FACTOR;
+
+        printf("%s: n_vocab = %d\n", __func__, hparams.n_vocab);
+        printf("%s: n_ctx   = %d\n", __func__, hparams.n_ctx);
+        printf("%s: n_embd  = %d\n", __func__, hparams.n_embd);
+        printf("%s: n_head  = %d\n", __func__, hparams.n_head);
+        printf("%s: n_layer = %d\n", __func__, hparams.n_layer);
+        printf("%s: n_rot   = %d\n", __func__, hparams.n_rot);
+        printf("%s: par_res = %d\n", __func__, hparams.par_res);
+        printf("%s: ftype   = %d\n", __func__, hparams.ftype);
+        printf("%s: qntvr   = %d\n", __func__, qntvr);
+
+        hparams.ftype %= GGML_QNT_VERSION_FACTOR;
+    }
+
+    // load vocab
+    {
+        const int32_t n_vocab = model.hparams.n_vocab;
+
+        std::string word;
+        std::vector<char> buf(128);
+
+        for (int i = 0; i < n_vocab; i++) {
+            uint32_t len;
+            fin.read((char *) &len, sizeof(len));
+
+            buf.resize(len);
+            fin.read((char *) buf.data(), len);
+            word.assign(buf.data(), len);
+
+            vocab.token_to_id[word] = i;
+            vocab.id_to_token[i] = word;
+        }
+
+        vocab.add_special_token("### End");
+        vocab.add_special_token("### Instruction:");
+        vocab.add_special_token("### Response:");
+    }
+
+    // for the big tensors, we have the option to store the data in 16-bit floats or quantized
+    // in order to save memory and also to speed up the computation
+    ggml_type wtype = ggml_ftype_to_ggml_type((ggml_ftype) (model.hparams.ftype));
+    if (wtype == GGML_TYPE_COUNT) {
+        fprintf(stderr, "%s: invalid model file '%s' (bad ftype value %d)\n",
+                __func__, fname.c_str(), model.hparams.ftype);
+        return false;
+    }
+
+    auto & ctx = model.ctx;
+
+    size_t ctx_size = 0;
+
+    {
+        const auto & hparams = model.hparams;
+
+        const int n_embd  = hparams.n_embd;
+        const int n_layer = hparams.n_layer;
+        const int n_ctx   = hparams.n_ctx;
+        const int n_vocab = hparams.n_vocab;
+
+        ctx_size += n_embd*ggml_type_sizef(GGML_TYPE_F32); // ln_f_g
+        ctx_size += n_embd*ggml_type_sizef(GGML_TYPE_F32); // ln_f_b
+
+        ctx_size += n_embd*n_vocab*ggml_type_sizef(wtype); // wte
+
+        ctx_size += n_embd*n_vocab*ggml_type_sizef(wtype);           // lmh_g
+        //ctx_size +=        n_vocab*ggml_type_sizef(GGML_TYPE_F32); // lmh_b
+
+        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_1_g
+        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_1_b
+
+        ctx_size += n_layer*(3*n_embd*n_embd*ggml_type_sizef(wtype));         // c_attn_attn_w
+        ctx_size += n_layer*(       3*n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_attn_attn_b
+
+        ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype));         // c_attn_proj_w
+        ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_attn_proj_b
+
+        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_2_g
+        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_2_b
+
+        ctx_size += n_layer*(4*n_embd*n_embd*ggml_type_sizef(wtype));         // c_mlp_fc_w
+        ctx_size += n_layer*(       4*n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_mlp_fc_b
+
+        ctx_size += n_layer*(4*n_embd*n_embd*ggml_type_sizef(wtype));         // c_mlp_proj_w
+        ctx_size += n_layer*(         n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_mlp_proj_b
+
+        ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(GGML_TYPE_F32); // memory_k
+        ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(GGML_TYPE_F32); // memory_v
+
+        ctx_size += (6 + 16*n_layer)*512; // object overhead
+
+        printf("%s: ggml ctx size = %6.2f MB\n", __func__, ctx_size/(1024.0*1024.0));
+    }
+
+    // create the ggml context
+    {
+        struct ggml_init_params params = {
+            /*.mem_size   =*/ ctx_size,
+            /*.mem_buffer =*/ NULL,
+            /*.no_alloc   =*/ false,
+        };
+
+        model.ctx = ggml_init(params);
+        if (!model.ctx) {
+            fprintf(stderr, "%s: ggml_init() failed\n", __func__);
+            return false;
+        }
+    }
+
+    // prepare memory for the weights
+    {
+        const auto & hparams = model.hparams;
+
+        const int n_embd  = hparams.n_embd;
+        const int n_layer = hparams.n_layer;
+        const int n_vocab = hparams.n_vocab;
+
+        model.layers.resize(n_layer);
+
+        model.wte    = ggml_new_tensor_2d(ctx, wtype,         n_embd, n_vocab);
+
+        model.ln_f_g = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
+        model.ln_f_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
+
+        model.lmh_g  = ggml_new_tensor_2d(ctx, wtype,         n_embd, n_vocab);
+        //model.lmh_b  = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_vocab);
+
+        // map by name
+        model.tensors["gpt_neox.embed_in.weight"] = model.wte;
+
+        model.tensors["gpt_neox.final_layer_norm.weight"] = model.ln_f_g;
+        model.tensors["gpt_neox.final_layer_norm.bias"]   = model.ln_f_b;
+
+        model.tensors["embed_out.weight"] = model.lmh_g;
+        //model.tensors["lm_head.bias"]   = model.lmh_b;
+
+        for (int i = 0; i < n_layer; ++i) {
+            auto & layer = model.layers[i];
+
+            layer.ln_1_g          = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+            layer.ln_1_b          = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+
+            layer.c_attn_attn_w   = ggml_new_tensor_2d(ctx, wtype,           n_embd, 3*n_embd);
+            layer.c_attn_attn_b   = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 3*n_embd);
+
+            layer.c_attn_proj_w   = ggml_new_tensor_2d(ctx, wtype,           n_embd,   n_embd);
+            layer.c_attn_proj_b   = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+
+            layer.ln_2_g          = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+            layer.ln_2_b          = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+
+            layer.c_mlp_fc_w      = ggml_new_tensor_2d(ctx, wtype,           n_embd, 4*n_embd);
+            layer.c_mlp_fc_b      = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 4*n_embd);
+
+            layer.c_mlp_proj_w    = ggml_new_tensor_2d(ctx, wtype,         4*n_embd,   n_embd);
+            layer.c_mlp_proj_b    = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+
+            // map by name
+
+            // unmapped: attention.rotary_emb, mlp.act
+
+            model.tensors["gpt_neox.layers." + std::to_string(i) + ".input_layernorm.weight"] = layer.ln_1_g;
+            model.tensors["gpt_neox.layers." + std::to_string(i) + ".input_layernorm.bias"]   = layer.ln_1_b;
+
+            model.tensors["gpt_neox.layers." + std::to_string(i) + ".attention.query_key_value.weight"] = layer.c_attn_attn_w;
+            model.tensors["gpt_neox.layers." + std::to_string(i) + ".attention.query_key_value.bias"]   = layer.c_attn_attn_b;
+
+            model.tensors["gpt_neox.layers." + std::to_string(i) + ".attention.dense.weight"] = layer.c_attn_proj_w;
+            model.tensors["gpt_neox.layers." + std::to_string(i) + ".attention.dense.bias"]   = layer.c_attn_proj_b;
+
+            model.tensors["gpt_neox.layers." + std::to_string(i) + ".post_attention_layernorm.weight"] = layer.ln_2_g;
+            model.tensors["gpt_neox.layers." + std::to_string(i) + ".post_attention_layernorm.bias"]   = layer.ln_2_b;
+
+            model.tensors["gpt_neox.layers." + std::to_string(i) + ".mlp.dense_h_to_4h.weight"] = layer.c_mlp_fc_w;
+            model.tensors["gpt_neox.layers." + std::to_string(i) + ".mlp.dense_h_to_4h.bias"]   = layer.c_mlp_fc_b;
+
+            model.tensors["gpt_neox.layers." + std::to_string(i) + ".mlp.dense_4h_to_h.weight"] = layer.c_mlp_proj_w;
+            model.tensors["gpt_neox.layers." + std::to_string(i) + ".mlp.dense_4h_to_h.bias"]   = layer.c_mlp_proj_b;
+        }
+    }
+
+    // key + value memory
+    {
+        const auto & hparams = model.hparams;
+
+        const int n_embd  = hparams.n_embd;
+        const int n_layer = hparams.n_layer;
+        const int n_ctx   = hparams.n_ctx;
+
+        const int64_t n_mem      = n_layer*n_ctx;
+        const int64_t n_elements = n_embd*n_mem;
+
+        model.memory_k = ggml_new_tensor_1d(ctx, GGML_TYPE_F16, n_elements);
+        model.memory_v = ggml_new_tensor_1d(ctx, GGML_TYPE_F16, n_elements);
+
+        const size_t memory_size = ggml_nbytes(model.memory_k) + ggml_nbytes(model.memory_v);
+
+        printf("%s: memory_size = %8.2f MB, n_mem = %" PRId64 "\n", __func__, memory_size/1024.0/1024.0, n_mem);
+    }
+
+    // load weights
+    {
+        int n_tensors = 0;
+        size_t total_size = 0;
+
+        printf("%s: ", __func__);
+
+        while (true) {
+            int32_t n_dims;
+            int32_t length;
+            int32_t ttype;
+
+            fin.read(reinterpret_cast<char *>(&n_dims), sizeof(n_dims));
+            fin.read(reinterpret_cast<char *>(&length), sizeof(length));
+            fin.read(reinterpret_cast<char *>(&ttype),  sizeof(ttype));
+
+            if (fin.eof()) {
+                break;
+            }
+
+            int32_t nelements = 1;
+            int32_t ne[2] = { 1, 1 };
+            for (int i = 0; i < n_dims; ++i) {
+                fin.read(reinterpret_cast<char *>(&ne[i]), sizeof(ne[i]));
+                nelements *= ne[i];
+            }
+
+            std::string name(length, 0);
+            fin.read(&name[0], length);
+
+            if (model.tensors.find(name) == model.tensors.end()) {
+                fprintf(stderr, "%s: unknown tensor '%s' in model file\n", __func__, name.c_str());
+                return false;
+            }
+
+            auto tensor = model.tensors[name];
+            if (ggml_nelements(tensor) != nelements) {
+                fprintf(stderr, "%s: tensor '%s' has wrong size in model file\n", __func__, name.c_str());
+                return false;
+            }
+
+            if (tensor->ne[0] != ne[0] || tensor->ne[1] != ne[1]) {
+                fprintf(stderr, "%s: tensor '%s' has wrong shape in model file: got [%5d, %5d], expected [%5d, %5d]\n",
+                        __func__, name.c_str(), (int) tensor->ne[0], (int) tensor->ne[1], ne[0], ne[1]);
+                return false;
+            }
+
+            // for debugging
+            if (0) {
+                printf("%24s - [%5d, %5d], type = %6s, %6.2f MB, %9zu bytes\n", name.c_str(), ne[0], ne[1], ggml_type_name(ggml_type(ttype)), ggml_nbytes(tensor)/1024.0/1024.0, ggml_nbytes(tensor));
+            }
+
+            const size_t bpe = ggml_type_size(ggml_type(ttype));
+
+            if ((nelements*bpe)/ggml_blck_size(tensor->type) != ggml_nbytes(tensor)) {
+                fprintf(stderr, "%s: tensor '%s' has wrong size in model file: got %zu, expected %zu\n",
+                        __func__, name.c_str(), ggml_nbytes(tensor), nelements*bpe);
+                return false;
+            }
+
+            fin.read(reinterpret_cast<char *>(tensor->data), ggml_nbytes(tensor));
+
+            total_size += ggml_nbytes(tensor);
+            if (++n_tensors % 8 == 0) {
+                printf(".");
+                fflush(stdout);
+            }
+        }
+
+        printf(" done\n");
+
+        printf("%s: model size = %8.2f MB / num tensors = %d\n", __func__, total_size/1024.0/1024.0, n_tensors);
+    }
+
+    fin.close();
+
+    return true;
+}
+
+// feed-forward network
+ggml_tensor * gpt_neox_ff(
+        const dollyv2_layer & layer,
+        ggml_context        * ctx0,
+        ggml_tensor         * inp,
+        float                 eps) {
+    ggml_tensor * cur = ggml_norm(ctx0, inp, eps);
+
+    cur = ggml_add(ctx0,
+        ggml_mul(ctx0,
+            ggml_repeat(ctx0, layer.ln_2_g, cur),
+            cur),
+        ggml_repeat(ctx0, layer.ln_2_b, cur));
+
+    cur = ggml_mul_mat(ctx0,
+            layer.c_mlp_fc_w,
+            cur);
+
+    cur = ggml_add(ctx0,
+            ggml_repeat(ctx0, layer.c_mlp_fc_b, cur),
+            cur);
+
+    // GELU activation
+    cur = ggml_gelu(ctx0, cur);
+
+    // projection
+    // cur = proj_w*cur + proj_b
+    cur = ggml_mul_mat(ctx0,
+            layer.c_mlp_proj_w,
+            cur);
+
+    cur = ggml_add(ctx0,
+            ggml_repeat(ctx0, layer.c_mlp_proj_b, cur),
+            cur);
+    return cur;
+}
+
+// evaluate the transformer
+//
+//   - model:     the model
+//   - n_threads: number of threads to use
+//   - n_past:    the context size so far
+//   - embd_inp:  the embeddings of the tokens in the context
+//   - embd_w:    the predicted logits for the next token
+//
+bool dollyv2_eval(
+        const dollyv2_model & model,
+        const int n_threads,
+        const int n_past,
+        const std::vector<gpt_vocab::id> & embd_inp,
+              std::vector<float>         & embd_w,
+              size_t                     & mem_per_token) {
+    const int N = embd_inp.size();
+
+    const auto & hparams = model.hparams;
+
+    const int n_embd  = hparams.n_embd;
+    const int n_layer = hparams.n_layer;
+    const int n_ctx   = hparams.n_ctx;
+    const int n_head  = hparams.n_head;
+    const int n_vocab = hparams.n_vocab;
+    const int n_rot   = hparams.n_rot;
+
+    static size_t buf_size = 256u*1024*1024;
+    static void * buf = malloc(buf_size);
+
+    if (mem_per_token > 0 && mem_per_token*N > buf_size) {
+        const size_t buf_size_new = 1.1*(mem_per_token*N); // add 10% to account for ggml object overhead
+        //printf("\n%s: reallocating buffer from %zu to %zu bytes\n", __func__, buf_size, buf_size_new);
+
+        // reallocate
+        buf_size = buf_size_new;
+        buf = realloc(buf, buf_size);
+        if (buf == nullptr) {
+            fprintf(stderr, "%s: failed to allocate %zu bytes\n", __func__, buf_size);
+            return false;
+        }
+    }
+
+    struct ggml_init_params params = {
+        /*.mem_size   =*/ buf_size,
+        /*.mem_buffer =*/ buf,
+        /*.no_alloc   =*/ false,
+    };
+
+    struct ggml_context * ctx0 = ggml_init(params);
+    struct ggml_cgraph gf = { };
+
+    struct ggml_tensor * embd = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
+    memcpy(embd->data, embd_inp.data(), N*ggml_element_size(embd));
+
+    // wte
+    struct ggml_tensor * inpL = ggml_get_rows(ctx0, model.wte, embd);
+
+    for (int il = 0; il < n_layer; ++il) {
+        struct ggml_tensor * cur;
+
+        // self-attention
+        {
+            {
+                cur = ggml_norm(ctx0, inpL, hparams.eps);
+
+                cur = ggml_add(ctx0,
+                        ggml_mul(ctx0,
+                            ggml_repeat(ctx0, model.layers[il].ln_1_g, cur),
+                            cur),
+                        ggml_repeat(ctx0, model.layers[il].ln_1_b, cur));
+            }
+
+            // compute QKV
+            {
+                cur = ggml_mul_mat(ctx0,
+                        model.layers[il].c_attn_attn_w,
+                        cur);
+
+                cur = ggml_add(ctx0,
+                        ggml_repeat(ctx0, model.layers[il].c_attn_attn_b, cur),
+                        cur);
+            }
+
+            struct ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_3d(ctx0, cur, n_embd/n_head, n_head, N, cur->nb[1]/n_head, cur->nb[1], 0*sizeof(float)*n_embd/n_head));
+            struct ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_3d(ctx0, cur, n_embd/n_head, n_head, N, cur->nb[1]/n_head, cur->nb[1], 1*sizeof(float)*n_embd/n_head));
+            struct ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_3d(ctx0, cur, n_embd/n_head, n_head, N, cur->nb[1]/n_head, cur->nb[1], 2*sizeof(float)*n_embd/n_head));
+
+            // using mode = 2 for GPT-NeoX mode
+            Qcur = ggml_rope_inplace(ctx0, Qcur, n_past, n_rot, 2, 0);
+            Kcur = ggml_rope_inplace(ctx0, Kcur, n_past, n_rot, 2, 0);
+
+            // store key and value to memory
+            {
+                Vcur = ggml_transpose(ctx0, ggml_reshape_2d(ctx0, Vcur, n_embd, N));
+
+                struct ggml_tensor * k = ggml_view_1d(ctx0, model.memory_k, N*n_embd, (ggml_element_size(model.memory_k)*n_embd)*(il*n_ctx + n_past));
+                struct ggml_tensor * v = ggml_view_2d(ctx0, model.memory_v, N, n_embd,
+                        (   n_ctx)*ggml_element_size(model.memory_v),
+                        (il*n_ctx)*ggml_element_size(model.memory_v)*n_embd + n_past*ggml_element_size(model.memory_v));
+
+                ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Kcur, k));
+                ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Vcur, v));
+            }
+
+            // Q = Qcur.contiguous().view(n_embd/n_head, n_head, N).permute(0, 2, 1, 3)
+            struct ggml_tensor * Q =
+                ggml_permute(ctx0,
+                        Qcur,
+                        0, 2, 1, 3);
+
+            // K = Kmem.view(n_embd/n_head, n_head, n_past + N).permute(0, 2, 1, 3)
+            struct ggml_tensor * K =
+                ggml_permute(ctx0,
+                        ggml_reshape_3d(ctx0,
+                            ggml_view_1d(ctx0, model.memory_k, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(model.memory_k)*n_embd),
+                            n_embd/n_head, n_head, n_past + N),
+                        0, 2, 1, 3);
+
+            // K * Q
+            struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);
+
+            // KQ_scaled = KQ / sqrt(n_embd/n_head)
+            struct ggml_tensor * KQ_scaled =
+                ggml_scale_inplace(ctx0,
+                        KQ,
+                        ggml_new_f32(ctx0, 1.0f/sqrt(float(n_embd)/n_head))
+                        );
+
+            // KQ_masked = mask_past(KQ_scaled)
+            struct ggml_tensor * KQ_masked = ggml_diag_mask_inf_inplace(ctx0, KQ_scaled, n_past);
+
+            // KQ = soft_max(KQ_masked)
+            struct ggml_tensor * KQ_soft_max = ggml_soft_max_inplace(ctx0, KQ_masked);
+
+            // V_trans = Vmem.view(n_embd/n_head, n_head, n_past + N).permute(1, 2, 0, 3).contiguous()
+            struct ggml_tensor * V =
+                ggml_view_3d(ctx0, model.memory_v,
+                        n_past + N, n_embd/n_head, n_head,
+                        n_ctx*ggml_element_size(model.memory_v),
+                        n_ctx*ggml_element_size(model.memory_v)*n_embd/n_head,
+                        il*n_ctx*ggml_element_size(model.memory_v)*n_embd);
+
+            // KQV = transpose(V) * KQ_soft_max
+            struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max);
+
+            // KQV_merged = KQV.permute(0, 2, 1, 3)
+            struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
+
+            // cur = KQV_merged.contiguous().view(n_embd, N)
+            cur = ggml_cpy(ctx0,
+                    KQV_merged,
+                    ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, N));
+
+            // projection
+            {
+                cur = ggml_mul_mat(ctx0,
+                        model.layers[il].c_attn_proj_w,
+                        cur);
+
+                cur = ggml_add(ctx0, ggml_repeat(ctx0, model.layers[il].c_attn_proj_b, cur), cur);
+            }
+        }
+
+        if (hparams.par_res == 0) {
+            struct ggml_tensor * inpFF = ggml_add(ctx0, cur, inpL);
+
+            cur = gpt_neox_ff(model.layers[il], ctx0, inpFF, hparams.eps);
+
+            // input for next layer
+            inpL = ggml_add(ctx0, cur, inpFF);
+        } else {
+            struct ggml_tensor * inpFF = cur;
+
+            // this is independent of the self-attention result, so it could be done in parallel to the self-attention
+            // note here we pass inpL instead of cur
+            cur = gpt_neox_ff(model.layers[il], ctx0, inpL, hparams.eps);
+
+            // layer input + FF
+            cur  = ggml_add(ctx0, cur, inpFF);
+
+            // input for next layer
+            inpL = ggml_add(ctx0, cur, inpL);
+        }
+
+    }
+
+    // norm
+    {
+        inpL = ggml_norm(ctx0, inpL, hparams.eps);
+
+        // inpL = ln_f_g*inpL + ln_f_b
+        inpL = ggml_add(ctx0,
+                ggml_mul(ctx0,
+                    ggml_repeat(ctx0, model.ln_f_g, inpL),
+                    inpL),
+                ggml_repeat(ctx0, model.ln_f_b, inpL));
+    }
+
+    // lm_head
+    {
+        inpL = ggml_mul_mat(ctx0, model.lmh_g, inpL);
+
+        //inpL = ggml_add(ctx0,
+        //        ggml_repeat(ctx0, model.lmh_b, inpL),
+        //        inpL);
+    }
+
+    // logits -> probs
+    //inpL = ggml_soft_max_inplace(ctx0, inpL);
+
+    // run the computation
+    ggml_build_forward_expand(&gf, inpL);
+    ggml_graph_compute_with_ctx(ctx0, &gf, n_threads);
+
+    //if (n_past%100 == 0) {
+    //    ggml_graph_print   (&gf);
+    //    ggml_graph_dump_dot(&gf, NULL, "gpt-2.dot");
+    //}
+
+    //embd_w.resize(n_vocab*N);
+    //memcpy(embd_w.data(), ggml_get_data(inpL), sizeof(float)*n_vocab*N);
+
+    // return result for just the last token
+    embd_w.resize(n_vocab);
+    memcpy(embd_w.data(), (float *) ggml_get_data(inpL) + (n_vocab*(N-1)), sizeof(float)*n_vocab);
+
+    if (mem_per_token == 0) {
+        mem_per_token = ggml_used_mem(ctx0)/N;
+    }
+    //printf("used_mem = %zu\n", ggml_used_mem(ctx0));
+
+    ggml_free(ctx0);
+
+    return true;
+}
+
+std::string execute_prompt(
+        const dollyv2_model &model,
+        gpt_vocab &vocab,
+        const std::string &prompt,
+        gpt_params &params,
+        std::mt19937 &rng,
+        int64_t t_load_us,
+        int64_t t_sample_us,
+        int64_t t_predict_us,
+        size_t mem_per_token,
+        int n_past,
+        bool stream_response_to_cout = false) {
+    std::string output = "";
+    std::vector<float> logits;
+
+    // tokenize the prompt
+    std::vector<gpt_vocab::id> embd_inp = ::gpt_tokenize(vocab, prompt);
+
+    params.n_predict = std::min(params.n_predict, model.hparams.n_ctx - (int)embd_inp.size());
+
+    printf("%s: number of tokens in prompt = %zu\n", __func__, embd_inp.size());
+    for (size_t i = 0; i < embd_inp.size(); i++) {
+        printf("%s: token[%zu] = %6d, %s\n", __func__, i, embd_inp[i], vocab.id_to_token.at(embd_inp[i]).c_str());
+    }
+    printf("\n");
+
+    std::vector<gpt_vocab::id> embd;
+
+    dollyv2_eval(model, params.n_threads, 0, {0, 1, 2, 3}, logits, mem_per_token);
+
+    const int32_t end_token = vocab.token_to_id["### End"];
+
+    for (size_t i = embd.size(); i < embd_inp.size() + params.n_predict; i++) {
+        // predict
+        if (embd.size() > 0) {
+            const int64_t t_start_us = ggml_time_us();
+
+            if (!dollyv2_eval(model, params.n_threads, n_past, embd, logits, mem_per_token)) {
+                printf("Failed to predict\n");
+                return output;
+            }
+
+            t_predict_us += ggml_time_us() - t_start_us;
+        }
+
+        n_past += embd.size();
+        embd.clear();
+
+        if (i >= embd_inp.size()) {
+            // sample next token
+            const int top_k = params.top_k;
+            const float top_p = params.top_p;
+            const float temp = params.temp;
+
+            const int n_vocab = model.hparams.n_vocab;
+
+            gpt_vocab::id id = 0;
+
+            {
+                const int64_t t_start_sample_us = ggml_time_us();
+
+                id = gpt_sample_top_k_top_p(vocab, logits.data() + (logits.size() - n_vocab), top_k, top_p, temp, rng);
+
+                t_sample_us += ggml_time_us() - t_start_sample_us;
+            }
+
+            // add it to the context
+            embd.push_back(id);
+        } else {
+            // if here, it means we are still processing the input prompt
+            for (size_t k = i; k < embd_inp.size(); k++) {
+                embd.push_back(embd_inp[k]);
+                if (int32_t(embd.size()) > params.n_batch) {
+                    break;
+                }
+            }
+            i += embd.size() - 1;
+        }
+
+        // display text
+        for (auto id : embd) {
+            output += vocab.id_to_token[id];
+            if (stream_response_to_cout) {
+                printf("%s", vocab.id_to_token[id].c_str());
+            }
+        }
+        if (stream_response_to_cout) {
+            fflush(stdout);
+        }
+
+        // end of text token
+        if (embd.back() == 0 || (end_token > 0 && embd.back() == end_token)) {
+            return output;
+        }
+    }
+    return output;
+}
+
+#if defined(DOLLY_INTERACTIVE_PORT)
+int setup_port(const int port) {
+    int sockfd = socket(AF_INET, SOCK_STREAM, 0);
+    if (sockfd < 0) {
+        fprintf(stderr, "%s: Failed to create new socket\n", __func__);
+        return -1;
+    }
+
+    sockaddr_in servaddr;
+    std::memset(&servaddr, 0, sizeof(servaddr));
+
+    servaddr.sin_family = AF_INET;
+    servaddr.sin_addr.s_addr = htonl(INADDR_ANY);
+    servaddr.sin_port = htons(port);
+
+    if (bind(sockfd, (struct sockaddr *)&servaddr, sizeof(servaddr)) < 0) {
+        fprintf(stderr, "%s: Failed to bind to port %i\n", __func__, port);
+        return -1;
+    }
+
+    if (listen(sockfd, 10) < 0) {
+        fprintf(stderr, "%s: Failed to listen to socket on port %i\n", __func__, port);
+        return -1;
+    }
+    return sockfd;
+}
+
+std::string read_from_port(int sockfd, int clientfd) {
+    if (clientfd < 0) {
+        fprintf(stderr, "%s: Failed to accept new connection\n", __func__);
+        return "";
+    }
+
+    char buffer[4096];
+    std::memset(buffer, 0, sizeof(buffer));
+
+    if (read(clientfd, buffer, sizeof(buffer)) < 0) {
+        fprintf(stderr, "%s: Failed to read from client\n", __func__);
+    } else {
+        std::cout << "Received: " << buffer;
+        return std::string(buffer);
+    }
+    return std::string("");
+}
+#endif
+
+int main(int argc, char ** argv) {
+    ggml_time_init();
+
+    const int64_t t_main_start_us = ggml_time_us();
+
+    gpt_params params;
+    params.model = "models/dolly-v2-3b/ggml-model-f16.bin";
+
+    if (gpt_params_parse(argc, argv, params) == false) {
+        return 1;
+    }
+
+    if (params.seed < 0) {
+        params.seed = time(NULL);
+    }
+
+    printf("%s: seed = %d\n", __func__, params.seed);
+
+    std::mt19937 rng(params.seed);
+
+    int64_t t_load_us = 0;
+    int64_t t_sample_us = 0;
+    int64_t t_predict_us = 0;
+
+    // determine the required inference memory per token:
+    size_t mem_per_token = 0;
+
+    int n_past = 0;
+
+    gpt_vocab vocab;
+    dollyv2_model model;
+
+    // load the model
+    {
+        const int64_t t_start_us = ggml_time_us();
+
+        if (!dollyv2_model_load(params.model, model, vocab)) {
+            fprintf(stderr, "%s: failed to load model from '%s'\n", __func__, params.model.c_str());
+            return 1;
+        }
+
+        t_load_us = ggml_time_us() - t_start_us;
+
+        test_gpt_tokenizer(vocab, params.token_test);
+    }
+
+#if defined(DOLLY_INTERACTIVE_PORT)
+    int sockfd = -1;
+    if (params.interactive_port != -1) {
+        sockfd = setup_port(params.interactive_port);
+        if (sockfd == -1) {
+            return 1;
+        }
+        fprintf(stdout, "Model is ready on port %i\n", params.interactive_port);
+        fflush(stdout);
+    }
+#endif
+
+    if (params.interactive || params.interactive_port != -1) {
+        while (true) {
+            std::string prompt_input;
+#if defined(DOLLY_INTERACTIVE_PORT)
+            int clientfd = -1;
+            if (params.interactive_port != -1) {
+                sockaddr_in clientaddr;
+                socklen_t clientaddrlen = sizeof(clientaddr);
+                clientfd = accept(sockfd, (struct sockaddr *)&clientaddr, &clientaddrlen);
+                prompt_input = read_from_port(sockfd, clientfd);
+            } else
+#endif
+            {
+                printf("Please enter your quesiton:\n>");
+                fflush(stdout);
+
+                std::getline(std::cin, prompt_input);
+            }
+
+            if (strcmp(prompt_input.c_str(), "exit") == 0) {
+                break;
+            }
+
+            const std::string prompt = prompt_for_generation(prompt_input);
+            // call the model
+            const std::string response = execute_prompt(model, vocab, prompt, params, rng, t_load_us, t_sample_us, t_predict_us, mem_per_token, n_past, true);
+
+#if defined(DOLLY_INTERACTIVE_PORT)
+            if (params.interactive_port != -1) {
+                if (write(clientfd, response.c_str(), response.size()) < 0) {
+                    fprintf(stderr, "%s: Failed to write answer '%s' to client\n", __func__, response.c_str());
+                }
+
+                if (close(clientfd) < 0) {
+                    fprintf(stderr, "%s: Failed to close client socket\n", __func__);
+                }
+            } else
+#endif
+            {
+                printf("%s\n\n", response.c_str());
+            }
+            fflush(stdout);
+        }
+    } else {
+        if (params.prompt.empty()) {
+            params.prompt = gpt_random_prompt(rng);
+        }
+
+        const std::string prompt = prompt_for_generation(params.prompt);
+        execute_prompt(model, vocab, prompt, params, rng, t_load_us, t_sample_us, t_predict_us, mem_per_token, n_past, true);
+    }
+
+    // report timing
+    {
+        const int64_t t_main_end_us = ggml_time_us();
+
+        printf("\n\n");
+        printf("%s: mem per token = %8zu bytes\n", __func__, mem_per_token);
+        printf("%s:     load time = %8.2f ms\n", __func__, t_load_us / 1000.0f);
+        printf("%s:   sample time = %8.2f ms\n", __func__, t_sample_us / 1000.0f);
+        printf("%s:  predict time = %8.2f ms / %.2f ms per token\n", __func__, t_predict_us / 1000.0f, t_predict_us / 1000.0f / n_past);
+        printf("%s:    total time = %8.2f ms\n", __func__, (t_main_end_us - t_main_start_us) / 1000.0f);
+    }
+
+    ggml_free(model.ctx);
+
+#if defined(DOLLY_INTERACTIVE_PORT)
+    if (params.interactive_port != -1 && close(sockfd) < 0) {
+        fprintf(stderr, "%s: Failed to close server socket\n", __func__);
+    }
+#endif
+
+    return 0;
+}

ggml/examples/dolly-v2/quantize.cpp

@@ -0,0 +1,178 @@
+#include "ggml/ggml.h"
+
+#include "common.h"
+#include "common-ggml.h"
+
+#include <cassert>
+#include <cmath>
+#include <cstdio>
+#include <cstring>
+#include <fstream>
+#include <map>
+#include <string>
+#include <vector>
+#include <regex>
+
+// default hparams (dollyv2 3B)
+struct dollyv2_hparams {
+    int32_t n_vocab = 50254; // tokenizer.vocab_size
+    int32_t n_ctx   = 2048;  // model.config.max_position_embeddings
+    int32_t n_embd  = 2560;  // model.config.hidden_size
+    int32_t n_head  = 32;    // model.config.num_attention_heads
+    int32_t n_layer = 32;    // model.config.num_hidden_layers
+    int32_t n_rot   = 20;    // rotary_pct[25%] * (n_embd / n_head)
+    int32_t par_res = 1; // 1 = true, 0 = false
+    int32_t ftype   = GGML_FTYPE_MOSTLY_F16;
+};
+
+// quantize a model
+bool dollyv2_model_quantize(const std::string & fname_inp, const std::string & fname_out, ggml_ftype ftype) {
+    gpt_vocab vocab;
+
+    printf("%s: loading model from '%s'\n", __func__, fname_inp.c_str());
+
+    auto finp = std::ifstream(fname_inp, std::ios::binary);
+    if (!finp) {
+        fprintf(stderr, "%s: failed to open '%s' for reading\n", __func__, fname_inp.c_str());
+        return false;
+    }
+
+    auto fout = std::ofstream(fname_out, std::ios::binary);
+    if (!fout) {
+        fprintf(stderr, "%s: failed to open '%s' for writing\n", __func__, fname_out.c_str());
+        return false;
+    }
+
+    // verify magic
+    {
+        uint32_t magic;
+        finp.read((char *) &magic, sizeof(magic));
+        if (magic != GGML_FILE_MAGIC) {
+            fprintf(stderr, "%s: invalid model file '%s' (bad magic)\n", __func__, fname_inp.c_str());
+            return false;
+        }
+
+        fout.write((char *) &magic, sizeof(magic));
+    }
+
+    dollyv2_hparams hparams;
+
+    // load hparams
+    {
+        finp.read((char *) &hparams.n_vocab, sizeof(hparams.n_vocab));
+        finp.read((char *) &hparams.n_ctx,   sizeof(hparams.n_ctx));
+        finp.read((char *) &hparams.n_embd,  sizeof(hparams.n_embd));
+        finp.read((char *) &hparams.n_head,  sizeof(hparams.n_head));
+        finp.read((char *) &hparams.n_layer, sizeof(hparams.n_layer));
+        finp.read((char *) &hparams.n_rot,   sizeof(hparams.n_rot));
+        finp.read((char *) &hparams.par_res, sizeof(hparams.par_res));
+        finp.read((char *) &hparams.ftype,   sizeof(hparams.ftype));
+
+        const int32_t qntvr_src =    hparams.ftype / GGML_QNT_VERSION_FACTOR;
+        const int32_t ftype_dst = GGML_QNT_VERSION * GGML_QNT_VERSION_FACTOR + ftype;
+
+        printf("%s: n_vocab     = %d\n", __func__, hparams.n_vocab);
+        printf("%s: n_ctx       = %d\n", __func__, hparams.n_ctx);
+        printf("%s: n_embd      = %d\n", __func__, hparams.n_embd);
+        printf("%s: n_head      = %d\n", __func__, hparams.n_head);
+        printf("%s: n_layer     = %d\n", __func__, hparams.n_layer);
+        printf("%s: par_res     = %d\n", __func__, hparams.par_res);
+        printf("%s: ftype (src) = %d\n", __func__, hparams.ftype);
+        printf("%s: qntvr (src) = %d\n", __func__, qntvr_src);
+        printf("%s: ftype (dst) = %d\n", __func__, ftype_dst);
+        printf("%s: qntvr (dst) = %d\n", __func__, GGML_QNT_VERSION);
+
+        fout.write((char *) &hparams.n_vocab, sizeof(hparams.n_vocab));
+        fout.write((char *) &hparams.n_ctx,   sizeof(hparams.n_ctx));
+        fout.write((char *) &hparams.n_embd,  sizeof(hparams.n_embd));
+        fout.write((char *) &hparams.n_head,  sizeof(hparams.n_head));
+        fout.write((char *) &hparams.n_layer, sizeof(hparams.n_layer));
+        fout.write((char *) &hparams.n_rot,   sizeof(hparams.n_rot));
+        fout.write((char *) &hparams.par_res, sizeof(hparams.par_res));
+        fout.write((char *) &ftype_dst,       sizeof(ftype_dst));
+    }
+
+    // load vocab
+    {
+        const int32_t n_vocab = hparams.n_vocab;
+
+        std::string word;
+        for (int i = 0; i < n_vocab; i++) {
+            uint32_t len;
+            finp.read ((char *) &len, sizeof(len));
+            fout.write((char *) &len, sizeof(len));
+
+            word.resize(len);
+            finp.read ((char *) word.data(), len);
+            fout.write((char *) word.data(), len);
+
+            vocab.token_to_id[word] = i;
+            vocab.id_to_token[i] = word;
+        }
+    }
+
+    // regexes of tensor names to be quantized
+    const std::vector<std::string> to_quant = {
+        ".*weight",
+    };
+
+    if (!ggml_common_quantize_0(finp, fout, ftype, to_quant, {})) {
+        fprintf(stderr, "%s: failed to quantize model '%s'\n", __func__, fname_inp.c_str());
+        return false;
+    }
+
+    finp.close();
+    fout.close();
+
+    return true;
+}
+
+// usage:
+//  ./dollyv2-quantize models/dolly-v2-3B/ggml-model.bin models/dolly-v2-3B/ggml-model-quant.bin type
+//
+int main(int argc, char ** argv) {
+    if (argc != 4) {
+        fprintf(stderr, "usage: %s model-f32.bin model-quant.bin type\n", argv[0]);
+        ggml_print_ftypes(stderr);
+        return 1;
+    }
+
+    // needed to initialize f16 tables
+    {
+        struct ggml_init_params params = { 0, NULL, false };
+        struct ggml_context * ctx = ggml_init(params);
+        ggml_free(ctx);
+    }
+
+    const std::string fname_inp = argv[1];
+    const std::string fname_out = argv[2];
+
+    const ggml_ftype ftype = ggml_parse_ftype(argv[3]);
+
+    const int64_t t_main_start_us = ggml_time_us();
+
+    int64_t t_quantize_us = 0;
+
+    // load the model
+    {
+        const int64_t t_start_us = ggml_time_us();
+
+        if (!dollyv2_model_quantize(fname_inp, fname_out, ggml_ftype(ftype))) {
+            fprintf(stderr, "%s: failed to quantize model from '%s'\n", __func__, fname_inp.c_str());
+            return 1;
+        }
+
+        t_quantize_us = ggml_time_us() - t_start_us;
+    }
+
+    // report timing
+    {
+        const int64_t t_main_end_us = ggml_time_us();
+
+        printf("\n");
+        printf("%s: quantize time = %8.2f ms\n", __func__, t_quantize_us/1000.0f);
+        printf("%s:    total time = %8.2f ms\n", __func__, (t_main_end_us - t_main_start_us)/1000.0f);
+    }
+
+    return 0;
+}

ggml/examples/dr_wav.h

@@ -0,0 +1,6434 @@
+/*
+WAV audio loader and writer. Choice of public domain or MIT-0. See license statements at the end of this file.
+dr_wav - v0.12.16 - 2020-12-02
+
+David Reid - mackron@gmail.com
+
+GitHub: https://github.com/mackron/dr_libs
+*/
+
+/*
+RELEASE NOTES - VERSION 0.12
+============================
+Version 0.12 includes breaking changes to custom chunk handling.
+
+
+Changes to Chunk Callback
+-------------------------
+dr_wav supports the ability to fire a callback when a chunk is encounted (except for WAVE and FMT chunks). The callback has been updated to include both the
+container (RIFF or Wave64) and the FMT chunk which contains information about the format of the data in the wave file.
+
+Previously, there was no direct way to determine the container, and therefore no way to discriminate against the different IDs in the chunk header (RIFF and
+Wave64 containers encode chunk ID's differently). The `container` parameter can be used to know which ID to use.
+
+Sometimes it can be useful to know the data format at the time the chunk callback is fired. A pointer to a `drwav_fmt` object is now passed into the chunk
+callback which will give you information about the data format. To determine the sample format, use `drwav_fmt_get_format()`. This will return one of the
+`DR_WAVE_FORMAT_*` tokens.
+*/
+
+/*
+Introduction
+============
+This is a single file library. To use it, do something like the following in one .c file.
+    
+    ```c
+    #define DR_WAV_IMPLEMENTATION
+    #include "dr_wav.h"
+    ```
+
+You can then #include this file in other parts of the program as you would with any other header file. Do something like the following to read audio data:
+
+    ```c
+    drwav wav;
+    if (!drwav_init_file(&wav, "my_song.wav", NULL)) {
+        // Error opening WAV file.
+    }
+
+    drwav_int32* pDecodedInterleavedPCMFrames = malloc(wav.totalPCMFrameCount * wav.channels * sizeof(drwav_int32));
+    size_t numberOfSamplesActuallyDecoded = drwav_read_pcm_frames_s32(&wav, wav.totalPCMFrameCount, pDecodedInterleavedPCMFrames);
+
+    ...
+
+    drwav_uninit(&wav);
+    ```
+
+If you just want to quickly open and read the audio data in a single operation you can do something like this:
+
+    ```c
+    unsigned int channels;
+    unsigned int sampleRate;
+    drwav_uint64 totalPCMFrameCount;
+    float* pSampleData = drwav_open_file_and_read_pcm_frames_f32("my_song.wav", &channels, &sampleRate, &totalPCMFrameCount, NULL);
+    if (pSampleData == NULL) {
+        // Error opening and reading WAV file.
+    }
+
+    ...
+
+    drwav_free(pSampleData);
+    ```
+
+The examples above use versions of the API that convert the audio data to a consistent format (32-bit signed PCM, in this case), but you can still output the
+audio data in its internal format (see notes below for supported formats):
+
+    ```c
+    size_t framesRead = drwav_read_pcm_frames(&wav, wav.totalPCMFrameCount, pDecodedInterleavedPCMFrames);
+    ```
+
+You can also read the raw bytes of audio data, which could be useful if dr_wav does not have native support for a particular data format:
+
+    ```c
+    size_t bytesRead = drwav_read_raw(&wav, bytesToRead, pRawDataBuffer);
+    ```
+
+dr_wav can also be used to output WAV files. This does not currently support compressed formats. To use this, look at `drwav_init_write()`,
+`drwav_init_file_write()`, etc. Use `drwav_write_pcm_frames()` to write samples, or `drwav_write_raw()` to write raw data in the "data" chunk.
+
+    ```c
+    drwav_data_format format;
+    format.container = drwav_container_riff;     // <-- drwav_container_riff = normal WAV files, drwav_container_w64 = Sony Wave64.
+    format.format = DR_WAVE_FORMAT_PCM;          // <-- Any of the DR_WAVE_FORMAT_* codes.
+    format.channels = 2;
+    format.sampleRate = 44100;
+    format.bitsPerSample = 16;
+    drwav_init_file_write(&wav, "data/recording.wav", &format, NULL);
+
+    ...
+
+    drwav_uint64 framesWritten = drwav_write_pcm_frames(pWav, frameCount, pSamples);
+    ```
+
+dr_wav has seamless support the Sony Wave64 format. The decoder will automatically detect it and it should Just Work without any manual intervention.
+
+
+Build Options
+=============
+#define these options before including this file.
+
+#define DR_WAV_NO_CONVERSION_API
+  Disables conversion APIs such as `drwav_read_pcm_frames_f32()` and `drwav_s16_to_f32()`.
+
+#define DR_WAV_NO_STDIO
+  Disables APIs that initialize a decoder from a file such as `drwav_init_file()`, `drwav_init_file_write()`, etc.
+
+
+
+Notes
+=====
+- Samples are always interleaved.
+- The default read function does not do any data conversion. Use `drwav_read_pcm_frames_f32()`, `drwav_read_pcm_frames_s32()` and `drwav_read_pcm_frames_s16()`
+  to read and convert audio data to 32-bit floating point, signed 32-bit integer and signed 16-bit integer samples respectively. Tested and supported internal
+  formats include the following:
+  - Unsigned 8-bit PCM
+  - Signed 12-bit PCM
+  - Signed 16-bit PCM
+  - Signed 24-bit PCM
+  - Signed 32-bit PCM
+  - IEEE 32-bit floating point
+  - IEEE 64-bit floating point
+  - A-law and u-law
+  - Microsoft ADPCM
+  - IMA ADPCM (DVI, format code 0x11)
+- dr_wav will try to read the WAV file as best it can, even if it's not strictly conformant to the WAV format.
+*/
+
+#ifndef dr_wav_h
+#define dr_wav_h
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define DRWAV_STRINGIFY(x)      #x
+#define DRWAV_XSTRINGIFY(x)     DRWAV_STRINGIFY(x)
+
+#define DRWAV_VERSION_MAJOR     0
+#define DRWAV_VERSION_MINOR     12
+#define DRWAV_VERSION_REVISION  16
+#define DRWAV_VERSION_STRING    DRWAV_XSTRINGIFY(DRWAV_VERSION_MAJOR) "." DRWAV_XSTRINGIFY(DRWAV_VERSION_MINOR) "." DRWAV_XSTRINGIFY(DRWAV_VERSION_REVISION)
+
+#include <stddef.h> /* For size_t. */
+
+/* Sized types. */
+typedef   signed char           drwav_int8;
+typedef unsigned char           drwav_uint8;
+typedef   signed short          drwav_int16;
+typedef unsigned short          drwav_uint16;
+typedef   signed int            drwav_int32;
+typedef unsigned int            drwav_uint32;
+#if defined(_MSC_VER)
+    typedef   signed __int64    drwav_int64;
+    typedef unsigned __int64    drwav_uint64;
+#else
+    #if defined(__clang__) || (defined(__GNUC__) && (__GNUC__ > 4 || (__GNUC__ == 4 && __GNUC_MINOR__ >= 6)))
+        #pragma GCC diagnostic push
+        #pragma GCC diagnostic ignored "-Wlong-long"
+        #if defined(__clang__)
+            #pragma GCC diagnostic ignored "-Wc++11-long-long"
+        #endif
+    #endif
+    typedef   signed long long  drwav_int64;
+    typedef unsigned long long  drwav_uint64;
+    #if defined(__clang__) || (defined(__GNUC__) && (__GNUC__ > 4 || (__GNUC__ == 4 && __GNUC_MINOR__ >= 6)))
+        #pragma GCC diagnostic pop
+    #endif
+#endif
+#if defined(__LP64__) || defined(_WIN64) || (defined(__x86_64__) && !defined(__ILP32__)) || defined(_M_X64) || defined(__ia64) || defined (_M_IA64) || defined(__aarch64__) || defined(__powerpc64__)
+    typedef drwav_uint64        drwav_uintptr;
+#else
+    typedef drwav_uint32        drwav_uintptr;
+#endif
+typedef drwav_uint8             drwav_bool8;
+typedef drwav_uint32            drwav_bool32;
+#define DRWAV_TRUE              1
+#define DRWAV_FALSE             0
+
+#if !defined(DRWAV_API)
+    #if defined(DRWAV_DLL)
+        #if defined(_WIN32)
+            #define DRWAV_DLL_IMPORT  __declspec(dllimport)
+            #define DRWAV_DLL_EXPORT  __declspec(dllexport)
+            #define DRWAV_DLL_PRIVATE static
+        #else
+            #if defined(__GNUC__) && __GNUC__ >= 4
+                #define DRWAV_DLL_IMPORT  __attribute__((visibility("default")))
+                #define DRWAV_DLL_EXPORT  __attribute__((visibility("default")))
+                #define DRWAV_DLL_PRIVATE __attribute__((visibility("hidden")))
+            #else
+                #define DRWAV_DLL_IMPORT
+                #define DRWAV_DLL_EXPORT
+                #define DRWAV_DLL_PRIVATE static
+            #endif
+        #endif
+
+        #if defined(DR_WAV_IMPLEMENTATION) || defined(DRWAV_IMPLEMENTATION)
+            #define DRWAV_API  DRWAV_DLL_EXPORT
+        #else
+            #define DRWAV_API  DRWAV_DLL_IMPORT
+        #endif
+        #define DRWAV_PRIVATE DRWAV_DLL_PRIVATE
+    #else
+        #define DRWAV_API extern
+        #define DRWAV_PRIVATE static
+    #endif
+#endif
+
+typedef drwav_int32 drwav_result;
+#define DRWAV_SUCCESS                        0
+#define DRWAV_ERROR                         -1   /* A generic error. */
+#define DRWAV_INVALID_ARGS                  -2
+#define DRWAV_INVALID_OPERATION             -3
+#define DRWAV_OUT_OF_MEMORY                 -4
+#define DRWAV_OUT_OF_RANGE                  -5
+#define DRWAV_ACCESS_DENIED                 -6
+#define DRWAV_DOES_NOT_EXIST                -7
+#define DRWAV_ALREADY_EXISTS                -8
+#define DRWAV_TOO_MANY_OPEN_FILES           -9
+#define DRWAV_INVALID_FILE                  -10
+#define DRWAV_TOO_BIG                       -11
+#define DRWAV_PATH_TOO_LONG                 -12
+#define DRWAV_NAME_TOO_LONG                 -13
+#define DRWAV_NOT_DIRECTORY                 -14
+#define DRWAV_IS_DIRECTORY                  -15
+#define DRWAV_DIRECTORY_NOT_EMPTY           -16
+#define DRWAV_END_OF_FILE                   -17
+#define DRWAV_NO_SPACE                      -18
+#define DRWAV_BUSY                          -19
+#define DRWAV_IO_ERROR                      -20
+#define DRWAV_INTERRUPT                     -21
+#define DRWAV_UNAVAILABLE                   -22
+#define DRWAV_ALREADY_IN_USE                -23
+#define DRWAV_BAD_ADDRESS                   -24
+#define DRWAV_BAD_SEEK                      -25
+#define DRWAV_BAD_PIPE                      -26
+#define DRWAV_DEADLOCK                      -27
+#define DRWAV_TOO_MANY_LINKS                -28
+#define DRWAV_NOT_IMPLEMENTED               -29
+#define DRWAV_NO_MESSAGE                    -30
+#define DRWAV_BAD_MESSAGE                   -31
+#define DRWAV_NO_DATA_AVAILABLE             -32
+#define DRWAV_INVALID_DATA                  -33
+#define DRWAV_TIMEOUT                       -34
+#define DRWAV_NO_NETWORK                    -35
+#define DRWAV_NOT_UNIQUE                    -36
+#define DRWAV_NOT_SOCKET                    -37
+#define DRWAV_NO_ADDRESS                    -38
+#define DRWAV_BAD_PROTOCOL                  -39
+#define DRWAV_PROTOCOL_UNAVAILABLE          -40
+#define DRWAV_PROTOCOL_NOT_SUPPORTED        -41
+#define DRWAV_PROTOCOL_FAMILY_NOT_SUPPORTED -42
+#define DRWAV_ADDRESS_FAMILY_NOT_SUPPORTED  -43
+#define DRWAV_SOCKET_NOT_SUPPORTED          -44
+#define DRWAV_CONNECTION_RESET              -45
+#define DRWAV_ALREADY_CONNECTED             -46
+#define DRWAV_NOT_CONNECTED                 -47
+#define DRWAV_CONNECTION_REFUSED            -48
+#define DRWAV_NO_HOST                       -49
+#define DRWAV_IN_PROGRESS                   -50
+#define DRWAV_CANCELLED                     -51
+#define DRWAV_MEMORY_ALREADY_MAPPED         -52
+#define DRWAV_AT_END                        -53
+
+/* Common data formats. */
+#define DR_WAVE_FORMAT_PCM          0x1
+#define DR_WAVE_FORMAT_ADPCM        0x2
+#define DR_WAVE_FORMAT_IEEE_FLOAT   0x3
+#define DR_WAVE_FORMAT_ALAW         0x6
+#define DR_WAVE_FORMAT_MULAW        0x7
+#define DR_WAVE_FORMAT_DVI_ADPCM    0x11
+#define DR_WAVE_FORMAT_EXTENSIBLE   0xFFFE
+
+/* Constants. */
+#ifndef DRWAV_MAX_SMPL_LOOPS
+#define DRWAV_MAX_SMPL_LOOPS        1
+#endif
+
+/* Flags to pass into drwav_init_ex(), etc. */
+#define DRWAV_SEQUENTIAL            0x00000001
+
+DRWAV_API void drwav_version(drwav_uint32* pMajor, drwav_uint32* pMinor, drwav_uint32* pRevision);
+DRWAV_API const char* drwav_version_string(void);
+
+typedef enum
+{
+    drwav_seek_origin_start,
+    drwav_seek_origin_current
+} drwav_seek_origin;
+
+typedef enum
+{
+    drwav_container_riff,
+    drwav_container_w64,
+    drwav_container_rf64
+} drwav_container;
+
+typedef struct
+{
+    union
+    {
+        drwav_uint8 fourcc[4];
+        drwav_uint8 guid[16];
+    } id;
+
+    /* The size in bytes of the chunk. */
+    drwav_uint64 sizeInBytes;
+
+    /*
+    RIFF = 2 byte alignment.
+    W64  = 8 byte alignment.
+    */
+    unsigned int paddingSize;
+} drwav_chunk_header;
+
+typedef struct
+{
+    /*
+    The format tag exactly as specified in the wave file's "fmt" chunk. This can be used by applications
+    that require support for data formats not natively supported by dr_wav.
+    */
+    drwav_uint16 formatTag;
+
+    /* The number of channels making up the audio data. When this is set to 1 it is mono, 2 is stereo, etc. */
+    drwav_uint16 channels;
+
+    /* The sample rate. Usually set to something like 44100. */
+    drwav_uint32 sampleRate;
+
+    /* Average bytes per second. You probably don't need this, but it's left here for informational purposes. */
+    drwav_uint32 avgBytesPerSec;
+
+    /* Block align. This is equal to the number of channels * bytes per sample. */
+    drwav_uint16 blockAlign;
+
+    /* Bits per sample. */
+    drwav_uint16 bitsPerSample;
+
+    /* The size of the extended data. Only used internally for validation, but left here for informational purposes. */
+    drwav_uint16 extendedSize;
+
+    /*
+    The number of valid bits per sample. When <formatTag> is equal to WAVE_FORMAT_EXTENSIBLE, <bitsPerSample>
+    is always rounded up to the nearest multiple of 8. This variable contains information about exactly how
+    many bits are valid per sample. Mainly used for informational purposes.
+    */
+    drwav_uint16 validBitsPerSample;
+
+    /* The channel mask. Not used at the moment. */
+    drwav_uint32 channelMask;
+
+    /* The sub-format, exactly as specified by the wave file. */
+    drwav_uint8 subFormat[16];
+} drwav_fmt;
+
+DRWAV_API drwav_uint16 drwav_fmt_get_format(const drwav_fmt* pFMT);
+
+
+/*
+Callback for when data is read. Return value is the number of bytes actually read.
+
+pUserData   [in]  The user data that was passed to drwav_init() and family.
+pBufferOut  [out] The output buffer.
+bytesToRead [in]  The number of bytes to read.
+
+Returns the number of bytes actually read.
+
+A return value of less than bytesToRead indicates the end of the stream. Do _not_ return from this callback until
+either the entire bytesToRead is filled or you have reached the end of the stream.
+*/
+typedef size_t (* drwav_read_proc)(void* pUserData, void* pBufferOut, size_t bytesToRead);
+
+/*
+Callback for when data is written. Returns value is the number of bytes actually written.
+
+pUserData    [in]  The user data that was passed to drwav_init_write() and family.
+pData        [out] A pointer to the data to write.
+bytesToWrite [in]  The number of bytes to write.
+
+Returns the number of bytes actually written.
+
+If the return value differs from bytesToWrite, it indicates an error.
+*/
+typedef size_t (* drwav_write_proc)(void* pUserData, const void* pData, size_t bytesToWrite);
+
+/*
+Callback for when data needs to be seeked.
+
+pUserData [in] The user data that was passed to drwav_init() and family.
+offset    [in] The number of bytes to move, relative to the origin. Will never be negative.
+origin    [in] The origin of the seek - the current position or the start of the stream.
+
+Returns whether or not the seek was successful.
+
+Whether or not it is relative to the beginning or current position is determined by the "origin" parameter which will be either drwav_seek_origin_start or
+drwav_seek_origin_current.
+*/
+typedef drwav_bool32 (* drwav_seek_proc)(void* pUserData, int offset, drwav_seek_origin origin);
+
+/*
+Callback for when drwav_init_ex() finds a chunk.
+
+pChunkUserData    [in] The user data that was passed to the pChunkUserData parameter of drwav_init_ex() and family.
+onRead            [in] A pointer to the function to call when reading.
+onSeek            [in] A pointer to the function to call when seeking.
+pReadSeekUserData [in] The user data that was passed to the pReadSeekUserData parameter of drwav_init_ex() and family.
+pChunkHeader      [in] A pointer to an object containing basic header information about the chunk. Use this to identify the chunk.
+container         [in] Whether or not the WAV file is a RIFF or Wave64 container. If you're unsure of the difference, assume RIFF.
+pFMT              [in] A pointer to the object containing the contents of the "fmt" chunk.
+
+Returns the number of bytes read + seeked.
+
+To read data from the chunk, call onRead(), passing in pReadSeekUserData as the first parameter. Do the same for seeking with onSeek(). The return value must
+be the total number of bytes you have read _plus_ seeked.
+
+Use the `container` argument to discriminate the fields in `pChunkHeader->id`. If the container is `drwav_container_riff` or `drwav_container_rf64` you should
+use `id.fourcc`, otherwise you should use `id.guid`.
+
+The `pFMT` parameter can be used to determine the data format of the wave file. Use `drwav_fmt_get_format()` to get the sample format, which will be one of the
+`DR_WAVE_FORMAT_*` identifiers. 
+
+The read pointer will be sitting on the first byte after the chunk's header. You must not attempt to read beyond the boundary of the chunk.
+*/
+typedef drwav_uint64 (* drwav_chunk_proc)(void* pChunkUserData, drwav_read_proc onRead, drwav_seek_proc onSeek, void* pReadSeekUserData, const drwav_chunk_header* pChunkHeader, drwav_container container, const drwav_fmt* pFMT);
+
+typedef struct
+{
+    void* pUserData;
+    void* (* onMalloc)(size_t sz, void* pUserData);
+    void* (* onRealloc)(void* p, size_t sz, void* pUserData);
+    void  (* onFree)(void* p, void* pUserData);
+} drwav_allocation_callbacks;
+
+/* Structure for internal use. Only used for loaders opened with drwav_init_memory(). */
+typedef struct
+{
+    const drwav_uint8* data;
+    size_t dataSize;
+    size_t currentReadPos;
+} drwav__memory_stream;
+
+/* Structure for internal use. Only used for writers opened with drwav_init_memory_write(). */
+typedef struct
+{
+    void** ppData;
+    size_t* pDataSize;
+    size_t dataSize;
+    size_t dataCapacity;
+    size_t currentWritePos;
+} drwav__memory_stream_write;
+
+typedef struct
+{
+    drwav_container container;  /* RIFF, W64. */
+    drwav_uint32 format;        /* DR_WAVE_FORMAT_* */
+    drwav_uint32 channels;
+    drwav_uint32 sampleRate;
+    drwav_uint32 bitsPerSample;
+} drwav_data_format;
+
+
+/* See the following for details on the 'smpl' chunk: https://sites.google.com/site/musicgapi/technical-documents/wav-file-format#smpl */
+typedef struct
+{
+    drwav_uint32 cuePointId;
+    drwav_uint32 type;
+    drwav_uint32 start;
+    drwav_uint32 end;
+    drwav_uint32 fraction;
+    drwav_uint32 playCount;
+} drwav_smpl_loop;
+
+ typedef struct
+{
+    drwav_uint32 manufacturer;
+    drwav_uint32 product;
+    drwav_uint32 samplePeriod;
+    drwav_uint32 midiUnityNotes;
+    drwav_uint32 midiPitchFraction;
+    drwav_uint32 smpteFormat;
+    drwav_uint32 smpteOffset;
+    drwav_uint32 numSampleLoops;
+    drwav_uint32 samplerData;
+    drwav_smpl_loop loops[DRWAV_MAX_SMPL_LOOPS];
+} drwav_smpl;
+
+typedef struct
+{
+    /* A pointer to the function to call when more data is needed. */
+    drwav_read_proc onRead;
+
+    /* A pointer to the function to call when data needs to be written. Only used when the drwav object is opened in write mode. */
+    drwav_write_proc onWrite;
+
+    /* A pointer to the function to call when the wav file needs to be seeked. */
+    drwav_seek_proc onSeek;
+
+    /* The user data to pass to callbacks. */
+    void* pUserData;
+
+    /* Allocation callbacks. */
+    drwav_allocation_callbacks allocationCallbacks;
+
+
+    /* Whether or not the WAV file is formatted as a standard RIFF file or W64. */
+    drwav_container container;
+
+
+    /* Structure containing format information exactly as specified by the wav file. */
+    drwav_fmt fmt;
+
+    /* The sample rate. Will be set to something like 44100. */
+    drwav_uint32 sampleRate;
+
+    /* The number of channels. This will be set to 1 for monaural streams, 2 for stereo, etc. */
+    drwav_uint16 channels;
+
+    /* The bits per sample. Will be set to something like 16, 24, etc. */
+    drwav_uint16 bitsPerSample;
+
+    /* Equal to fmt.formatTag, or the value specified by fmt.subFormat if fmt.formatTag is equal to 65534 (WAVE_FORMAT_EXTENSIBLE). */
+    drwav_uint16 translatedFormatTag;
+
+    /* The total number of PCM frames making up the audio data. */
+    drwav_uint64 totalPCMFrameCount;
+
+
+    /* The size in bytes of the data chunk. */
+    drwav_uint64 dataChunkDataSize;
+    
+    /* The position in the stream of the first byte of the data chunk. This is used for seeking. */
+    drwav_uint64 dataChunkDataPos;
+
+    /* The number of bytes remaining in the data chunk. */
+    drwav_uint64 bytesRemaining;
+
+
+    /*
+    Only used in sequential write mode. Keeps track of the desired size of the "data" chunk at the point of initialization time. Always
+    set to 0 for non-sequential writes and when the drwav object is opened in read mode. Used for validation.
+    */
+    drwav_uint64 dataChunkDataSizeTargetWrite;
+
+    /* Keeps track of whether or not the wav writer was initialized in sequential mode. */
+    drwav_bool32 isSequentialWrite;
+
+
+    /* smpl chunk. */
+    drwav_smpl smpl;
+
+
+    /* A hack to avoid a DRWAV_MALLOC() when opening a decoder with drwav_init_memory(). */
+    drwav__memory_stream memoryStream;
+    drwav__memory_stream_write memoryStreamWrite;
+
+    /* Generic data for compressed formats. This data is shared across all block-compressed formats. */
+    struct
+    {
+        drwav_uint64 iCurrentPCMFrame;  /* The index of the next PCM frame that will be read by drwav_read_*(). This is used with "totalPCMFrameCount" to ensure we don't read excess samples at the end of the last block. */
+    } compressed;
+    
+    /* Microsoft ADPCM specific data. */
+    struct
+    {
+        drwav_uint32 bytesRemainingInBlock;
+        drwav_uint16 predictor[2];
+        drwav_int32  delta[2];
+        drwav_int32  cachedFrames[4];  /* Samples are stored in this cache during decoding. */
+        drwav_uint32 cachedFrameCount;
+        drwav_int32  prevFrames[2][2]; /* The previous 2 samples for each channel (2 channels at most). */
+    } msadpcm;
+
+    /* IMA ADPCM specific data. */
+    struct
+    {
+        drwav_uint32 bytesRemainingInBlock;
+        drwav_int32  predictor[2];
+        drwav_int32  stepIndex[2];
+        drwav_int32  cachedFrames[16]; /* Samples are stored in this cache during decoding. */
+        drwav_uint32 cachedFrameCount;
+    } ima;
+} drwav;
+
+
+/*
+Initializes a pre-allocated drwav object for reading.
+
+pWav                         [out]          A pointer to the drwav object being initialized.
+onRead                       [in]           The function to call when data needs to be read from the client.
+onSeek                       [in]           The function to call when the read position of the client data needs to move.
+onChunk                      [in, optional] The function to call when a chunk is enumerated at initialized time.
+pUserData, pReadSeekUserData [in, optional] A pointer to application defined data that will be passed to onRead and onSeek.
+pChunkUserData               [in, optional] A pointer to application defined data that will be passed to onChunk.
+flags                        [in, optional] A set of flags for controlling how things are loaded.
+
+Returns true if successful; false otherwise.
+
+Close the loader with drwav_uninit().
+
+This is the lowest level function for initializing a WAV file. You can also use drwav_init_file() and drwav_init_memory()
+to open the stream from a file or from a block of memory respectively.
+
+Possible values for flags:
+  DRWAV_SEQUENTIAL: Never perform a backwards seek while loading. This disables the chunk callback and will cause this function
+                    to return as soon as the data chunk is found. Any chunks after the data chunk will be ignored.
+
+drwav_init() is equivalent to "drwav_init_ex(pWav, onRead, onSeek, NULL, pUserData, NULL, 0);".
+
+The onChunk callback is not called for the WAVE or FMT chunks. The contents of the FMT chunk can be read from pWav->fmt
+after the function returns.
+
+See also: drwav_init_file(), drwav_init_memory(), drwav_uninit()
+*/
+DRWAV_API drwav_bool32 drwav_init(drwav* pWav, drwav_read_proc onRead, drwav_seek_proc onSeek, void* pUserData, const drwav_allocation_callbacks* pAllocationCallbacks);
+DRWAV_API drwav_bool32 drwav_init_ex(drwav* pWav, drwav_read_proc onRead, drwav_seek_proc onSeek, drwav_chunk_proc onChunk, void* pReadSeekUserData, void* pChunkUserData, drwav_uint32 flags, const drwav_allocation_callbacks* pAllocationCallbacks);
+
+/*
+Initializes a pre-allocated drwav object for writing.
+
+onWrite   [in]           The function to call when data needs to be written.
+onSeek    [in]           The function to call when the write position needs to move.
+pUserData [in, optional] A pointer to application defined data that will be passed to onWrite and onSeek.
+
+Returns true if successful; false otherwise.
+
+Close the writer with drwav_uninit().
+
+This is the lowest level function for initializing a WAV file. You can also use drwav_init_file_write() and drwav_init_memory_write()
+to open the stream from a file or from a block of memory respectively.
+
+If the total sample count is known, you can use drwav_init_write_sequential(). This avoids the need for dr_wav to perform
+a post-processing step for storing the total sample count and the size of the data chunk which requires a backwards seek.
+
+See also: drwav_init_file_write(), drwav_init_memory_write(), drwav_uninit()
+*/
+DRWAV_API drwav_bool32 drwav_init_write(drwav* pWav, const drwav_data_format* pFormat, drwav_write_proc onWrite, drwav_seek_proc onSeek, void* pUserData, const drwav_allocation_callbacks* pAllocationCallbacks);
+DRWAV_API drwav_bool32 drwav_init_write_sequential(drwav* pWav, const drwav_data_format* pFormat, drwav_uint64 totalSampleCount, drwav_write_proc onWrite, void* pUserData, const drwav_allocation_callbacks* pAllocationCallbacks);
+DRWAV_API drwav_bool32 drwav_init_write_sequential_pcm_frames(drwav* pWav, const drwav_data_format* pFormat, drwav_uint64 totalPCMFrameCount, drwav_write_proc onWrite, void* pUserData, const drwav_allocation_callbacks* pAllocationCallbacks);
+
+/*
+Utility function to determine the target size of the entire data to be written (including all headers and chunks).
+
+Returns the target size in bytes.
+
+Useful if the application needs to know the size to allocate.
+
+Only writing to the RIFF chunk and one data chunk is currently supported.
+
+See also: drwav_init_write(), drwav_init_file_write(), drwav_init_memory_write()
+*/
+DRWAV_API drwav_uint64 drwav_target_write_size_bytes(const drwav_data_format* pFormat, drwav_uint64 totalSampleCount);
+
+/*
+Uninitializes the given drwav object.
+
+Use this only for objects initialized with drwav_init*() functions (drwav_init(), drwav_init_ex(), drwav_init_write(), drwav_init_write_sequential()).
+*/
+DRWAV_API drwav_result drwav_uninit(drwav* pWav);
+
+
+/*
+Reads raw audio data.
+
+This is the lowest level function for reading audio data. It simply reads the given number of
+bytes of the raw internal sample data.
+
+Consider using drwav_read_pcm_frames_s16(), drwav_read_pcm_frames_s32() or drwav_read_pcm_frames_f32() for
+reading sample data in a consistent format.
+
+pBufferOut can be NULL in which case a seek will be performed.
+
+Returns the number of bytes actually read.
+*/
+DRWAV_API size_t drwav_read_raw(drwav* pWav, size_t bytesToRead, void* pBufferOut);
+
+/*
+Reads up to the specified number of PCM frames from the WAV file.
+
+The output data will be in the file's internal format, converted to native-endian byte order. Use
+drwav_read_pcm_frames_s16/f32/s32() to read data in a specific format.
+
+If the return value is less than <framesToRead> it means the end of the file has been reached or
+you have requested more PCM frames than can possibly fit in the output buffer.
+
+This function will only work when sample data is of a fixed size and uncompressed. If you are
+using a compressed format consider using drwav_read_raw() or drwav_read_pcm_frames_s16/s32/f32().
+
+pBufferOut can be NULL in which case a seek will be performed.
+*/
+DRWAV_API drwav_uint64 drwav_read_pcm_frames(drwav* pWav, drwav_uint64 framesToRead, void* pBufferOut);
+DRWAV_API drwav_uint64 drwav_read_pcm_frames_le(drwav* pWav, drwav_uint64 framesToRead, void* pBufferOut);
+DRWAV_API drwav_uint64 drwav_read_pcm_frames_be(drwav* pWav, drwav_uint64 framesToRead, void* pBufferOut);
+
+/*
+Seeks to the given PCM frame.
+
+Returns true if successful; false otherwise.
+*/
+DRWAV_API drwav_bool32 drwav_seek_to_pcm_frame(drwav* pWav, drwav_uint64 targetFrameIndex);
+
+
+/*
+Writes raw audio data.
+
+Returns the number of bytes actually written. If this differs from bytesToWrite, it indicates an error.
+*/
+DRWAV_API size_t drwav_write_raw(drwav* pWav, size_t bytesToWrite, const void* pData);
+
+/*
+Writes PCM frames.
+
+Returns the number of PCM frames written.
+
+Input samples need to be in native-endian byte order. On big-endian architectures the input data will be converted to
+little-endian. Use drwav_write_raw() to write raw audio data without performing any conversion.
+*/
+DRWAV_API drwav_uint64 drwav_write_pcm_frames(drwav* pWav, drwav_uint64 framesToWrite, const void* pData);
+DRWAV_API drwav_uint64 drwav_write_pcm_frames_le(drwav* pWav, drwav_uint64 framesToWrite, const void* pData);
+DRWAV_API drwav_uint64 drwav_write_pcm_frames_be(drwav* pWav, drwav_uint64 framesToWrite, const void* pData);
+
+
+/* Conversion Utilities */
+#ifndef DR_WAV_NO_CONVERSION_API
+
+/*
+Reads a chunk of audio data and converts it to signed 16-bit PCM samples.
+
+pBufferOut can be NULL in which case a seek will be performed.
+
+Returns the number of PCM frames actually read.
+
+If the return value is less than <framesToRead> it means the end of the file has been reached.
+*/
+DRWAV_API drwav_uint64 drwav_read_pcm_frames_s16(drwav* pWav, drwav_uint64 framesToRead, drwav_int16* pBufferOut);
+DRWAV_API drwav_uint64 drwav_read_pcm_frames_s16le(drwav* pWav, drwav_uint64 framesToRead, drwav_int16* pBufferOut);
+DRWAV_API drwav_uint64 drwav_read_pcm_frames_s16be(drwav* pWav, drwav_uint64 framesToRead, drwav_int16* pBufferOut);
+
+/* Low-level function for converting unsigned 8-bit PCM samples to signed 16-bit PCM samples. */
+DRWAV_API void drwav_u8_to_s16(drwav_int16* pOut, const drwav_uint8* pIn, size_t sampleCount);
+
+/* Low-level function for converting signed 24-bit PCM samples to signed 16-bit PCM samples. */
+DRWAV_API void drwav_s24_to_s16(drwav_int16* pOut, const drwav_uint8* pIn, size_t sampleCount);
+
+/* Low-level function for converting signed 32-bit PCM samples to signed 16-bit PCM samples. */
+DRWAV_API void drwav_s32_to_s16(drwav_int16* pOut, const drwav_int32* pIn, size_t sampleCount);
+
+/* Low-level function for converting IEEE 32-bit floating point samples to signed 16-bit PCM samples. */
+DRWAV_API void drwav_f32_to_s16(drwav_int16* pOut, const float* pIn, size_t sampleCount);
+
+/* Low-level function for converting IEEE 64-bit floating point samples to signed 16-bit PCM samples. */
+DRWAV_API void drwav_f64_to_s16(drwav_int16* pOut, const double* pIn, size_t sampleCount);
+
+/* Low-level function for converting A-law samples to signed 16-bit PCM samples. */
+DRWAV_API void drwav_alaw_to_s16(drwav_int16* pOut, const drwav_uint8* pIn, size_t sampleCount);
+
+/* Low-level function for converting u-law samples to signed 16-bit PCM samples. */
+DRWAV_API void drwav_mulaw_to_s16(drwav_int16* pOut, const drwav_uint8* pIn, size_t sampleCount);
+
+
+/*
+Reads a chunk of audio data and converts it to IEEE 32-bit floating point samples.
+
+pBufferOut can be NULL in which case a seek will be performed.
+
+Returns the number of PCM frames actually read.
+
+If the return value is less than <framesToRead> it means the end of the file has been reached.
+*/
+DRWAV_API drwav_uint64 drwav_read_pcm_frames_f32(drwav* pWav, drwav_uint64 framesToRead, float* pBufferOut);
+DRWAV_API drwav_uint64 drwav_read_pcm_frames_f32le(drwav* pWav, drwav_uint64 framesToRead, float* pBufferOut);
+DRWAV_API drwav_uint64 drwav_read_pcm_frames_f32be(drwav* pWav, drwav_uint64 framesToRead, float* pBufferOut);
+
+/* Low-level function for converting unsigned 8-bit PCM samples to IEEE 32-bit floating point samples. */
+DRWAV_API void drwav_u8_to_f32(float* pOut, const drwav_uint8* pIn, size_t sampleCount);
+
+/* Low-level function for converting signed 16-bit PCM samples to IEEE 32-bit floating point samples. */
+DRWAV_API void drwav_s16_to_f32(float* pOut, const drwav_int16* pIn, size_t sampleCount);
+
+/* Low-level function for converting signed 24-bit PCM samples to IEEE 32-bit floating point samples. */
+DRWAV_API void drwav_s24_to_f32(float* pOut, const drwav_uint8* pIn, size_t sampleCount);
+
+/* Low-level function for converting signed 32-bit PCM samples to IEEE 32-bit floating point samples. */
+DRWAV_API void drwav_s32_to_f32(float* pOut, const drwav_int32* pIn, size_t sampleCount);
+
+/* Low-level function for converting IEEE 64-bit floating point samples to IEEE 32-bit floating point samples. */
+DRWAV_API void drwav_f64_to_f32(float* pOut, const double* pIn, size_t sampleCount);
+
+/* Low-level function for converting A-law samples to IEEE 32-bit floating point samples. */
+DRWAV_API void drwav_alaw_to_f32(float* pOut, const drwav_uint8* pIn, size_t sampleCount);
+
+/* Low-level function for converting u-law samples to IEEE 32-bit floating point samples. */
+DRWAV_API void drwav_mulaw_to_f32(float* pOut, const drwav_uint8* pIn, size_t sampleCount);
+
+
+/*
+Reads a chunk of audio data and converts it to signed 32-bit PCM samples.
+
+pBufferOut can be NULL in which case a seek will be performed.
+
+Returns the number of PCM frames actually read.
+
+If the return value is less than <framesToRead> it means the end of the file has been reached.
+*/
+DRWAV_API drwav_uint64 drwav_read_pcm_frames_s32(drwav* pWav, drwav_uint64 framesToRead, drwav_int32* pBufferOut);
+DRWAV_API drwav_uint64 drwav_read_pcm_frames_s32le(drwav* pWav, drwav_uint64 framesToRead, drwav_int32* pBufferOut);
+DRWAV_API drwav_uint64 drwav_read_pcm_frames_s32be(drwav* pWav, drwav_uint64 framesToRead, drwav_int32* pBufferOut);
+
+/* Low-level function for converting unsigned 8-bit PCM samples to signed 32-bit PCM samples. */
+DRWAV_API void drwav_u8_to_s32(drwav_int32* pOut, const drwav_uint8* pIn, size_t sampleCount);
+
+/* Low-level function for converting signed 16-bit PCM samples to signed 32-bit PCM samples. */
+DRWAV_API void drwav_s16_to_s32(drwav_int32* pOut, const drwav_int16* pIn, size_t sampleCount);
+
+/* Low-level function for converting signed 24-bit PCM samples to signed 32-bit PCM samples. */
+DRWAV_API void drwav_s24_to_s32(drwav_int32* pOut, const drwav_uint8* pIn, size_t sampleCount);
+
+/* Low-level function for converting IEEE 32-bit floating point samples to signed 32-bit PCM samples. */
+DRWAV_API void drwav_f32_to_s32(drwav_int32* pOut, const float* pIn, size_t sampleCount);
+
+/* Low-level function for converting IEEE 64-bit floating point samples to signed 32-bit PCM samples. */
+DRWAV_API void drwav_f64_to_s32(drwav_int32* pOut, const double* pIn, size_t sampleCount);
+
+/* Low-level function for converting A-law samples to signed 32-bit PCM samples. */
+DRWAV_API void drwav_alaw_to_s32(drwav_int32* pOut, const drwav_uint8* pIn, size_t sampleCount);
+
+/* Low-level function for converting u-law samples to signed 32-bit PCM samples. */
+DRWAV_API void drwav_mulaw_to_s32(drwav_int32* pOut, const drwav_uint8* pIn, size_t sampleCount);
+
+#endif  /* DR_WAV_NO_CONVERSION_API */
+
+
+/* High-Level Convenience Helpers */
+
+#ifndef DR_WAV_NO_STDIO
+/*
+Helper for initializing a wave file for reading using stdio.
+
+This holds the internal FILE object until drwav_uninit() is called. Keep this in mind if you're caching drwav
+objects because the operating system may restrict the number of file handles an application can have open at
+any given time.
+*/
+DRWAV_API drwav_bool32 drwav_init_file(drwav* pWav, const char* filename, const drwav_allocation_callbacks* pAllocationCallbacks);
+DRWAV_API drwav_bool32 drwav_init_file_ex(drwav* pWav, const char* filename, drwav_chunk_proc onChunk, void* pChunkUserData, drwav_uint32 flags, const drwav_allocation_callbacks* pAllocationCallbacks);
+DRWAV_API drwav_bool32 drwav_init_file_w(drwav* pWav, const wchar_t* filename, const drwav_allocation_callbacks* pAllocationCallbacks);
+DRWAV_API drwav_bool32 drwav_init_file_ex_w(drwav* pWav, const wchar_t* filename, drwav_chunk_proc onChunk, void* pChunkUserData, drwav_uint32 flags, const drwav_allocation_callbacks* pAllocationCallbacks);
+
+/*
+Helper for initializing a wave file for writing using stdio.
+
+This holds the internal FILE object until drwav_uninit() is called. Keep this in mind if you're caching drwav
+objects because the operating system may restrict the number of file handles an application can have open at
+any given time.
+*/
+DRWAV_API drwav_bool32 drwav_init_file_write(drwav* pWav, const char* filename, const drwav_data_format* pFormat, const drwav_allocation_callbacks* pAllocationCallbacks);
+DRWAV_API drwav_bool32 drwav_init_file_write_sequential(drwav* pWav, const char* filename, const drwav_data_format* pFormat, drwav_uint64 totalSampleCount, const drwav_allocation_callbacks* pAllocationCallbacks);
+DRWAV_API drwav_bool32 drwav_init_file_write_sequential_pcm_frames(drwav* pWav, const char* filename, const drwav_data_format* pFormat, drwav_uint64 totalPCMFrameCount, const drwav_allocation_callbacks* pAllocationCallbacks);
+DRWAV_API drwav_bool32 drwav_init_file_write_w(drwav* pWav, const wchar_t* filename, const drwav_data_format* pFormat, const drwav_allocation_callbacks* pAllocationCallbacks);
+DRWAV_API drwav_bool32 drwav_init_file_write_sequential_w(drwav* pWav, const wchar_t* filename, const drwav_data_format* pFormat, drwav_uint64 totalSampleCount, const drwav_allocation_callbacks* pAllocationCallbacks);
+DRWAV_API drwav_bool32 drwav_init_file_write_sequential_pcm_frames_w(drwav* pWav, const wchar_t* filename, const drwav_data_format* pFormat, drwav_uint64 totalPCMFrameCount, const drwav_allocation_callbacks* pAllocationCallbacks);
+#endif  /* DR_WAV_NO_STDIO */
+
+/*
+Helper for initializing a loader from a pre-allocated memory buffer.
+
+This does not create a copy of the data. It is up to the application to ensure the buffer remains valid for
+the lifetime of the drwav object.
+
+The buffer should contain the contents of the entire wave file, not just the sample data.
+*/
+DRWAV_API drwav_bool32 drwav_init_memory(drwav* pWav, const void* data, size_t dataSize, const drwav_allocation_callbacks* pAllocationCallbacks);
+DRWAV_API drwav_bool32 drwav_init_memory_ex(drwav* pWav, const void* data, size_t dataSize, drwav_chunk_proc onChunk, void* pChunkUserData, drwav_uint32 flags, const drwav_allocation_callbacks* pAllocationCallbacks);
+
+/*
+Helper for initializing a writer which outputs data to a memory buffer.
+
+dr_wav will manage the memory allocations, however it is up to the caller to free the data with drwav_free().
+
+The buffer will remain allocated even after drwav_uninit() is called. The buffer should not be considered valid
+until after drwav_uninit() has been called.
+*/
+DRWAV_API drwav_bool32 drwav_init_memory_write(drwav* pWav, void** ppData, size_t* pDataSize, const drwav_data_format* pFormat, const drwav_allocation_callbacks* pAllocationCallbacks);
+DRWAV_API drwav_bool32 drwav_init_memory_write_sequential(drwav* pWav, void** ppData, size_t* pDataSize, const drwav_data_format* pFormat, drwav_uint64 totalSampleCount, const drwav_allocation_callbacks* pAllocationCallbacks);
+DRWAV_API drwav_bool32 drwav_init_memory_write_sequential_pcm_frames(drwav* pWav, void** ppData, size_t* pDataSize, const drwav_data_format* pFormat, drwav_uint64 totalPCMFrameCount, const drwav_allocation_callbacks* pAllocationCallbacks);
+
+
+#ifndef DR_WAV_NO_CONVERSION_API
+/*
+Opens and reads an entire wav file in a single operation.
+
+The return value is a heap-allocated buffer containing the audio data. Use drwav_free() to free the buffer.
+*/
+DRWAV_API drwav_int16* drwav_open_and_read_pcm_frames_s16(drwav_read_proc onRead, drwav_seek_proc onSeek, void* pUserData, unsigned int* channelsOut, unsigned int* sampleRateOut, drwav_uint64* totalFrameCountOut, const drwav_allocation_callbacks* pAllocationCallbacks);
+DRWAV_API float* drwav_open_and_read_pcm_frames_f32(drwav_read_proc onRead, drwav_seek_proc onSeek, void* pUserData, unsigned int* channelsOut, unsigned int* sampleRateOut, drwav_uint64* totalFrameCountOut, const drwav_allocation_callbacks* pAllocationCallbacks);
+DRWAV_API drwav_int32* drwav_open_and_read_pcm_frames_s32(drwav_read_proc onRead, drwav_seek_proc onSeek, void* pUserData, unsigned int* channelsOut, unsigned int* sampleRateOut, drwav_uint64* totalFrameCountOut, const drwav_allocation_callbacks* pAllocationCallbacks);
+#ifndef DR_WAV_NO_STDIO
+/*
+Opens and decodes an entire wav file in a single operation.
+
+The return value is a heap-allocated buffer containing the audio data. Use drwav_free() to free the buffer.
+*/
+DRWAV_API drwav_int16* drwav_open_file_and_read_pcm_frames_s16(const char* filename, unsigned int* channelsOut, unsigned int* sampleRateOut, drwav_uint64* totalFrameCountOut, const drwav_allocation_callbacks* pAllocationCallbacks);
+DRWAV_API float* drwav_open_file_and_read_pcm_frames_f32(const char* filename, unsigned int* channelsOut, unsigned int* sampleRateOut, drwav_uint64* totalFrameCountOut, const drwav_allocation_callbacks* pAllocationCallbacks);
+DRWAV_API drwav_int32* drwav_open_file_and_read_pcm_frames_s32(const char* filename, unsigned int* channelsOut, unsigned int* sampleRateOut, drwav_uint64* totalFrameCountOut, const drwav_allocation_callbacks* pAllocationCallbacks);
+DRWAV_API drwav_int16* drwav_open_file_and_read_pcm_frames_s16_w(const wchar_t* filename, unsigned int* channelsOut, unsigned int* sampleRateOut, drwav_uint64* totalFrameCountOut, const drwav_allocation_callbacks* pAllocationCallbacks);
+DRWAV_API float* drwav_open_file_and_read_pcm_frames_f32_w(const wchar_t* filename, unsigned int* channelsOut, unsigned int* sampleRateOut, drwav_uint64* totalFrameCountOut, const drwav_allocation_callbacks* pAllocationCallbacks);
+DRWAV_API drwav_int32* drwav_open_file_and_read_pcm_frames_s32_w(const wchar_t* filename, unsigned int* channelsOut, unsigned int* sampleRateOut, drwav_uint64* totalFrameCountOut, const drwav_allocation_callbacks* pAllocationCallbacks);
+#endif
+/*
+Opens and decodes an entire wav file from a block of memory in a single operation.
+
+The return value is a heap-allocated buffer containing the audio data. Use drwav_free() to free the buffer.
+*/
+DRWAV_API drwav_int16* drwav_open_memory_and_read_pcm_frames_s16(const void* data, size_t dataSize, unsigned int* channelsOut, unsigned int* sampleRateOut, drwav_uint64* totalFrameCountOut, const drwav_allocation_callbacks* pAllocationCallbacks);
+DRWAV_API float* drwav_open_memory_and_read_pcm_frames_f32(const void* data, size_t dataSize, unsigned int* channelsOut, unsigned int* sampleRateOut, drwav_uint64* totalFrameCountOut, const drwav_allocation_callbacks* pAllocationCallbacks);
+DRWAV_API drwav_int32* drwav_open_memory_and_read_pcm_frames_s32(const void* data, size_t dataSize, unsigned int* channelsOut, unsigned int* sampleRateOut, drwav_uint64* totalFrameCountOut, const drwav_allocation_callbacks* pAllocationCallbacks);
+#endif
+
+/* Frees data that was allocated internally by dr_wav. */
+DRWAV_API void drwav_free(void* p, const drwav_allocation_callbacks* pAllocationCallbacks);
+
+/* Converts bytes from a wav stream to a sized type of native endian. */
+DRWAV_API drwav_uint16 drwav_bytes_to_u16(const drwav_uint8* data);
+DRWAV_API drwav_int16 drwav_bytes_to_s16(const drwav_uint8* data);
+DRWAV_API drwav_uint32 drwav_bytes_to_u32(const drwav_uint8* data);
+DRWAV_API drwav_int32 drwav_bytes_to_s32(const drwav_uint8* data);
+DRWAV_API drwav_uint64 drwav_bytes_to_u64(const drwav_uint8* data);
+DRWAV_API drwav_int64 drwav_bytes_to_s64(const drwav_uint8* data);
+
+/* Compares a GUID for the purpose of checking the type of a Wave64 chunk. */
+DRWAV_API drwav_bool32 drwav_guid_equal(const drwav_uint8 a[16], const drwav_uint8 b[16]);
+
+/* Compares a four-character-code for the purpose of checking the type of a RIFF chunk. */
+DRWAV_API drwav_bool32 drwav_fourcc_equal(const drwav_uint8* a, const char* b);
+
+#ifdef __cplusplus
+}
+#endif
+#endif  /* dr_wav_h */
+
+
+/************************************************************************************************************************************************************
+ ************************************************************************************************************************************************************
+
+ IMPLEMENTATION
+
+ ************************************************************************************************************************************************************
+ ************************************************************************************************************************************************************/
+#if defined(DR_WAV_IMPLEMENTATION) || defined(DRWAV_IMPLEMENTATION)
+#ifndef dr_wav_c
+#define dr_wav_c
+
+#include <stdlib.h>
+#include <string.h> /* For memcpy(), memset() */
+#include <limits.h> /* For INT_MAX */
+
+#ifndef DR_WAV_NO_STDIO
+#include <stdio.h>
+#include <wchar.h>
+#endif
+
+/* Standard library stuff. */
+#ifndef DRWAV_ASSERT
+#include <assert.h>
+#define DRWAV_ASSERT(expression)           assert(expression)
+#endif
+#ifndef DRWAV_MALLOC
+#define DRWAV_MALLOC(sz)                   malloc((sz))
+#endif
+#ifndef DRWAV_REALLOC
+#define DRWAV_REALLOC(p, sz)               realloc((p), (sz))
+#endif
+#ifndef DRWAV_FREE
+#define DRWAV_FREE(p)                      free((p))
+#endif
+#ifndef DRWAV_COPY_MEMORY
+#define DRWAV_COPY_MEMORY(dst, src, sz)    memcpy((dst), (src), (sz))
+#endif
+#ifndef DRWAV_ZERO_MEMORY
+#define DRWAV_ZERO_MEMORY(p, sz)           memset((p), 0, (sz))
+#endif
+#ifndef DRWAV_ZERO_OBJECT
+#define DRWAV_ZERO_OBJECT(p)               DRWAV_ZERO_MEMORY((p), sizeof(*p))
+#endif
+
+#define drwav_countof(x)                   (sizeof(x) / sizeof(x[0]))
+#define drwav_align(x, a)                  ((((x) + (a) - 1) / (a)) * (a))
+#define drwav_min(a, b)                    (((a) < (b)) ? (a) : (b))
+#define drwav_max(a, b)                    (((a) > (b)) ? (a) : (b))
+#define drwav_clamp(x, lo, hi)             (drwav_max((lo), drwav_min((hi), (x))))
+
+#define DRWAV_MAX_SIMD_VECTOR_SIZE         64  /* 64 for AVX-512 in the future. */
+
+/* CPU architecture. */
+#if defined(__x86_64__) || defined(_M_X64)
+    #define DRWAV_X64
+#elif defined(__i386) || defined(_M_IX86)
+    #define DRWAV_X86
+#elif defined(__arm__) || defined(_M_ARM)
+    #define DRWAV_ARM
+#endif
+
+#ifdef _MSC_VER
+    #define DRWAV_INLINE __forceinline
+#elif defined(__GNUC__)
+    /*
+    I've had a bug report where GCC is emitting warnings about functions possibly not being inlineable. This warning happens when
+    the __attribute__((always_inline)) attribute is defined without an "inline" statement. I think therefore there must be some
+    case where "__inline__" is not always defined, thus the compiler emitting these warnings. When using -std=c89 or -ansi on the
+    command line, we cannot use the "inline" keyword and instead need to use "__inline__". In an attempt to work around this issue
+    I am using "__inline__" only when we're compiling in strict ANSI mode.
+    */
+    #if defined(__STRICT_ANSI__)
+        #define DRWAV_INLINE __inline__ __attribute__((always_inline))
+    #else
+        #define DRWAV_INLINE inline __attribute__((always_inline))
+    #endif
+#elif defined(__WATCOMC__)
+    #define DRWAV_INLINE __inline
+#else
+    #define DRWAV_INLINE
+#endif
+
+#if defined(SIZE_MAX)
+    #define DRWAV_SIZE_MAX  SIZE_MAX
+#else
+    #if defined(_WIN64) || defined(_LP64) || defined(__LP64__)
+        #define DRWAV_SIZE_MAX  ((drwav_uint64)0xFFFFFFFFFFFFFFFF)
+    #else
+        #define DRWAV_SIZE_MAX  0xFFFFFFFF
+    #endif
+#endif
+
+#if defined(_MSC_VER) && _MSC_VER >= 1400
+    #define DRWAV_HAS_BYTESWAP16_INTRINSIC
+    #define DRWAV_HAS_BYTESWAP32_INTRINSIC
+    #define DRWAV_HAS_BYTESWAP64_INTRINSIC
+#elif defined(__clang__)
+    #if defined(__has_builtin)
+        #if __has_builtin(__builtin_bswap16)
+            #define DRWAV_HAS_BYTESWAP16_INTRINSIC
+        #endif
+        #if __has_builtin(__builtin_bswap32)
+            #define DRWAV_HAS_BYTESWAP32_INTRINSIC
+        #endif
+        #if __has_builtin(__builtin_bswap64)
+            #define DRWAV_HAS_BYTESWAP64_INTRINSIC
+        #endif
+    #endif
+#elif defined(__GNUC__)
+    #if ((__GNUC__ > 4) || (__GNUC__ == 4 && __GNUC_MINOR__ >= 3))
+        #define DRWAV_HAS_BYTESWAP32_INTRINSIC
+        #define DRWAV_HAS_BYTESWAP64_INTRINSIC
+    #endif
+    #if ((__GNUC__ > 4) || (__GNUC__ == 4 && __GNUC_MINOR__ >= 8))
+        #define DRWAV_HAS_BYTESWAP16_INTRINSIC
+    #endif
+#endif
+
+DRWAV_API void drwav_version(drwav_uint32* pMajor, drwav_uint32* pMinor, drwav_uint32* pRevision)
+{
+    if (pMajor) {
+        *pMajor = DRWAV_VERSION_MAJOR;
+    }
+
+    if (pMinor) {
+        *pMinor = DRWAV_VERSION_MINOR;
+    }
+
+    if (pRevision) {
+        *pRevision = DRWAV_VERSION_REVISION;
+    }
+}
+
+DRWAV_API const char* drwav_version_string(void)
+{
+    return DRWAV_VERSION_STRING;
+}
+
+/*
+These limits are used for basic validation when initializing the decoder. If you exceed these limits, first of all: what on Earth are
+you doing?! (Let me know, I'd be curious!) Second, you can adjust these by #define-ing them before the dr_wav implementation.
+*/
+#ifndef DRWAV_MAX_SAMPLE_RATE
+#define DRWAV_MAX_SAMPLE_RATE       384000
+#endif
+#ifndef DRWAV_MAX_CHANNELS
+#define DRWAV_MAX_CHANNELS          256
+#endif
+#ifndef DRWAV_MAX_BITS_PER_SAMPLE
+#define DRWAV_MAX_BITS_PER_SAMPLE   64
+#endif
+
+static const drwav_uint8 drwavGUID_W64_RIFF[16] = {0x72,0x69,0x66,0x66, 0x2E,0x91, 0xCF,0x11, 0xA5,0xD6, 0x28,0xDB,0x04,0xC1,0x00,0x00};    /* 66666972-912E-11CF-A5D6-28DB04C10000 */
+static const drwav_uint8 drwavGUID_W64_WAVE[16] = {0x77,0x61,0x76,0x65, 0xF3,0xAC, 0xD3,0x11, 0x8C,0xD1, 0x00,0xC0,0x4F,0x8E,0xDB,0x8A};    /* 65766177-ACF3-11D3-8CD1-00C04F8EDB8A */
+/*static const drwav_uint8 drwavGUID_W64_JUNK[16] = {0x6A,0x75,0x6E,0x6B, 0xF3,0xAC, 0xD3,0x11, 0x8C,0xD1, 0x00,0xC0,0x4F,0x8E,0xDB,0x8A};*/    /* 6B6E756A-ACF3-11D3-8CD1-00C04F8EDB8A */
+static const drwav_uint8 drwavGUID_W64_FMT [16] = {0x66,0x6D,0x74,0x20, 0xF3,0xAC, 0xD3,0x11, 0x8C,0xD1, 0x00,0xC0,0x4F,0x8E,0xDB,0x8A};    /* 20746D66-ACF3-11D3-8CD1-00C04F8EDB8A */
+static const drwav_uint8 drwavGUID_W64_FACT[16] = {0x66,0x61,0x63,0x74, 0xF3,0xAC, 0xD3,0x11, 0x8C,0xD1, 0x00,0xC0,0x4F,0x8E,0xDB,0x8A};    /* 74636166-ACF3-11D3-8CD1-00C04F8EDB8A */
+static const drwav_uint8 drwavGUID_W64_DATA[16] = {0x64,0x61,0x74,0x61, 0xF3,0xAC, 0xD3,0x11, 0x8C,0xD1, 0x00,0xC0,0x4F,0x8E,0xDB,0x8A};    /* 61746164-ACF3-11D3-8CD1-00C04F8EDB8A */
+static const drwav_uint8 drwavGUID_W64_SMPL[16] = {0x73,0x6D,0x70,0x6C, 0xF3,0xAC, 0xD3,0x11, 0x8C,0xD1, 0x00,0xC0,0x4F,0x8E,0xDB,0x8A};    /* 6C706D73-ACF3-11D3-8CD1-00C04F8EDB8A */
+
+static DRWAV_INLINE drwav_bool32 drwav__guid_equal(const drwav_uint8 a[16], const drwav_uint8 b[16])
+{
+    int i;
+    for (i = 0; i < 16; i += 1) {
+        if (a[i] != b[i]) {
+            return DRWAV_FALSE;
+        }
+    }
+
+    return DRWAV_TRUE;
+}
+
+static DRWAV_INLINE drwav_bool32 drwav__fourcc_equal(const drwav_uint8* a, const char* b)
+{
+    return
+        a[0] == b[0] &&
+        a[1] == b[1] &&
+        a[2] == b[2] &&
+        a[3] == b[3];
+}
+
+
+
+static DRWAV_INLINE int drwav__is_little_endian(void)
+{
+#if defined(DRWAV_X86) || defined(DRWAV_X64)
+    return DRWAV_TRUE;
+#elif defined(__BYTE_ORDER) && defined(__LITTLE_ENDIAN) && __BYTE_ORDER == __LITTLE_ENDIAN
+    return DRWAV_TRUE;
+#else
+    int n = 1;
+    return (*(char*)&n) == 1;
+#endif
+}
+
+static DRWAV_INLINE drwav_uint16 drwav__bytes_to_u16(const drwav_uint8* data)
+{
+    return (data[0] << 0) | (data[1] << 8);
+}
+
+static DRWAV_INLINE drwav_int16 drwav__bytes_to_s16(const drwav_uint8* data)
+{
+    return (short)drwav__bytes_to_u16(data);
+}
+
+static DRWAV_INLINE drwav_uint32 drwav__bytes_to_u32(const drwav_uint8* data)
+{
+    return (data[0] << 0) | (data[1] << 8) | (data[2] << 16) | (data[3] << 24);
+}
+
+static DRWAV_INLINE drwav_int32 drwav__bytes_to_s32(const drwav_uint8* data)
+{
+    return (drwav_int32)drwav__bytes_to_u32(data);
+}
+
+static DRWAV_INLINE drwav_uint64 drwav__bytes_to_u64(const drwav_uint8* data)
+{
+    return
+        ((drwav_uint64)data[0] <<  0) | ((drwav_uint64)data[1] <<  8) | ((drwav_uint64)data[2] << 16) | ((drwav_uint64)data[3] << 24) |
+        ((drwav_uint64)data[4] << 32) | ((drwav_uint64)data[5] << 40) | ((drwav_uint64)data[6] << 48) | ((drwav_uint64)data[7] << 56);
+}
+
+static DRWAV_INLINE drwav_int64 drwav__bytes_to_s64(const drwav_uint8* data)
+{
+    return (drwav_int64)drwav__bytes_to_u64(data);
+}
+
+static DRWAV_INLINE void drwav__bytes_to_guid(const drwav_uint8* data, drwav_uint8* guid)
+{
+    int i;
+    for (i = 0; i < 16; ++i) {
+        guid[i] = data[i];
+    }
+}
+
+
+static DRWAV_INLINE drwav_uint16 drwav__bswap16(drwav_uint16 n)
+{
+#ifdef DRWAV_HAS_BYTESWAP16_INTRINSIC
+    #if defined(_MSC_VER)
+        return _byteswap_ushort(n);
+    #elif defined(__GNUC__) || defined(__clang__)
+        return __builtin_bswap16(n);
+    #else
+        #error "This compiler does not support the byte swap intrinsic."
+    #endif
+#else
+    return ((n & 0xFF00) >> 8) |
+           ((n & 0x00FF) << 8);
+#endif
+}
+
+static DRWAV_INLINE drwav_uint32 drwav__bswap32(drwav_uint32 n)
+{
+#ifdef DRWAV_HAS_BYTESWAP32_INTRINSIC
+    #if defined(_MSC_VER)
+        return _byteswap_ulong(n);
+    #elif defined(__GNUC__) || defined(__clang__)
+        #if defined(DRWAV_ARM) && (defined(__ARM_ARCH) && __ARM_ARCH >= 6) && !defined(DRWAV_64BIT)   /* <-- 64-bit inline assembly has not been tested, so disabling for now. */
+            /* Inline assembly optimized implementation for ARM. In my testing, GCC does not generate optimized code with __builtin_bswap32(). */
+            drwav_uint32 r;
+            __asm__ __volatile__ (
+            #if defined(DRWAV_64BIT)
+                "rev %w[out], %w[in]" : [out]"=r"(r) : [in]"r"(n)   /* <-- This is untested. If someone in the community could test this, that would be appreciated! */
+            #else
+                "rev %[out], %[in]" : [out]"=r"(r) : [in]"r"(n)
+            #endif
+            );
+            return r;
+        #else
+            return __builtin_bswap32(n);
+        #endif
+    #else
+        #error "This compiler does not support the byte swap intrinsic."
+    #endif
+#else
+    return ((n & 0xFF000000) >> 24) |
+           ((n & 0x00FF0000) >>  8) |
+           ((n & 0x0000FF00) <<  8) |
+           ((n & 0x000000FF) << 24);
+#endif
+}
+
+static DRWAV_INLINE drwav_uint64 drwav__bswap64(drwav_uint64 n)
+{
+#ifdef DRWAV_HAS_BYTESWAP64_INTRINSIC
+    #if defined(_MSC_VER)
+        return _byteswap_uint64(n);
+    #elif defined(__GNUC__) || defined(__clang__)
+        return __builtin_bswap64(n);
+    #else
+        #error "This compiler does not support the byte swap intrinsic."
+    #endif
+#else
+    /* Weird "<< 32" bitshift is required for C89 because it doesn't support 64-bit constants. Should be optimized out by a good compiler. */
+    return ((n & ((drwav_uint64)0xFF000000 << 32)) >> 56) |
+           ((n & ((drwav_uint64)0x00FF0000 << 32)) >> 40) |
+           ((n & ((drwav_uint64)0x0000FF00 << 32)) >> 24) |
+           ((n & ((drwav_uint64)0x000000FF << 32)) >>  8) |
+           ((n & ((drwav_uint64)0xFF000000      )) <<  8) |
+           ((n & ((drwav_uint64)0x00FF0000      )) << 24) |
+           ((n & ((drwav_uint64)0x0000FF00      )) << 40) |
+           ((n & ((drwav_uint64)0x000000FF      )) << 56);
+#endif
+}
+
+
+static DRWAV_INLINE drwav_int16 drwav__bswap_s16(drwav_int16 n)
+{
+    return (drwav_int16)drwav__bswap16((drwav_uint16)n);
+}
+
+static DRWAV_INLINE void drwav__bswap_samples_s16(drwav_int16* pSamples, drwav_uint64 sampleCount)
+{
+    drwav_uint64 iSample;
+    for (iSample = 0; iSample < sampleCount; iSample += 1) {
+        pSamples[iSample] = drwav__bswap_s16(pSamples[iSample]);
+    }
+}
+
+
+static DRWAV_INLINE void drwav__bswap_s24(drwav_uint8* p)
+{
+    drwav_uint8 t;
+    t = p[0];
+    p[0] = p[2];
+    p[2] = t;
+}
+
+static DRWAV_INLINE void drwav__bswap_samples_s24(drwav_uint8* pSamples, drwav_uint64 sampleCount)
+{
+    drwav_uint64 iSample;
+    for (iSample = 0; iSample < sampleCount; iSample += 1) {
+        drwav_uint8* pSample = pSamples + (iSample*3);
+        drwav__bswap_s24(pSample);
+    }
+}
+
+
+static DRWAV_INLINE drwav_int32 drwav__bswap_s32(drwav_int32 n)
+{
+    return (drwav_int32)drwav__bswap32((drwav_uint32)n);
+}
+
+static DRWAV_INLINE void drwav__bswap_samples_s32(drwav_int32* pSamples, drwav_uint64 sampleCount)
+{
+    drwav_uint64 iSample;
+    for (iSample = 0; iSample < sampleCount; iSample += 1) {
+        pSamples[iSample] = drwav__bswap_s32(pSamples[iSample]);
+    }
+}
+
+
+static DRWAV_INLINE float drwav__bswap_f32(float n)
+{
+    union {
+        drwav_uint32 i;
+        float f;
+    } x;
+    x.f = n;
+    x.i = drwav__bswap32(x.i);
+
+    return x.f;
+}
+
+static DRWAV_INLINE void drwav__bswap_samples_f32(float* pSamples, drwav_uint64 sampleCount)
+{
+    drwav_uint64 iSample;
+    for (iSample = 0; iSample < sampleCount; iSample += 1) {
+        pSamples[iSample] = drwav__bswap_f32(pSamples[iSample]);
+    }
+}
+
+
+static DRWAV_INLINE double drwav__bswap_f64(double n)
+{
+    union {
+        drwav_uint64 i;
+        double f;
+    } x;
+    x.f = n;
+    x.i = drwav__bswap64(x.i);
+
+    return x.f;
+}
+
+static DRWAV_INLINE void drwav__bswap_samples_f64(double* pSamples, drwav_uint64 sampleCount)
+{
+    drwav_uint64 iSample;
+    for (iSample = 0; iSample < sampleCount; iSample += 1) {
+        pSamples[iSample] = drwav__bswap_f64(pSamples[iSample]);
+    }
+}
+
+
+static DRWAV_INLINE void drwav__bswap_samples_pcm(void* pSamples, drwav_uint64 sampleCount, drwav_uint32 bytesPerSample)
+{
+    /* Assumes integer PCM. Floating point PCM is done in drwav__bswap_samples_ieee(). */
+    switch (bytesPerSample)
+    {
+        case 2: /* s16, s12 (loosely packed) */
+        {
+            drwav__bswap_samples_s16((drwav_int16*)pSamples, sampleCount);
+        } break;
+        case 3: /* s24 */
+        {
+            drwav__bswap_samples_s24((drwav_uint8*)pSamples, sampleCount);
+        } break;
+        case 4: /* s32 */
+        {
+            drwav__bswap_samples_s32((drwav_int32*)pSamples, sampleCount);
+        } break;
+        default:
+        {
+            /* Unsupported format. */
+            DRWAV_ASSERT(DRWAV_FALSE);
+        } break;
+    }
+}
+
+static DRWAV_INLINE void drwav__bswap_samples_ieee(void* pSamples, drwav_uint64 sampleCount, drwav_uint32 bytesPerSample)
+{
+    switch (bytesPerSample)
+    {
+    #if 0   /* Contributions welcome for f16 support. */
+        case 2: /* f16 */
+        {
+            drwav__bswap_samples_f16((drwav_float16*)pSamples, sampleCount);
+        } break;
+    #endif
+        case 4: /* f32 */
+        {
+            drwav__bswap_samples_f32((float*)pSamples, sampleCount);
+        } break;
+        case 8: /* f64 */
+        {
+            drwav__bswap_samples_f64((double*)pSamples, sampleCount);
+        } break;
+        default:
+        {
+            /* Unsupported format. */
+            DRWAV_ASSERT(DRWAV_FALSE);
+        } break;
+    }
+}
+
+static DRWAV_INLINE void drwav__bswap_samples(void* pSamples, drwav_uint64 sampleCount, drwav_uint32 bytesPerSample, drwav_uint16 format)
+{
+    switch (format)
+    {
+        case DR_WAVE_FORMAT_PCM:
+        {
+            drwav__bswap_samples_pcm(pSamples, sampleCount, bytesPerSample);
+        } break;
+
+        case DR_WAVE_FORMAT_IEEE_FLOAT:
+        {
+            drwav__bswap_samples_ieee(pSamples, sampleCount, bytesPerSample);
+        } break;
+
+        case DR_WAVE_FORMAT_ALAW:
+        case DR_WAVE_FORMAT_MULAW:
+        {
+            drwav__bswap_samples_s16((drwav_int16*)pSamples, sampleCount);
+        } break;
+
+        case DR_WAVE_FORMAT_ADPCM:
+        case DR_WAVE_FORMAT_DVI_ADPCM:
+        default:
+        {
+            /* Unsupported format. */
+            DRWAV_ASSERT(DRWAV_FALSE);
+        } break;
+    }
+}
+
+
+static void* drwav__malloc_default(size_t sz, void* pUserData)
+{
+    (void)pUserData;
+    return DRWAV_MALLOC(sz);
+}
+
+static void* drwav__realloc_default(void* p, size_t sz, void* pUserData)
+{
+    (void)pUserData;
+    return DRWAV_REALLOC(p, sz);
+}
+
+static void drwav__free_default(void* p, void* pUserData)
+{
+    (void)pUserData;
+    DRWAV_FREE(p);
+}
+
+
+static void* drwav__malloc_from_callbacks(size_t sz, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    if (pAllocationCallbacks == NULL) {
+        return NULL;
+    }
+
+    if (pAllocationCallbacks->onMalloc != NULL) {
+        return pAllocationCallbacks->onMalloc(sz, pAllocationCallbacks->pUserData);
+    }
+
+    /* Try using realloc(). */
+    if (pAllocationCallbacks->onRealloc != NULL) {
+        return pAllocationCallbacks->onRealloc(NULL, sz, pAllocationCallbacks->pUserData);
+    }
+
+    return NULL;
+}
+
+static void* drwav__realloc_from_callbacks(void* p, size_t szNew, size_t szOld, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    if (pAllocationCallbacks == NULL) {
+        return NULL;
+    }
+
+    if (pAllocationCallbacks->onRealloc != NULL) {
+        return pAllocationCallbacks->onRealloc(p, szNew, pAllocationCallbacks->pUserData);
+    }
+
+    /* Try emulating realloc() in terms of malloc()/free(). */
+    if (pAllocationCallbacks->onMalloc != NULL && pAllocationCallbacks->onFree != NULL) {
+        void* p2;
+
+        p2 = pAllocationCallbacks->onMalloc(szNew, pAllocationCallbacks->pUserData);
+        if (p2 == NULL) {
+            return NULL;
+        }
+
+        if (p != NULL) {
+            DRWAV_COPY_MEMORY(p2, p, szOld);
+            pAllocationCallbacks->onFree(p, pAllocationCallbacks->pUserData);
+        }
+
+        return p2;
+    }
+
+    return NULL;
+}
+
+static void drwav__free_from_callbacks(void* p, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    if (p == NULL || pAllocationCallbacks == NULL) {
+        return;
+    }
+
+    if (pAllocationCallbacks->onFree != NULL) {
+        pAllocationCallbacks->onFree(p, pAllocationCallbacks->pUserData);
+    }
+}
+
+
+static drwav_allocation_callbacks drwav_copy_allocation_callbacks_or_defaults(const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    if (pAllocationCallbacks != NULL) {
+        /* Copy. */
+        return *pAllocationCallbacks;
+    } else {
+        /* Defaults. */
+        drwav_allocation_callbacks allocationCallbacks;
+        allocationCallbacks.pUserData = NULL;
+        allocationCallbacks.onMalloc  = drwav__malloc_default;
+        allocationCallbacks.onRealloc = drwav__realloc_default;
+        allocationCallbacks.onFree    = drwav__free_default;
+        return allocationCallbacks;
+    }
+}
+
+
+static DRWAV_INLINE drwav_bool32 drwav__is_compressed_format_tag(drwav_uint16 formatTag)
+{
+    return
+        formatTag == DR_WAVE_FORMAT_ADPCM ||
+        formatTag == DR_WAVE_FORMAT_DVI_ADPCM;
+}
+
+static unsigned int drwav__chunk_padding_size_riff(drwav_uint64 chunkSize)
+{
+    return (unsigned int)(chunkSize % 2);
+}
+
+static unsigned int drwav__chunk_padding_size_w64(drwav_uint64 chunkSize)
+{
+    return (unsigned int)(chunkSize % 8);
+}
+
+static drwav_uint64 drwav_read_pcm_frames_s16__msadpcm(drwav* pWav, drwav_uint64 samplesToRead, drwav_int16* pBufferOut);
+static drwav_uint64 drwav_read_pcm_frames_s16__ima(drwav* pWav, drwav_uint64 samplesToRead, drwav_int16* pBufferOut);
+static drwav_bool32 drwav_init_write__internal(drwav* pWav, const drwav_data_format* pFormat, drwav_uint64 totalSampleCount);
+
+static drwav_result drwav__read_chunk_header(drwav_read_proc onRead, void* pUserData, drwav_container container, drwav_uint64* pRunningBytesReadOut, drwav_chunk_header* pHeaderOut)
+{
+    if (container == drwav_container_riff || container == drwav_container_rf64) {
+        drwav_uint8 sizeInBytes[4];
+
+        if (onRead(pUserData, pHeaderOut->id.fourcc, 4) != 4) {
+            return DRWAV_AT_END;
+        }
+
+        if (onRead(pUserData, sizeInBytes, 4) != 4) {
+            return DRWAV_INVALID_FILE;
+        }
+
+        pHeaderOut->sizeInBytes = drwav__bytes_to_u32(sizeInBytes);
+        pHeaderOut->paddingSize = drwav__chunk_padding_size_riff(pHeaderOut->sizeInBytes);
+        *pRunningBytesReadOut += 8;
+    } else {
+        drwav_uint8 sizeInBytes[8];
+
+        if (onRead(pUserData, pHeaderOut->id.guid, 16) != 16) {
+            return DRWAV_AT_END;
+        }
+
+        if (onRead(pUserData, sizeInBytes, 8) != 8) {
+            return DRWAV_INVALID_FILE;
+        }
+
+        pHeaderOut->sizeInBytes = drwav__bytes_to_u64(sizeInBytes) - 24;    /* <-- Subtract 24 because w64 includes the size of the header. */
+        pHeaderOut->paddingSize = drwav__chunk_padding_size_w64(pHeaderOut->sizeInBytes);
+        *pRunningBytesReadOut += 24;
+    }
+
+    return DRWAV_SUCCESS;
+}
+
+static drwav_bool32 drwav__seek_forward(drwav_seek_proc onSeek, drwav_uint64 offset, void* pUserData)
+{
+    drwav_uint64 bytesRemainingToSeek = offset;
+    while (bytesRemainingToSeek > 0) {
+        if (bytesRemainingToSeek > 0x7FFFFFFF) {
+            if (!onSeek(pUserData, 0x7FFFFFFF, drwav_seek_origin_current)) {
+                return DRWAV_FALSE;
+            }
+            bytesRemainingToSeek -= 0x7FFFFFFF;
+        } else {
+            if (!onSeek(pUserData, (int)bytesRemainingToSeek, drwav_seek_origin_current)) {
+                return DRWAV_FALSE;
+            }
+            bytesRemainingToSeek = 0;
+        }
+    }
+
+    return DRWAV_TRUE;
+}
+
+static drwav_bool32 drwav__seek_from_start(drwav_seek_proc onSeek, drwav_uint64 offset, void* pUserData)
+{
+    if (offset <= 0x7FFFFFFF) {
+        return onSeek(pUserData, (int)offset, drwav_seek_origin_start);
+    }
+
+    /* Larger than 32-bit seek. */
+    if (!onSeek(pUserData, 0x7FFFFFFF, drwav_seek_origin_start)) {
+        return DRWAV_FALSE;
+    }
+    offset -= 0x7FFFFFFF;
+
+    for (;;) {
+        if (offset <= 0x7FFFFFFF) {
+            return onSeek(pUserData, (int)offset, drwav_seek_origin_current);
+        }
+
+        if (!onSeek(pUserData, 0x7FFFFFFF, drwav_seek_origin_current)) {
+            return DRWAV_FALSE;
+        }
+        offset -= 0x7FFFFFFF;
+    }
+
+    /* Should never get here. */
+    /*return DRWAV_TRUE; */
+}
+
+
+static drwav_bool32 drwav__read_fmt(drwav_read_proc onRead, drwav_seek_proc onSeek, void* pUserData, drwav_container container, drwav_uint64* pRunningBytesReadOut, drwav_fmt* fmtOut)
+{
+    drwav_chunk_header header;
+    drwav_uint8 fmt[16];
+
+    if (drwav__read_chunk_header(onRead, pUserData, container, pRunningBytesReadOut, &header) != DRWAV_SUCCESS) {
+        return DRWAV_FALSE;
+    }
+
+
+    /* Skip non-fmt chunks. */
+    while (((container == drwav_container_riff || container == drwav_container_rf64) && !drwav__fourcc_equal(header.id.fourcc, "fmt ")) || (container == drwav_container_w64 && !drwav__guid_equal(header.id.guid, drwavGUID_W64_FMT))) {
+        if (!drwav__seek_forward(onSeek, header.sizeInBytes + header.paddingSize, pUserData)) {
+            return DRWAV_FALSE;
+        }
+        *pRunningBytesReadOut += header.sizeInBytes + header.paddingSize;
+
+        /* Try the next header. */
+        if (drwav__read_chunk_header(onRead, pUserData, container, pRunningBytesReadOut, &header) != DRWAV_SUCCESS) {
+            return DRWAV_FALSE;
+        }
+    }
+
+
+    /* Validation. */
+    if (container == drwav_container_riff || container == drwav_container_rf64) {
+        if (!drwav__fourcc_equal(header.id.fourcc, "fmt ")) {
+            return DRWAV_FALSE;
+        }
+    } else {
+        if (!drwav__guid_equal(header.id.guid, drwavGUID_W64_FMT)) {
+            return DRWAV_FALSE;
+        }
+    }
+
+
+    if (onRead(pUserData, fmt, sizeof(fmt)) != sizeof(fmt)) {
+        return DRWAV_FALSE;
+    }
+    *pRunningBytesReadOut += sizeof(fmt);
+
+    fmtOut->formatTag      = drwav__bytes_to_u16(fmt + 0);
+    fmtOut->channels       = drwav__bytes_to_u16(fmt + 2);
+    fmtOut->sampleRate     = drwav__bytes_to_u32(fmt + 4);
+    fmtOut->avgBytesPerSec = drwav__bytes_to_u32(fmt + 8);
+    fmtOut->blockAlign     = drwav__bytes_to_u16(fmt + 12);
+    fmtOut->bitsPerSample  = drwav__bytes_to_u16(fmt + 14);
+
+    fmtOut->extendedSize       = 0;
+    fmtOut->validBitsPerSample = 0;
+    fmtOut->channelMask        = 0;
+    memset(fmtOut->subFormat, 0, sizeof(fmtOut->subFormat));
+
+    if (header.sizeInBytes > 16) {
+        drwav_uint8 fmt_cbSize[2];
+        int bytesReadSoFar = 0;
+
+        if (onRead(pUserData, fmt_cbSize, sizeof(fmt_cbSize)) != sizeof(fmt_cbSize)) {
+            return DRWAV_FALSE;    /* Expecting more data. */
+        }
+        *pRunningBytesReadOut += sizeof(fmt_cbSize);
+
+        bytesReadSoFar = 18;
+
+        fmtOut->extendedSize = drwav__bytes_to_u16(fmt_cbSize);
+        if (fmtOut->extendedSize > 0) {
+            /* Simple validation. */
+            if (fmtOut->formatTag == DR_WAVE_FORMAT_EXTENSIBLE) {
+                if (fmtOut->extendedSize != 22) {
+                    return DRWAV_FALSE;
+                }
+            }
+
+            if (fmtOut->formatTag == DR_WAVE_FORMAT_EXTENSIBLE) {
+                drwav_uint8 fmtext[22];
+                if (onRead(pUserData, fmtext, fmtOut->extendedSize) != fmtOut->extendedSize) {
+                    return DRWAV_FALSE;    /* Expecting more data. */
+                }
+
+                fmtOut->validBitsPerSample = drwav__bytes_to_u16(fmtext + 0);
+                fmtOut->channelMask        = drwav__bytes_to_u32(fmtext + 2);
+                drwav__bytes_to_guid(fmtext + 6, fmtOut->subFormat);
+            } else {
+                if (!onSeek(pUserData, fmtOut->extendedSize, drwav_seek_origin_current)) {
+                    return DRWAV_FALSE;
+                }
+            }
+            *pRunningBytesReadOut += fmtOut->extendedSize;
+
+            bytesReadSoFar += fmtOut->extendedSize;
+        }
+
+        /* Seek past any leftover bytes. For w64 the leftover will be defined based on the chunk size. */
+        if (!onSeek(pUserData, (int)(header.sizeInBytes - bytesReadSoFar), drwav_seek_origin_current)) {
+            return DRWAV_FALSE;
+        }
+        *pRunningBytesReadOut += (header.sizeInBytes - bytesReadSoFar);
+    }
+
+    if (header.paddingSize > 0) {
+        if (!onSeek(pUserData, header.paddingSize, drwav_seek_origin_current)) {
+            return DRWAV_FALSE;
+        }
+        *pRunningBytesReadOut += header.paddingSize;
+    }
+
+    return DRWAV_TRUE;
+}
+
+
+static size_t drwav__on_read(drwav_read_proc onRead, void* pUserData, void* pBufferOut, size_t bytesToRead, drwav_uint64* pCursor)
+{
+    size_t bytesRead;
+
+    DRWAV_ASSERT(onRead != NULL);
+    DRWAV_ASSERT(pCursor != NULL);
+
+    bytesRead = onRead(pUserData, pBufferOut, bytesToRead);
+    *pCursor += bytesRead;
+    return bytesRead;
+}
+
+#if 0
+static drwav_bool32 drwav__on_seek(drwav_seek_proc onSeek, void* pUserData, int offset, drwav_seek_origin origin, drwav_uint64* pCursor)
+{
+    DRWAV_ASSERT(onSeek != NULL);
+    DRWAV_ASSERT(pCursor != NULL);
+
+    if (!onSeek(pUserData, offset, origin)) {
+        return DRWAV_FALSE;
+    }
+
+    if (origin == drwav_seek_origin_start) {
+        *pCursor = offset;
+    } else {
+        *pCursor += offset;
+    }
+
+    return DRWAV_TRUE;
+}
+#endif
+
+
+
+static drwav_uint32 drwav_get_bytes_per_pcm_frame(drwav* pWav)
+{
+    /*
+    The bytes per frame is a bit ambiguous. It can be either be based on the bits per sample, or the block align. The way I'm doing it here
+    is that if the bits per sample is a multiple of 8, use floor(bitsPerSample*channels/8), otherwise fall back to the block align.
+    */
+    if ((pWav->bitsPerSample & 0x7) == 0) {
+        /* Bits per sample is a multiple of 8. */
+        return (pWav->bitsPerSample * pWav->fmt.channels) >> 3;
+    } else {
+        return pWav->fmt.blockAlign;
+    }
+}
+
+DRWAV_API drwav_uint16 drwav_fmt_get_format(const drwav_fmt* pFMT)
+{
+    if (pFMT == NULL) {
+        return 0;
+    }
+
+    if (pFMT->formatTag != DR_WAVE_FORMAT_EXTENSIBLE) {
+        return pFMT->formatTag;
+    } else {
+        return drwav__bytes_to_u16(pFMT->subFormat);    /* Only the first two bytes are required. */
+    }
+}
+
+static drwav_bool32 drwav_preinit(drwav* pWav, drwav_read_proc onRead, drwav_seek_proc onSeek, void* pReadSeekUserData, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    if (pWav == NULL || onRead == NULL || onSeek == NULL) {
+        return DRWAV_FALSE;
+    }
+
+    DRWAV_ZERO_MEMORY(pWav, sizeof(*pWav));
+    pWav->onRead    = onRead;
+    pWav->onSeek    = onSeek;
+    pWav->pUserData = pReadSeekUserData;
+    pWav->allocationCallbacks = drwav_copy_allocation_callbacks_or_defaults(pAllocationCallbacks);
+
+    if (pWav->allocationCallbacks.onFree == NULL || (pWav->allocationCallbacks.onMalloc == NULL && pWav->allocationCallbacks.onRealloc == NULL)) {
+        return DRWAV_FALSE;    /* Invalid allocation callbacks. */
+    }
+
+    return DRWAV_TRUE;
+}
+
+static drwav_bool32 drwav_init__internal(drwav* pWav, drwav_chunk_proc onChunk, void* pChunkUserData, drwav_uint32 flags)
+{
+    /* This function assumes drwav_preinit() has been called beforehand. */
+
+    drwav_uint64 cursor;    /* <-- Keeps track of the byte position so we can seek to specific locations. */
+    drwav_bool32 sequential;
+    drwav_uint8 riff[4];
+    drwav_fmt fmt;
+    unsigned short translatedFormatTag;
+    drwav_bool32 foundDataChunk;
+    drwav_uint64 dataChunkSize = 0; /* <-- Important! Don't explicitly set this to 0 anywhere else. Calculation of the size of the data chunk is performed in different paths depending on the container. */
+    drwav_uint64 sampleCountFromFactChunk = 0;  /* Same as dataChunkSize - make sure this is the only place this is initialized to 0. */
+    drwav_uint64 chunkSize;
+
+    cursor = 0;
+    sequential = (flags & DRWAV_SEQUENTIAL) != 0;
+
+    /* The first 4 bytes should be the RIFF identifier. */
+    if (drwav__on_read(pWav->onRead, pWav->pUserData, riff, sizeof(riff), &cursor) != sizeof(riff)) {
+        return DRWAV_FALSE;
+    }
+
+    /*
+    The first 4 bytes can be used to identify the container. For RIFF files it will start with "RIFF" and for
+    w64 it will start with "riff".
+    */
+    if (drwav__fourcc_equal(riff, "RIFF")) {
+        pWav->container = drwav_container_riff;
+    } else if (drwav__fourcc_equal(riff, "riff")) {
+        int i;
+        drwav_uint8 riff2[12];
+
+        pWav->container = drwav_container_w64;
+
+        /* Check the rest of the GUID for validity. */
+        if (drwav__on_read(pWav->onRead, pWav->pUserData, riff2, sizeof(riff2), &cursor) != sizeof(riff2)) {
+            return DRWAV_FALSE;
+        }
+
+        for (i = 0; i < 12; ++i) {
+            if (riff2[i] != drwavGUID_W64_RIFF[i+4]) {
+                return DRWAV_FALSE;
+            }
+        }
+    } else if (drwav__fourcc_equal(riff, "RF64")) {
+        pWav->container = drwav_container_rf64;
+    } else {
+        return DRWAV_FALSE;   /* Unknown or unsupported container. */
+    }
+
+
+    if (pWav->container == drwav_container_riff || pWav->container == drwav_container_rf64) {
+        drwav_uint8 chunkSizeBytes[4];
+        drwav_uint8 wave[4];
+
+        /* RIFF/WAVE */
+        if (drwav__on_read(pWav->onRead, pWav->pUserData, chunkSizeBytes, sizeof(chunkSizeBytes), &cursor) != sizeof(chunkSizeBytes)) {
+            return DRWAV_FALSE;
+        }
+
+        if (pWav->container == drwav_container_riff) {
+            if (drwav__bytes_to_u32(chunkSizeBytes) < 36) {
+                return DRWAV_FALSE;    /* Chunk size should always be at least 36 bytes. */
+            }
+        } else {
+            if (drwav__bytes_to_u32(chunkSizeBytes) != 0xFFFFFFFF) {
+                return DRWAV_FALSE;    /* Chunk size should always be set to -1/0xFFFFFFFF for RF64. The actual size is retrieved later. */
+            }
+        }
+
+        if (drwav__on_read(pWav->onRead, pWav->pUserData, wave, sizeof(wave), &cursor) != sizeof(wave)) {
+            return DRWAV_FALSE;
+        }
+
+        if (!drwav__fourcc_equal(wave, "WAVE")) {
+            return DRWAV_FALSE;    /* Expecting "WAVE". */
+        }
+    } else {
+        drwav_uint8 chunkSizeBytes[8];
+        drwav_uint8 wave[16];
+
+        /* W64 */
+        if (drwav__on_read(pWav->onRead, pWav->pUserData, chunkSizeBytes, sizeof(chunkSizeBytes), &cursor) != sizeof(chunkSizeBytes)) {
+            return DRWAV_FALSE;
+        }
+
+        if (drwav__bytes_to_u64(chunkSizeBytes) < 80) {
+            return DRWAV_FALSE;
+        }
+
+        if (drwav__on_read(pWav->onRead, pWav->pUserData, wave, sizeof(wave), &cursor) != sizeof(wave)) {
+            return DRWAV_FALSE;
+        }
+
+        if (!drwav__guid_equal(wave, drwavGUID_W64_WAVE)) {
+            return DRWAV_FALSE;
+        }
+    }
+
+
+    /* For RF64, the "ds64" chunk must come next, before the "fmt " chunk. */
+    if (pWav->container == drwav_container_rf64) {
+        drwav_uint8 sizeBytes[8];
+        drwav_uint64 bytesRemainingInChunk;
+        drwav_chunk_header header;
+        drwav_result result = drwav__read_chunk_header(pWav->onRead, pWav->pUserData, pWav->container, &cursor, &header);
+        if (result != DRWAV_SUCCESS) {
+            return DRWAV_FALSE;
+        }
+
+        if (!drwav__fourcc_equal(header.id.fourcc, "ds64")) {
+            return DRWAV_FALSE; /* Expecting "ds64". */
+        }
+
+        bytesRemainingInChunk = header.sizeInBytes + header.paddingSize;
+
+        /* We don't care about the size of the RIFF chunk - skip it. */
+        if (!drwav__seek_forward(pWav->onSeek, 8, pWav->pUserData)) {
+            return DRWAV_FALSE;
+        }
+        bytesRemainingInChunk -= 8;
+        cursor += 8;
+
+
+        /* Next 8 bytes is the size of the "data" chunk. */
+        if (drwav__on_read(pWav->onRead, pWav->pUserData, sizeBytes, sizeof(sizeBytes), &cursor) != sizeof(sizeBytes)) {
+            return DRWAV_FALSE;
+        }
+        bytesRemainingInChunk -= 8;
+        dataChunkSize = drwav__bytes_to_u64(sizeBytes);
+
+
+        /* Next 8 bytes is the same count which we would usually derived from the FACT chunk if it was available. */
+        if (drwav__on_read(pWav->onRead, pWav->pUserData, sizeBytes, sizeof(sizeBytes), &cursor) != sizeof(sizeBytes)) {
+            return DRWAV_FALSE;
+        }
+        bytesRemainingInChunk -= 8;
+        sampleCountFromFactChunk = drwav__bytes_to_u64(sizeBytes);
+
+
+        /* Skip over everything else. */
+        if (!drwav__seek_forward(pWav->onSeek, bytesRemainingInChunk, pWav->pUserData)) {
+            return DRWAV_FALSE;
+        }
+        cursor += bytesRemainingInChunk;
+    }
+
+
+    /* The next bytes should be the "fmt " chunk. */
+    if (!drwav__read_fmt(pWav->onRead, pWav->onSeek, pWav->pUserData, pWav->container, &cursor, &fmt)) {
+        return DRWAV_FALSE;    /* Failed to read the "fmt " chunk. */
+    }
+
+    /* Basic validation. */
+    if ((fmt.sampleRate    == 0 || fmt.sampleRate    > DRWAV_MAX_SAMPLE_RATE)     ||
+        (fmt.channels      == 0 || fmt.channels      > DRWAV_MAX_CHANNELS)        ||
+        (fmt.bitsPerSample == 0 || fmt.bitsPerSample > DRWAV_MAX_BITS_PER_SAMPLE) ||
+        fmt.blockAlign == 0) {
+        return DRWAV_FALSE; /* Probably an invalid WAV file. */
+    }
+
+
+    /* Translate the internal format. */
+    translatedFormatTag = fmt.formatTag;
+    if (translatedFormatTag == DR_WAVE_FORMAT_EXTENSIBLE) {
+        translatedFormatTag = drwav__bytes_to_u16(fmt.subFormat + 0);
+    }
+
+
+    /*
+    We need to enumerate over each chunk for two reasons:
+      1) The "data" chunk may not be the next one
+      2) We may want to report each chunk back to the client
+    
+    In order to correctly report each chunk back to the client we will need to keep looping until the end of the file.
+    */
+    foundDataChunk = DRWAV_FALSE;
+
+    /* The next chunk we care about is the "data" chunk. This is not necessarily the next chunk so we'll need to loop. */
+    for (;;)
+    {
+        drwav_chunk_header header;
+        drwav_result result = drwav__read_chunk_header(pWav->onRead, pWav->pUserData, pWav->container, &cursor, &header);
+        if (result != DRWAV_SUCCESS) {
+            if (!foundDataChunk) {
+                return DRWAV_FALSE;
+            } else {
+                break;  /* Probably at the end of the file. Get out of the loop. */
+            }
+        }
+
+        /* Tell the client about this chunk. */
+        if (!sequential && onChunk != NULL) {
+            drwav_uint64 callbackBytesRead = onChunk(pChunkUserData, pWav->onRead, pWav->onSeek, pWav->pUserData, &header, pWav->container, &fmt);
+
+            /*
+            dr_wav may need to read the contents of the chunk, so we now need to seek back to the position before
+            we called the callback.
+            */
+            if (callbackBytesRead > 0) {
+                if (!drwav__seek_from_start(pWav->onSeek, cursor, pWav->pUserData)) {
+                    return DRWAV_FALSE;
+                }
+            }
+        }
+        
+
+        if (!foundDataChunk) {
+            pWav->dataChunkDataPos = cursor;
+        }
+
+        chunkSize = header.sizeInBytes;
+        if (pWav->container == drwav_container_riff || pWav->container == drwav_container_rf64) {
+            if (drwav__fourcc_equal(header.id.fourcc, "data")) {
+                foundDataChunk = DRWAV_TRUE;
+                if (pWav->container != drwav_container_rf64) {  /* The data chunk size for RF64 will always be set to 0xFFFFFFFF here. It was set to it's true value earlier. */
+                    dataChunkSize = chunkSize;
+                }
+            }
+        } else {
+            if (drwav__guid_equal(header.id.guid, drwavGUID_W64_DATA)) {
+                foundDataChunk = DRWAV_TRUE;
+                dataChunkSize = chunkSize;
+            }
+        }
+
+        /*
+        If at this point we have found the data chunk and we're running in sequential mode, we need to break out of this loop. The reason for
+        this is that we would otherwise require a backwards seek which sequential mode forbids.
+        */
+        if (foundDataChunk && sequential) {
+            break;
+        }
+
+        /* Optional. Get the total sample count from the FACT chunk. This is useful for compressed formats. */
+        if (pWav->container == drwav_container_riff) {
+            if (drwav__fourcc_equal(header.id.fourcc, "fact")) {
+                drwav_uint32 sampleCount;
+                if (drwav__on_read(pWav->onRead, pWav->pUserData, &sampleCount, 4, &cursor) != 4) {
+                    return DRWAV_FALSE;
+                }
+                chunkSize -= 4;
+
+                if (!foundDataChunk) {
+                    pWav->dataChunkDataPos = cursor;
+                }
+
+                /*
+                The sample count in the "fact" chunk is either unreliable, or I'm not understanding it properly. For now I am only enabling this
+                for Microsoft ADPCM formats.
+                */
+                if (pWav->translatedFormatTag == DR_WAVE_FORMAT_ADPCM) {
+                    sampleCountFromFactChunk = sampleCount;
+                } else {
+                    sampleCountFromFactChunk = 0;
+                }
+            }
+        } else if (pWav->container == drwav_container_w64) {
+            if (drwav__guid_equal(header.id.guid, drwavGUID_W64_FACT)) {
+                if (drwav__on_read(pWav->onRead, pWav->pUserData, &sampleCountFromFactChunk, 8, &cursor) != 8) {
+                    return DRWAV_FALSE;
+                }
+                chunkSize -= 8;
+
+                if (!foundDataChunk) {
+                    pWav->dataChunkDataPos = cursor;
+                }
+            }
+        } else if (pWav->container == drwav_container_rf64) {
+            /* We retrieved the sample count from the ds64 chunk earlier so no need to do that here. */
+        }
+
+        /* "smpl" chunk. */
+        if (pWav->container == drwav_container_riff || pWav->container == drwav_container_rf64) {
+            if (drwav__fourcc_equal(header.id.fourcc, "smpl")) {
+                drwav_uint8 smplHeaderData[36];    /* 36 = size of the smpl header section, not including the loop data. */
+                if (chunkSize >= sizeof(smplHeaderData)) {
+                    drwav_uint64 bytesJustRead = drwav__on_read(pWav->onRead, pWav->pUserData, smplHeaderData, sizeof(smplHeaderData), &cursor);
+                    chunkSize -= bytesJustRead;
+
+                    if (bytesJustRead == sizeof(smplHeaderData)) {
+                        drwav_uint32 iLoop;
+
+                        pWav->smpl.manufacturer      = drwav__bytes_to_u32(smplHeaderData+0);
+                        pWav->smpl.product           = drwav__bytes_to_u32(smplHeaderData+4);
+                        pWav->smpl.samplePeriod      = drwav__bytes_to_u32(smplHeaderData+8);
+                        pWav->smpl.midiUnityNotes    = drwav__bytes_to_u32(smplHeaderData+12);
+                        pWav->smpl.midiPitchFraction = drwav__bytes_to_u32(smplHeaderData+16);
+                        pWav->smpl.smpteFormat       = drwav__bytes_to_u32(smplHeaderData+20);
+                        pWav->smpl.smpteOffset       = drwav__bytes_to_u32(smplHeaderData+24);
+                        pWav->smpl.numSampleLoops    = drwav__bytes_to_u32(smplHeaderData+28);
+                        pWav->smpl.samplerData       = drwav__bytes_to_u32(smplHeaderData+32);
+
+                        for (iLoop = 0; iLoop < pWav->smpl.numSampleLoops && iLoop < drwav_countof(pWav->smpl.loops); ++iLoop) {
+                            drwav_uint8 smplLoopData[24];  /* 24 = size of a loop section in the smpl chunk. */
+                            bytesJustRead = drwav__on_read(pWav->onRead, pWav->pUserData, smplLoopData, sizeof(smplLoopData), &cursor);
+                            chunkSize -= bytesJustRead;
+
+                            if (bytesJustRead == sizeof(smplLoopData)) {
+                                pWav->smpl.loops[iLoop].cuePointId = drwav__bytes_to_u32(smplLoopData+0);
+                                pWav->smpl.loops[iLoop].type       = drwav__bytes_to_u32(smplLoopData+4);
+                                pWav->smpl.loops[iLoop].start      = drwav__bytes_to_u32(smplLoopData+8);
+                                pWav->smpl.loops[iLoop].end        = drwav__bytes_to_u32(smplLoopData+12);
+                                pWav->smpl.loops[iLoop].fraction   = drwav__bytes_to_u32(smplLoopData+16);
+                                pWav->smpl.loops[iLoop].playCount  = drwav__bytes_to_u32(smplLoopData+20);
+                            } else {
+                                break;  /* Break from the smpl loop for loop. */
+                            }
+                        }
+                    }
+                } else {
+                    /* Looks like invalid data. Ignore the chunk. */
+                }
+            }
+        } else {
+            if (drwav__guid_equal(header.id.guid, drwavGUID_W64_SMPL)) {
+                /*
+                This path will be hit when a W64 WAV file contains a smpl chunk. I don't have a sample file to test this path, so a contribution
+                is welcome to add support for this.
+                */
+            }
+        }
+
+        /* Make sure we seek past the padding. */
+        chunkSize += header.paddingSize;
+        if (!drwav__seek_forward(pWav->onSeek, chunkSize, pWav->pUserData)) {
+            break;
+        }
+        cursor += chunkSize;
+
+        if (!foundDataChunk) {
+            pWav->dataChunkDataPos = cursor;
+        }
+    }
+
+    /* If we haven't found a data chunk, return an error. */
+    if (!foundDataChunk) {
+        return DRWAV_FALSE;
+    }
+
+    /* We may have moved passed the data chunk. If so we need to move back. If running in sequential mode we can assume we are already sitting on the data chunk. */
+    if (!sequential) {
+        if (!drwav__seek_from_start(pWav->onSeek, pWav->dataChunkDataPos, pWav->pUserData)) {
+            return DRWAV_FALSE;
+        }
+        cursor = pWav->dataChunkDataPos;
+    }
+    
+
+    /* At this point we should be sitting on the first byte of the raw audio data. */
+
+    pWav->fmt                 = fmt;
+    pWav->sampleRate          = fmt.sampleRate;
+    pWav->channels            = fmt.channels;
+    pWav->bitsPerSample       = fmt.bitsPerSample;
+    pWav->bytesRemaining      = dataChunkSize;
+    pWav->translatedFormatTag = translatedFormatTag;
+    pWav->dataChunkDataSize   = dataChunkSize;
+
+    if (sampleCountFromFactChunk != 0) {
+        pWav->totalPCMFrameCount = sampleCountFromFactChunk;
+    } else {
+        pWav->totalPCMFrameCount = dataChunkSize / drwav_get_bytes_per_pcm_frame(pWav);
+
+        if (pWav->translatedFormatTag == DR_WAVE_FORMAT_ADPCM) {
+            drwav_uint64 totalBlockHeaderSizeInBytes;
+            drwav_uint64 blockCount = dataChunkSize / fmt.blockAlign;
+
+            /* Make sure any trailing partial block is accounted for. */
+            if ((blockCount * fmt.blockAlign) < dataChunkSize) {
+                blockCount += 1;
+            }
+
+            /* We decode two samples per byte. There will be blockCount headers in the data chunk. This is enough to know how to calculate the total PCM frame count. */
+            totalBlockHeaderSizeInBytes = blockCount * (6*fmt.channels);
+            pWav->totalPCMFrameCount = ((dataChunkSize - totalBlockHeaderSizeInBytes) * 2) / fmt.channels;
+        }
+        if (pWav->translatedFormatTag == DR_WAVE_FORMAT_DVI_ADPCM) {
+            drwav_uint64 totalBlockHeaderSizeInBytes;
+            drwav_uint64 blockCount = dataChunkSize / fmt.blockAlign;
+
+            /* Make sure any trailing partial block is accounted for. */
+            if ((blockCount * fmt.blockAlign) < dataChunkSize) {
+                blockCount += 1;
+            }
+
+            /* We decode two samples per byte. There will be blockCount headers in the data chunk. This is enough to know how to calculate the total PCM frame count. */
+            totalBlockHeaderSizeInBytes = blockCount * (4*fmt.channels);
+            pWav->totalPCMFrameCount = ((dataChunkSize - totalBlockHeaderSizeInBytes) * 2) / fmt.channels;
+
+            /* The header includes a decoded sample for each channel which acts as the initial predictor sample. */
+            pWav->totalPCMFrameCount += blockCount;
+        }
+    }
+
+    /* Some formats only support a certain number of channels. */
+    if (pWav->translatedFormatTag == DR_WAVE_FORMAT_ADPCM || pWav->translatedFormatTag == DR_WAVE_FORMAT_DVI_ADPCM) {
+        if (pWav->channels > 2) {
+            return DRWAV_FALSE;
+        }
+    }
+
+#ifdef DR_WAV_LIBSNDFILE_COMPAT
+    /*
+    I use libsndfile as a benchmark for testing, however in the version I'm using (from the Windows installer on the libsndfile website),
+    it appears the total sample count libsndfile uses for MS-ADPCM is incorrect. It would seem they are computing the total sample count
+    from the number of blocks, however this results in the inclusion of extra silent samples at the end of the last block. The correct
+    way to know the total sample count is to inspect the "fact" chunk, which should always be present for compressed formats, and should
+    always include the sample count. This little block of code below is only used to emulate the libsndfile logic so I can properly run my
+    correctness tests against libsndfile, and is disabled by default.
+    */
+    if (pWav->translatedFormatTag == DR_WAVE_FORMAT_ADPCM) {
+        drwav_uint64 blockCount = dataChunkSize / fmt.blockAlign;
+        pWav->totalPCMFrameCount = (((blockCount * (fmt.blockAlign - (6*pWav->channels))) * 2)) / fmt.channels;  /* x2 because two samples per byte. */
+    }
+    if (pWav->translatedFormatTag == DR_WAVE_FORMAT_DVI_ADPCM) {
+        drwav_uint64 blockCount = dataChunkSize / fmt.blockAlign;
+        pWav->totalPCMFrameCount = (((blockCount * (fmt.blockAlign - (4*pWav->channels))) * 2) + (blockCount * pWav->channels)) / fmt.channels;
+    }
+#endif
+
+    return DRWAV_TRUE;
+}
+
+DRWAV_API drwav_bool32 drwav_init(drwav* pWav, drwav_read_proc onRead, drwav_seek_proc onSeek, void* pUserData, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    return drwav_init_ex(pWav, onRead, onSeek, NULL, pUserData, NULL, 0, pAllocationCallbacks);
+}
+
+DRWAV_API drwav_bool32 drwav_init_ex(drwav* pWav, drwav_read_proc onRead, drwav_seek_proc onSeek, drwav_chunk_proc onChunk, void* pReadSeekUserData, void* pChunkUserData, drwav_uint32 flags, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    if (!drwav_preinit(pWav, onRead, onSeek, pReadSeekUserData, pAllocationCallbacks)) {
+        return DRWAV_FALSE;
+    }
+
+    return drwav_init__internal(pWav, onChunk, pChunkUserData, flags);
+}
+
+
+static drwav_uint32 drwav__riff_chunk_size_riff(drwav_uint64 dataChunkSize)
+{
+    drwav_uint64 chunkSize = 4 + 24 + dataChunkSize + drwav__chunk_padding_size_riff(dataChunkSize); /* 4 = "WAVE". 24 = "fmt " chunk. */
+    if (chunkSize > 0xFFFFFFFFUL) {
+        chunkSize = 0xFFFFFFFFUL;
+    }
+
+    return (drwav_uint32)chunkSize; /* Safe cast due to the clamp above. */
+}
+
+static drwav_uint32 drwav__data_chunk_size_riff(drwav_uint64 dataChunkSize)
+{
+    if (dataChunkSize <= 0xFFFFFFFFUL) {
+        return (drwav_uint32)dataChunkSize;
+    } else {
+        return 0xFFFFFFFFUL;
+    }
+}
+
+static drwav_uint64 drwav__riff_chunk_size_w64(drwav_uint64 dataChunkSize)
+{
+    drwav_uint64 dataSubchunkPaddingSize = drwav__chunk_padding_size_w64(dataChunkSize);
+
+    return 80 + 24 + dataChunkSize + dataSubchunkPaddingSize;   /* +24 because W64 includes the size of the GUID and size fields. */
+}
+
+static drwav_uint64 drwav__data_chunk_size_w64(drwav_uint64 dataChunkSize)
+{
+    return 24 + dataChunkSize;        /* +24 because W64 includes the size of the GUID and size fields. */
+}
+
+static drwav_uint64 drwav__riff_chunk_size_rf64(drwav_uint64 dataChunkSize)
+{
+    drwav_uint64 chunkSize = 4 + 36 + 24 + dataChunkSize + drwav__chunk_padding_size_riff(dataChunkSize); /* 4 = "WAVE". 36 = "ds64" chunk. 24 = "fmt " chunk. */
+    if (chunkSize > 0xFFFFFFFFUL) {
+        chunkSize = 0xFFFFFFFFUL;
+    }
+
+    return chunkSize;
+}
+
+static drwav_uint64 drwav__data_chunk_size_rf64(drwav_uint64 dataChunkSize)
+{
+    return dataChunkSize;
+}
+
+
+static size_t drwav__write(drwav* pWav, const void* pData, size_t dataSize)
+{
+    DRWAV_ASSERT(pWav          != NULL);
+    DRWAV_ASSERT(pWav->onWrite != NULL);
+
+    /* Generic write. Assumes no byte reordering required. */
+    return pWav->onWrite(pWav->pUserData, pData, dataSize);
+}
+
+static size_t drwav__write_u16ne_to_le(drwav* pWav, drwav_uint16 value)
+{
+    DRWAV_ASSERT(pWav          != NULL);
+    DRWAV_ASSERT(pWav->onWrite != NULL);
+
+    if (!drwav__is_little_endian()) {
+        value = drwav__bswap16(value);
+    }
+
+    return drwav__write(pWav, &value, 2);
+}
+
+static size_t drwav__write_u32ne_to_le(drwav* pWav, drwav_uint32 value)
+{
+    DRWAV_ASSERT(pWav          != NULL);
+    DRWAV_ASSERT(pWav->onWrite != NULL);
+
+    if (!drwav__is_little_endian()) {
+        value = drwav__bswap32(value);
+    }
+
+    return drwav__write(pWav, &value, 4);
+}
+
+static size_t drwav__write_u64ne_to_le(drwav* pWav, drwav_uint64 value)
+{
+    DRWAV_ASSERT(pWav          != NULL);
+    DRWAV_ASSERT(pWav->onWrite != NULL);
+
+    if (!drwav__is_little_endian()) {
+        value = drwav__bswap64(value);
+    }
+
+    return drwav__write(pWav, &value, 8);
+}
+
+
+static drwav_bool32 drwav_preinit_write(drwav* pWav, const drwav_data_format* pFormat, drwav_bool32 isSequential, drwav_write_proc onWrite, drwav_seek_proc onSeek, void* pUserData, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    if (pWav == NULL || onWrite == NULL) {
+        return DRWAV_FALSE;
+    }
+
+    if (!isSequential && onSeek == NULL) {
+        return DRWAV_FALSE; /* <-- onSeek is required when in non-sequential mode. */
+    }
+
+    /* Not currently supporting compressed formats. Will need to add support for the "fact" chunk before we enable this. */
+    if (pFormat->format == DR_WAVE_FORMAT_EXTENSIBLE) {
+        return DRWAV_FALSE;
+    }
+    if (pFormat->format == DR_WAVE_FORMAT_ADPCM || pFormat->format == DR_WAVE_FORMAT_DVI_ADPCM) {
+        return DRWAV_FALSE;
+    }
+
+    DRWAV_ZERO_MEMORY(pWav, sizeof(*pWav));
+    pWav->onWrite   = onWrite;
+    pWav->onSeek    = onSeek;
+    pWav->pUserData = pUserData;
+    pWav->allocationCallbacks = drwav_copy_allocation_callbacks_or_defaults(pAllocationCallbacks);
+
+    if (pWav->allocationCallbacks.onFree == NULL || (pWav->allocationCallbacks.onMalloc == NULL && pWav->allocationCallbacks.onRealloc == NULL)) {
+        return DRWAV_FALSE;    /* Invalid allocation callbacks. */
+    }
+
+    pWav->fmt.formatTag = (drwav_uint16)pFormat->format;
+    pWav->fmt.channels = (drwav_uint16)pFormat->channels;
+    pWav->fmt.sampleRate = pFormat->sampleRate;
+    pWav->fmt.avgBytesPerSec = (drwav_uint32)((pFormat->bitsPerSample * pFormat->sampleRate * pFormat->channels) / 8);
+    pWav->fmt.blockAlign = (drwav_uint16)((pFormat->channels * pFormat->bitsPerSample) / 8);
+    pWav->fmt.bitsPerSample = (drwav_uint16)pFormat->bitsPerSample;
+    pWav->fmt.extendedSize = 0;
+    pWav->isSequentialWrite = isSequential;
+
+    return DRWAV_TRUE;
+}
+
+static drwav_bool32 drwav_init_write__internal(drwav* pWav, const drwav_data_format* pFormat, drwav_uint64 totalSampleCount)
+{
+    /* The function assumes drwav_preinit_write() was called beforehand. */
+
+    size_t runningPos = 0;
+    drwav_uint64 initialDataChunkSize = 0;
+    drwav_uint64 chunkSizeFMT;
+
+    /*
+    The initial values for the "RIFF" and "data" chunks depends on whether or not we are initializing in sequential mode or not. In
+    sequential mode we set this to its final values straight away since they can be calculated from the total sample count. In non-
+    sequential mode we initialize it all to zero and fill it out in drwav_uninit() using a backwards seek.
+    */
+    if (pWav->isSequentialWrite) {
+        initialDataChunkSize = (totalSampleCount * pWav->fmt.bitsPerSample) / 8;
+
+        /*
+        The RIFF container has a limit on the number of samples. drwav is not allowing this. There's no practical limits for Wave64
+        so for the sake of simplicity I'm not doing any validation for that.
+        */
+        if (pFormat->container == drwav_container_riff) {
+            if (initialDataChunkSize > (0xFFFFFFFFUL - 36)) {
+                return DRWAV_FALSE; /* Not enough room to store every sample. */
+            }
+        }
+    }
+
+    pWav->dataChunkDataSizeTargetWrite = initialDataChunkSize;
+
+
+    /* "RIFF" chunk. */
+    if (pFormat->container == drwav_container_riff) {
+        drwav_uint32 chunkSizeRIFF = 28 + (drwav_uint32)initialDataChunkSize;   /* +28 = "WAVE" + [sizeof "fmt " chunk] */
+        runningPos += drwav__write(pWav, "RIFF", 4);
+        runningPos += drwav__write_u32ne_to_le(pWav, chunkSizeRIFF);
+        runningPos += drwav__write(pWav, "WAVE", 4);
+    } else if (pFormat->container == drwav_container_w64) {
+        drwav_uint64 chunkSizeRIFF = 80 + 24 + initialDataChunkSize;            /* +24 because W64 includes the size of the GUID and size fields. */
+        runningPos += drwav__write(pWav, drwavGUID_W64_RIFF, 16);
+        runningPos += drwav__write_u64ne_to_le(pWav, chunkSizeRIFF);
+        runningPos += drwav__write(pWav, drwavGUID_W64_WAVE, 16);
+    } else if (pFormat->container == drwav_container_rf64) {
+        runningPos += drwav__write(pWav, "RF64", 4);
+        runningPos += drwav__write_u32ne_to_le(pWav, 0xFFFFFFFF);               /* Always 0xFFFFFFFF for RF64. Set to a proper value in the "ds64" chunk. */
+        runningPos += drwav__write(pWav, "WAVE", 4);
+    }
+
+    
+    /* "ds64" chunk (RF64 only). */
+    if (pFormat->container == drwav_container_rf64) {
+        drwav_uint32 initialds64ChunkSize = 28;                                 /* 28 = [Size of RIFF (8 bytes)] + [Size of DATA (8 bytes)] + [Sample Count (8 bytes)] + [Table Length (4 bytes)]. Table length always set to 0. */
+        drwav_uint64 initialRiffChunkSize = 8 + initialds64ChunkSize + initialDataChunkSize;    /* +8 for the ds64 header. */
+
+        runningPos += drwav__write(pWav, "ds64", 4);
+        runningPos += drwav__write_u32ne_to_le(pWav, initialds64ChunkSize);     /* Size of ds64. */
+        runningPos += drwav__write_u64ne_to_le(pWav, initialRiffChunkSize);     /* Size of RIFF. Set to true value at the end. */
+        runningPos += drwav__write_u64ne_to_le(pWav, initialDataChunkSize);     /* Size of DATA. Set to true value at the end. */
+        runningPos += drwav__write_u64ne_to_le(pWav, totalSampleCount);         /* Sample count. */
+        runningPos += drwav__write_u32ne_to_le(pWav, 0);                        /* Table length. Always set to zero in our case since we're not doing any other chunks than "DATA". */
+    }
+
+
+    /* "fmt " chunk. */
+    if (pFormat->container == drwav_container_riff || pFormat->container == drwav_container_rf64) {
+        chunkSizeFMT = 16;
+        runningPos += drwav__write(pWav, "fmt ", 4);
+        runningPos += drwav__write_u32ne_to_le(pWav, (drwav_uint32)chunkSizeFMT);
+    } else if (pFormat->container == drwav_container_w64) {
+        chunkSizeFMT = 40;
+        runningPos += drwav__write(pWav, drwavGUID_W64_FMT, 16);
+        runningPos += drwav__write_u64ne_to_le(pWav, chunkSizeFMT);
+    }
+
+    runningPos += drwav__write_u16ne_to_le(pWav, pWav->fmt.formatTag);
+    runningPos += drwav__write_u16ne_to_le(pWav, pWav->fmt.channels);
+    runningPos += drwav__write_u32ne_to_le(pWav, pWav->fmt.sampleRate);
+    runningPos += drwav__write_u32ne_to_le(pWav, pWav->fmt.avgBytesPerSec);
+    runningPos += drwav__write_u16ne_to_le(pWav, pWav->fmt.blockAlign);
+    runningPos += drwav__write_u16ne_to_le(pWav, pWav->fmt.bitsPerSample);
+
+    pWav->dataChunkDataPos = runningPos;
+
+    /* "data" chunk. */
+    if (pFormat->container == drwav_container_riff) {
+        drwav_uint32 chunkSizeDATA = (drwav_uint32)initialDataChunkSize;
+        runningPos += drwav__write(pWav, "data", 4);
+        runningPos += drwav__write_u32ne_to_le(pWav, chunkSizeDATA);
+    } else if (pFormat->container == drwav_container_w64) {
+        drwav_uint64 chunkSizeDATA = 24 + initialDataChunkSize;     /* +24 because W64 includes the size of the GUID and size fields. */
+        runningPos += drwav__write(pWav, drwavGUID_W64_DATA, 16);
+        runningPos += drwav__write_u64ne_to_le(pWav, chunkSizeDATA);
+    } else if (pFormat->container == drwav_container_rf64) {
+        runningPos += drwav__write(pWav, "data", 4);
+        runningPos += drwav__write_u32ne_to_le(pWav, 0xFFFFFFFF);   /* Always set to 0xFFFFFFFF for RF64. The true size of the data chunk is specified in the ds64 chunk. */
+    }
+
+    /*
+    The runningPos variable is incremented in the section above but is left unused which is causing some static analysis tools to detect it
+    as a dead store. I'm leaving this as-is for safety just in case I want to expand this function later to include other tags and want to
+    keep track of the running position for whatever reason. The line below should silence the static analysis tools.
+    */
+    (void)runningPos;
+
+    /* Set some properties for the client's convenience. */
+    pWav->container = pFormat->container;
+    pWav->channels = (drwav_uint16)pFormat->channels;
+    pWav->sampleRate = pFormat->sampleRate;
+    pWav->bitsPerSample = (drwav_uint16)pFormat->bitsPerSample;
+    pWav->translatedFormatTag = (drwav_uint16)pFormat->format;
+
+    return DRWAV_TRUE;
+}
+
+
+DRWAV_API drwav_bool32 drwav_init_write(drwav* pWav, const drwav_data_format* pFormat, drwav_write_proc onWrite, drwav_seek_proc onSeek, void* pUserData, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    if (!drwav_preinit_write(pWav, pFormat, DRWAV_FALSE, onWrite, onSeek, pUserData, pAllocationCallbacks)) {
+        return DRWAV_FALSE;
+    }
+
+    return drwav_init_write__internal(pWav, pFormat, 0);               /* DRWAV_FALSE = Not Sequential */
+}
+
+DRWAV_API drwav_bool32 drwav_init_write_sequential(drwav* pWav, const drwav_data_format* pFormat, drwav_uint64 totalSampleCount, drwav_write_proc onWrite, void* pUserData, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    if (!drwav_preinit_write(pWav, pFormat, DRWAV_TRUE, onWrite, NULL, pUserData, pAllocationCallbacks)) {
+        return DRWAV_FALSE;
+    }
+
+    return drwav_init_write__internal(pWav, pFormat, totalSampleCount); /* DRWAV_TRUE = Sequential */
+}
+
+DRWAV_API drwav_bool32 drwav_init_write_sequential_pcm_frames(drwav* pWav, const drwav_data_format* pFormat, drwav_uint64 totalPCMFrameCount, drwav_write_proc onWrite, void* pUserData, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    if (pFormat == NULL) {
+        return DRWAV_FALSE;
+    }
+
+    return drwav_init_write_sequential(pWav, pFormat, totalPCMFrameCount*pFormat->channels, onWrite, pUserData, pAllocationCallbacks);
+}
+
+DRWAV_API drwav_uint64 drwav_target_write_size_bytes(const drwav_data_format* pFormat, drwav_uint64 totalSampleCount)
+{
+    /* Casting totalSampleCount to drwav_int64 for VC6 compatibility. No issues in practice because nobody is going to exhaust the whole 63 bits. */
+    drwav_uint64 targetDataSizeBytes = (drwav_uint64)((drwav_int64)totalSampleCount * pFormat->channels * pFormat->bitsPerSample/8.0);
+    drwav_uint64 riffChunkSizeBytes;
+    drwav_uint64 fileSizeBytes = 0;
+
+    if (pFormat->container == drwav_container_riff) {
+        riffChunkSizeBytes = drwav__riff_chunk_size_riff(targetDataSizeBytes);
+        fileSizeBytes = (8 + riffChunkSizeBytes);   /* +8 because WAV doesn't include the size of the ChunkID and ChunkSize fields. */
+    } else if (pFormat->container == drwav_container_w64) {
+        riffChunkSizeBytes = drwav__riff_chunk_size_w64(targetDataSizeBytes);
+        fileSizeBytes = riffChunkSizeBytes;
+    } else if (pFormat->container == drwav_container_rf64) {
+        riffChunkSizeBytes = drwav__riff_chunk_size_rf64(targetDataSizeBytes);
+        fileSizeBytes = (8 + riffChunkSizeBytes);   /* +8 because WAV doesn't include the size of the ChunkID and ChunkSize fields. */
+    }
+
+    return fileSizeBytes;
+}
+
+
+#ifndef DR_WAV_NO_STDIO
+
+/* drwav_result_from_errno() is only used for fopen() and wfopen() so putting it inside DR_WAV_NO_STDIO for now. If something else needs this later we can move it out. */
+#include <errno.h>
+static drwav_result drwav_result_from_errno(int e)
+{
+    switch (e)
+    {
+        case 0: return DRWAV_SUCCESS;
+    #ifdef EPERM
+        case EPERM: return DRWAV_INVALID_OPERATION;
+    #endif
+    #ifdef ENOENT
+        case ENOENT: return DRWAV_DOES_NOT_EXIST;
+    #endif
+    #ifdef ESRCH
+        case ESRCH: return DRWAV_DOES_NOT_EXIST;
+    #endif
+    #ifdef EINTR
+        case EINTR: return DRWAV_INTERRUPT;
+    #endif
+    #ifdef EIO
+        case EIO: return DRWAV_IO_ERROR;
+    #endif
+    #ifdef ENXIO
+        case ENXIO: return DRWAV_DOES_NOT_EXIST;
+    #endif
+    #ifdef E2BIG
+        case E2BIG: return DRWAV_INVALID_ARGS;
+    #endif
+    #ifdef ENOEXEC
+        case ENOEXEC: return DRWAV_INVALID_FILE;
+    #endif
+    #ifdef EBADF
+        case EBADF: return DRWAV_INVALID_FILE;
+    #endif
+    #ifdef ECHILD
+        case ECHILD: return DRWAV_ERROR;
+    #endif
+    #ifdef EAGAIN
+        case EAGAIN: return DRWAV_UNAVAILABLE;
+    #endif
+    #ifdef ENOMEM
+        case ENOMEM: return DRWAV_OUT_OF_MEMORY;
+    #endif
+    #ifdef EACCES
+        case EACCES: return DRWAV_ACCESS_DENIED;
+    #endif
+    #ifdef EFAULT
+        case EFAULT: return DRWAV_BAD_ADDRESS;
+    #endif
+    #ifdef ENOTBLK
+        case ENOTBLK: return DRWAV_ERROR;
+    #endif
+    #ifdef EBUSY
+        case EBUSY: return DRWAV_BUSY;
+    #endif
+    #ifdef EEXIST
+        case EEXIST: return DRWAV_ALREADY_EXISTS;
+    #endif
+    #ifdef EXDEV
+        case EXDEV: return DRWAV_ERROR;
+    #endif
+    #ifdef ENODEV
+        case ENODEV: return DRWAV_DOES_NOT_EXIST;
+    #endif
+    #ifdef ENOTDIR
+        case ENOTDIR: return DRWAV_NOT_DIRECTORY;
+    #endif
+    #ifdef EISDIR
+        case EISDIR: return DRWAV_IS_DIRECTORY;
+    #endif
+    #ifdef EINVAL
+        case EINVAL: return DRWAV_INVALID_ARGS;
+    #endif
+    #ifdef ENFILE
+        case ENFILE: return DRWAV_TOO_MANY_OPEN_FILES;
+    #endif
+    #ifdef EMFILE
+        case EMFILE: return DRWAV_TOO_MANY_OPEN_FILES;
+    #endif
+    #ifdef ENOTTY
+        case ENOTTY: return DRWAV_INVALID_OPERATION;
+    #endif
+    #ifdef ETXTBSY
+        case ETXTBSY: return DRWAV_BUSY;
+    #endif
+    #ifdef EFBIG
+        case EFBIG: return DRWAV_TOO_BIG;
+    #endif
+    #ifdef ENOSPC
+        case ENOSPC: return DRWAV_NO_SPACE;
+    #endif
+    #ifdef ESPIPE
+        case ESPIPE: return DRWAV_BAD_SEEK;
+    #endif
+    #ifdef EROFS
+        case EROFS: return DRWAV_ACCESS_DENIED;
+    #endif
+    #ifdef EMLINK
+        case EMLINK: return DRWAV_TOO_MANY_LINKS;
+    #endif
+    #ifdef EPIPE
+        case EPIPE: return DRWAV_BAD_PIPE;
+    #endif
+    #ifdef EDOM
+        case EDOM: return DRWAV_OUT_OF_RANGE;
+    #endif
+    #ifdef ERANGE
+        case ERANGE: return DRWAV_OUT_OF_RANGE;
+    #endif
+    #ifdef EDEADLK
+        case EDEADLK: return DRWAV_DEADLOCK;
+    #endif
+    #ifdef ENAMETOOLONG
+        case ENAMETOOLONG: return DRWAV_PATH_TOO_LONG;
+    #endif
+    #ifdef ENOLCK
+        case ENOLCK: return DRWAV_ERROR;
+    #endif
+    #ifdef ENOSYS
+        case ENOSYS: return DRWAV_NOT_IMPLEMENTED;
+    #endif
+    #ifdef ENOTEMPTY
+        case ENOTEMPTY: return DRWAV_DIRECTORY_NOT_EMPTY;
+    #endif
+    #ifdef ELOOP
+        case ELOOP: return DRWAV_TOO_MANY_LINKS;
+    #endif
+    #ifdef ENOMSG
+        case ENOMSG: return DRWAV_NO_MESSAGE;
+    #endif
+    #ifdef EIDRM
+        case EIDRM: return DRWAV_ERROR;
+    #endif
+    #ifdef ECHRNG
+        case ECHRNG: return DRWAV_ERROR;
+    #endif
+    #ifdef EL2NSYNC
+        case EL2NSYNC: return DRWAV_ERROR;
+    #endif
+    #ifdef EL3HLT
+        case EL3HLT: return DRWAV_ERROR;
+    #endif
+    #ifdef EL3RST
+        case EL3RST: return DRWAV_ERROR;
+    #endif
+    #ifdef ELNRNG
+        case ELNRNG: return DRWAV_OUT_OF_RANGE;
+    #endif
+    #ifdef EUNATCH
+        case EUNATCH: return DRWAV_ERROR;
+    #endif
+    #ifdef ENOCSI
+        case ENOCSI: return DRWAV_ERROR;
+    #endif
+    #ifdef EL2HLT
+        case EL2HLT: return DRWAV_ERROR;
+    #endif
+    #ifdef EBADE
+        case EBADE: return DRWAV_ERROR;
+    #endif
+    #ifdef EBADR
+        case EBADR: return DRWAV_ERROR;
+    #endif
+    #ifdef EXFULL
+        case EXFULL: return DRWAV_ERROR;
+    #endif
+    #ifdef ENOANO
+        case ENOANO: return DRWAV_ERROR;
+    #endif
+    #ifdef EBADRQC
+        case EBADRQC: return DRWAV_ERROR;
+    #endif
+    #ifdef EBADSLT
+        case EBADSLT: return DRWAV_ERROR;
+    #endif
+    #ifdef EBFONT
+        case EBFONT: return DRWAV_INVALID_FILE;
+    #endif
+    #ifdef ENOSTR
+        case ENOSTR: return DRWAV_ERROR;
+    #endif
+    #ifdef ENODATA
+        case ENODATA: return DRWAV_NO_DATA_AVAILABLE;
+    #endif
+    #ifdef ETIME
+        case ETIME: return DRWAV_TIMEOUT;
+    #endif
+    #ifdef ENOSR
+        case ENOSR: return DRWAV_NO_DATA_AVAILABLE;
+    #endif
+    #ifdef ENONET
+        case ENONET: return DRWAV_NO_NETWORK;
+    #endif
+    #ifdef ENOPKG
+        case ENOPKG: return DRWAV_ERROR;
+    #endif
+    #ifdef EREMOTE
+        case EREMOTE: return DRWAV_ERROR;
+    #endif
+    #ifdef ENOLINK
+        case ENOLINK: return DRWAV_ERROR;
+    #endif
+    #ifdef EADV
+        case EADV: return DRWAV_ERROR;
+    #endif
+    #ifdef ESRMNT
+        case ESRMNT: return DRWAV_ERROR;
+    #endif
+    #ifdef ECOMM
+        case ECOMM: return DRWAV_ERROR;
+    #endif
+    #ifdef EPROTO
+        case EPROTO: return DRWAV_ERROR;
+    #endif
+    #ifdef EMULTIHOP
+        case EMULTIHOP: return DRWAV_ERROR;
+    #endif
+    #ifdef EDOTDOT
+        case EDOTDOT: return DRWAV_ERROR;
+    #endif
+    #ifdef EBADMSG
+        case EBADMSG: return DRWAV_BAD_MESSAGE;
+    #endif
+    #ifdef EOVERFLOW
+        case EOVERFLOW: return DRWAV_TOO_BIG;
+    #endif
+    #ifdef ENOTUNIQ
+        case ENOTUNIQ: return DRWAV_NOT_UNIQUE;
+    #endif
+    #ifdef EBADFD
+        case EBADFD: return DRWAV_ERROR;
+    #endif
+    #ifdef EREMCHG
+        case EREMCHG: return DRWAV_ERROR;
+    #endif
+    #ifdef ELIBACC
+        case ELIBACC: return DRWAV_ACCESS_DENIED;
+    #endif
+    #ifdef ELIBBAD
+        case ELIBBAD: return DRWAV_INVALID_FILE;
+    #endif
+    #ifdef ELIBSCN
+        case ELIBSCN: return DRWAV_INVALID_FILE;
+    #endif
+    #ifdef ELIBMAX
+        case ELIBMAX: return DRWAV_ERROR;
+    #endif
+    #ifdef ELIBEXEC
+        case ELIBEXEC: return DRWAV_ERROR;
+    #endif
+    #ifdef EILSEQ
+        case EILSEQ: return DRWAV_INVALID_DATA;
+    #endif
+    #ifdef ERESTART
+        case ERESTART: return DRWAV_ERROR;
+    #endif
+    #ifdef ESTRPIPE
+        case ESTRPIPE: return DRWAV_ERROR;
+    #endif
+    #ifdef EUSERS
+        case EUSERS: return DRWAV_ERROR;
+    #endif
+    #ifdef ENOTSOCK
+        case ENOTSOCK: return DRWAV_NOT_SOCKET;
+    #endif
+    #ifdef EDESTADDRREQ
+        case EDESTADDRREQ: return DRWAV_NO_ADDRESS;
+    #endif
+    #ifdef EMSGSIZE
+        case EMSGSIZE: return DRWAV_TOO_BIG;
+    #endif
+    #ifdef EPROTOTYPE
+        case EPROTOTYPE: return DRWAV_BAD_PROTOCOL;
+    #endif
+    #ifdef ENOPROTOOPT
+        case ENOPROTOOPT: return DRWAV_PROTOCOL_UNAVAILABLE;
+    #endif
+    #ifdef EPROTONOSUPPORT
+        case EPROTONOSUPPORT: return DRWAV_PROTOCOL_NOT_SUPPORTED;
+    #endif
+    #ifdef ESOCKTNOSUPPORT
+        case ESOCKTNOSUPPORT: return DRWAV_SOCKET_NOT_SUPPORTED;
+    #endif
+    #ifdef EOPNOTSUPP
+        case EOPNOTSUPP: return DRWAV_INVALID_OPERATION;
+    #endif
+    #ifdef EPFNOSUPPORT
+        case EPFNOSUPPORT: return DRWAV_PROTOCOL_FAMILY_NOT_SUPPORTED;
+    #endif
+    #ifdef EAFNOSUPPORT
+        case EAFNOSUPPORT: return DRWAV_ADDRESS_FAMILY_NOT_SUPPORTED;
+    #endif
+    #ifdef EADDRINUSE
+        case EADDRINUSE: return DRWAV_ALREADY_IN_USE;
+    #endif
+    #ifdef EADDRNOTAVAIL
+        case EADDRNOTAVAIL: return DRWAV_ERROR;
+    #endif
+    #ifdef ENETDOWN
+        case ENETDOWN: return DRWAV_NO_NETWORK;
+    #endif
+    #ifdef ENETUNREACH
+        case ENETUNREACH: return DRWAV_NO_NETWORK;
+    #endif
+    #ifdef ENETRESET
+        case ENETRESET: return DRWAV_NO_NETWORK;
+    #endif
+    #ifdef ECONNABORTED
+        case ECONNABORTED: return DRWAV_NO_NETWORK;
+    #endif
+    #ifdef ECONNRESET
+        case ECONNRESET: return DRWAV_CONNECTION_RESET;
+    #endif
+    #ifdef ENOBUFS
+        case ENOBUFS: return DRWAV_NO_SPACE;
+    #endif
+    #ifdef EISCONN
+        case EISCONN: return DRWAV_ALREADY_CONNECTED;
+    #endif
+    #ifdef ENOTCONN
+        case ENOTCONN: return DRWAV_NOT_CONNECTED;
+    #endif
+    #ifdef ESHUTDOWN
+        case ESHUTDOWN: return DRWAV_ERROR;
+    #endif
+    #ifdef ETOOMANYREFS
+        case ETOOMANYREFS: return DRWAV_ERROR;
+    #endif
+    #ifdef ETIMEDOUT
+        case ETIMEDOUT: return DRWAV_TIMEOUT;
+    #endif
+    #ifdef ECONNREFUSED
+        case ECONNREFUSED: return DRWAV_CONNECTION_REFUSED;
+    #endif
+    #ifdef EHOSTDOWN
+        case EHOSTDOWN: return DRWAV_NO_HOST;
+    #endif
+    #ifdef EHOSTUNREACH
+        case EHOSTUNREACH: return DRWAV_NO_HOST;
+    #endif
+    #ifdef EALREADY
+        case EALREADY: return DRWAV_IN_PROGRESS;
+    #endif
+    #ifdef EINPROGRESS
+        case EINPROGRESS: return DRWAV_IN_PROGRESS;
+    #endif
+    #ifdef ESTALE
+        case ESTALE: return DRWAV_INVALID_FILE;
+    #endif
+    #ifdef EUCLEAN
+        case EUCLEAN: return DRWAV_ERROR;
+    #endif
+    #ifdef ENOTNAM
+        case ENOTNAM: return DRWAV_ERROR;
+    #endif
+    #ifdef ENAVAIL
+        case ENAVAIL: return DRWAV_ERROR;
+    #endif
+    #ifdef EISNAM
+        case EISNAM: return DRWAV_ERROR;
+    #endif
+    #ifdef EREMOTEIO
+        case EREMOTEIO: return DRWAV_IO_ERROR;
+    #endif
+    #ifdef EDQUOT
+        case EDQUOT: return DRWAV_NO_SPACE;
+    #endif
+    #ifdef ENOMEDIUM
+        case ENOMEDIUM: return DRWAV_DOES_NOT_EXIST;
+    #endif
+    #ifdef EMEDIUMTYPE
+        case EMEDIUMTYPE: return DRWAV_ERROR;
+    #endif
+    #ifdef ECANCELED
+        case ECANCELED: return DRWAV_CANCELLED;
+    #endif
+    #ifdef ENOKEY
+        case ENOKEY: return DRWAV_ERROR;
+    #endif
+    #ifdef EKEYEXPIRED
+        case EKEYEXPIRED: return DRWAV_ERROR;
+    #endif
+    #ifdef EKEYREVOKED
+        case EKEYREVOKED: return DRWAV_ERROR;
+    #endif
+    #ifdef EKEYREJECTED
+        case EKEYREJECTED: return DRWAV_ERROR;
+    #endif
+    #ifdef EOWNERDEAD
+        case EOWNERDEAD: return DRWAV_ERROR;
+    #endif
+    #ifdef ENOTRECOVERABLE
+        case ENOTRECOVERABLE: return DRWAV_ERROR;
+    #endif
+    #ifdef ERFKILL
+        case ERFKILL: return DRWAV_ERROR;
+    #endif
+    #ifdef EHWPOISON
+        case EHWPOISON: return DRWAV_ERROR;
+    #endif
+        default: return DRWAV_ERROR;
+    }
+}
+
+static drwav_result drwav_fopen(FILE** ppFile, const char* pFilePath, const char* pOpenMode)
+{
+#if _MSC_VER && _MSC_VER >= 1400
+    errno_t err;
+#endif
+
+    if (ppFile != NULL) {
+        *ppFile = NULL;  /* Safety. */
+    }
+
+    if (pFilePath == NULL || pOpenMode == NULL || ppFile == NULL) {
+        return DRWAV_INVALID_ARGS;
+    }
+
+#if _MSC_VER && _MSC_VER >= 1400
+    err = fopen_s(ppFile, pFilePath, pOpenMode);
+    if (err != 0) {
+        return drwav_result_from_errno(err);
+    }
+#else
+#if defined(_WIN32) || defined(__APPLE__)
+    *ppFile = fopen(pFilePath, pOpenMode);
+#else
+    #if defined(_FILE_OFFSET_BITS) && _FILE_OFFSET_BITS == 64 && defined(_LARGEFILE64_SOURCE)
+        *ppFile = fopen64(pFilePath, pOpenMode);
+    #else
+        *ppFile = fopen(pFilePath, pOpenMode);
+    #endif
+#endif
+    if (*ppFile == NULL) {
+        drwav_result result = drwav_result_from_errno(errno);
+        if (result == DRWAV_SUCCESS) {
+            result = DRWAV_ERROR;   /* Just a safety check to make sure we never ever return success when pFile == NULL. */
+        }
+
+        return result;
+    }
+#endif
+
+    return DRWAV_SUCCESS;
+}
+
+/*
+_wfopen() isn't always available in all compilation environments.
+
+    * Windows only.
+    * MSVC seems to support it universally as far back as VC6 from what I can tell (haven't checked further back).
+    * MinGW-64 (both 32- and 64-bit) seems to support it.
+    * MinGW wraps it in !defined(__STRICT_ANSI__).
+    * OpenWatcom wraps it in !defined(_NO_EXT_KEYS).
+
+This can be reviewed as compatibility issues arise. The preference is to use _wfopen_s() and _wfopen() as opposed to the wcsrtombs()
+fallback, so if you notice your compiler not detecting this properly I'm happy to look at adding support.
+*/
+#if defined(_WIN32)
+    #if defined(_MSC_VER) || defined(__MINGW64__) || (!defined(__STRICT_ANSI__) && !defined(_NO_EXT_KEYS))
+        #define DRWAV_HAS_WFOPEN
+    #endif
+#endif
+
+static drwav_result drwav_wfopen(FILE** ppFile, const wchar_t* pFilePath, const wchar_t* pOpenMode, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    if (ppFile != NULL) {
+        *ppFile = NULL;  /* Safety. */
+    }
+
+    if (pFilePath == NULL || pOpenMode == NULL || ppFile == NULL) {
+        return DRWAV_INVALID_ARGS;
+    }
+
+#if defined(DRWAV_HAS_WFOPEN)
+    {
+        /* Use _wfopen() on Windows. */
+    #if defined(_MSC_VER) && _MSC_VER >= 1400
+        errno_t err = _wfopen_s(ppFile, pFilePath, pOpenMode);
+        if (err != 0) {
+            return drwav_result_from_errno(err);
+        }
+    #else
+        *ppFile = _wfopen(pFilePath, pOpenMode);
+        if (*ppFile == NULL) {
+            return drwav_result_from_errno(errno);
+        }
+    #endif
+        (void)pAllocationCallbacks;
+    }
+#else
+    /*
+    Use fopen() on anything other than Windows. Requires a conversion. This is annoying because fopen() is locale specific. The only real way I can
+    think of to do this is with wcsrtombs(). Note that wcstombs() is apparently not thread-safe because it uses a static global mbstate_t object for
+    maintaining state. I've checked this with -std=c89 and it works, but if somebody get's a compiler error I'll look into improving compatibility.
+    */
+    {
+        mbstate_t mbs;
+        size_t lenMB;
+        const wchar_t* pFilePathTemp = pFilePath;
+        char* pFilePathMB = NULL;
+        char pOpenModeMB[32] = {0};
+
+        /* Get the length first. */
+        DRWAV_ZERO_OBJECT(&mbs);
+        lenMB = wcsrtombs(NULL, &pFilePathTemp, 0, &mbs);
+        if (lenMB == (size_t)-1) {
+            return drwav_result_from_errno(errno);
+        }
+
+        pFilePathMB = (char*)drwav__malloc_from_callbacks(lenMB + 1, pAllocationCallbacks);
+        if (pFilePathMB == NULL) {
+            return DRWAV_OUT_OF_MEMORY;
+        }
+
+        pFilePathTemp = pFilePath;
+        DRWAV_ZERO_OBJECT(&mbs);
+        wcsrtombs(pFilePathMB, &pFilePathTemp, lenMB + 1, &mbs);
+
+        /* The open mode should always consist of ASCII characters so we should be able to do a trivial conversion. */
+        {
+            size_t i = 0;
+            for (;;) {
+                if (pOpenMode[i] == 0) {
+                    pOpenModeMB[i] = '\0';
+                    break;
+                }
+
+                pOpenModeMB[i] = (char)pOpenMode[i];
+                i += 1;
+            }
+        }
+
+        *ppFile = fopen(pFilePathMB, pOpenModeMB);
+
+        drwav__free_from_callbacks(pFilePathMB, pAllocationCallbacks);
+    }
+
+    if (*ppFile == NULL) {
+        return DRWAV_ERROR;
+    }
+#endif
+
+    return DRWAV_SUCCESS;
+}
+
+
+static size_t drwav__on_read_stdio(void* pUserData, void* pBufferOut, size_t bytesToRead)
+{
+    return fread(pBufferOut, 1, bytesToRead, (FILE*)pUserData);
+}
+
+static size_t drwav__on_write_stdio(void* pUserData, const void* pData, size_t bytesToWrite)
+{
+    return fwrite(pData, 1, bytesToWrite, (FILE*)pUserData);
+}
+
+static drwav_bool32 drwav__on_seek_stdio(void* pUserData, int offset, drwav_seek_origin origin)
+{
+    return fseek((FILE*)pUserData, offset, (origin == drwav_seek_origin_current) ? SEEK_CUR : SEEK_SET) == 0;
+}
+
+DRWAV_API drwav_bool32 drwav_init_file(drwav* pWav, const char* filename, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    return drwav_init_file_ex(pWav, filename, NULL, NULL, 0, pAllocationCallbacks);
+}
+
+
+static drwav_bool32 drwav_init_file__internal_FILE(drwav* pWav, FILE* pFile, drwav_chunk_proc onChunk, void* pChunkUserData, drwav_uint32 flags, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    drwav_bool32 result;
+
+    result = drwav_preinit(pWav, drwav__on_read_stdio, drwav__on_seek_stdio, (void*)pFile, pAllocationCallbacks);
+    if (result != DRWAV_TRUE) {
+        fclose(pFile);
+        return result;
+    }
+
+    result = drwav_init__internal(pWav, onChunk, pChunkUserData, flags);
+    if (result != DRWAV_TRUE) {
+        fclose(pFile);
+        return result;
+    }
+
+    return DRWAV_TRUE;
+}
+
+DRWAV_API drwav_bool32 drwav_init_file_ex(drwav* pWav, const char* filename, drwav_chunk_proc onChunk, void* pChunkUserData, drwav_uint32 flags, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    FILE* pFile;
+    if (drwav_fopen(&pFile, filename, "rb") != DRWAV_SUCCESS) {
+        return DRWAV_FALSE;
+    }
+
+    /* This takes ownership of the FILE* object. */
+    return drwav_init_file__internal_FILE(pWav, pFile, onChunk, pChunkUserData, flags, pAllocationCallbacks);
+}
+
+DRWAV_API drwav_bool32 drwav_init_file_w(drwav* pWav, const wchar_t* filename, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    return drwav_init_file_ex_w(pWav, filename, NULL, NULL, 0, pAllocationCallbacks);
+}
+
+DRWAV_API drwav_bool32 drwav_init_file_ex_w(drwav* pWav, const wchar_t* filename, drwav_chunk_proc onChunk, void* pChunkUserData, drwav_uint32 flags, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    FILE* pFile;
+    if (drwav_wfopen(&pFile, filename, L"rb", pAllocationCallbacks) != DRWAV_SUCCESS) {
+        return DRWAV_FALSE;
+    }
+
+    /* This takes ownership of the FILE* object. */
+    return drwav_init_file__internal_FILE(pWav, pFile, onChunk, pChunkUserData, flags, pAllocationCallbacks);
+}
+
+
+static drwav_bool32 drwav_init_file_write__internal_FILE(drwav* pWav, FILE* pFile, const drwav_data_format* pFormat, drwav_uint64 totalSampleCount, drwav_bool32 isSequential, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    drwav_bool32 result;
+
+    result = drwav_preinit_write(pWav, pFormat, isSequential, drwav__on_write_stdio, drwav__on_seek_stdio, (void*)pFile, pAllocationCallbacks);
+    if (result != DRWAV_TRUE) {
+        fclose(pFile);
+        return result;
+    }
+
+    result = drwav_init_write__internal(pWav, pFormat, totalSampleCount);
+    if (result != DRWAV_TRUE) {
+        fclose(pFile);
+        return result;
+    }
+
+    return DRWAV_TRUE;
+}
+
+static drwav_bool32 drwav_init_file_write__internal(drwav* pWav, const char* filename, const drwav_data_format* pFormat, drwav_uint64 totalSampleCount, drwav_bool32 isSequential, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    FILE* pFile;
+    if (drwav_fopen(&pFile, filename, "wb") != DRWAV_SUCCESS) {
+        return DRWAV_FALSE;
+    }
+
+    /* This takes ownership of the FILE* object. */
+    return drwav_init_file_write__internal_FILE(pWav, pFile, pFormat, totalSampleCount, isSequential, pAllocationCallbacks);
+}
+
+static drwav_bool32 drwav_init_file_write_w__internal(drwav* pWav, const wchar_t* filename, const drwav_data_format* pFormat, drwav_uint64 totalSampleCount, drwav_bool32 isSequential, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    FILE* pFile;
+    if (drwav_wfopen(&pFile, filename, L"wb", pAllocationCallbacks) != DRWAV_SUCCESS) {
+        return DRWAV_FALSE;
+    }
+
+    /* This takes ownership of the FILE* object. */
+    return drwav_init_file_write__internal_FILE(pWav, pFile, pFormat, totalSampleCount, isSequential, pAllocationCallbacks);
+}
+
+DRWAV_API drwav_bool32 drwav_init_file_write(drwav* pWav, const char* filename, const drwav_data_format* pFormat, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    return drwav_init_file_write__internal(pWav, filename, pFormat, 0, DRWAV_FALSE, pAllocationCallbacks);
+}
+
+DRWAV_API drwav_bool32 drwav_init_file_write_sequential(drwav* pWav, const char* filename, const drwav_data_format* pFormat, drwav_uint64 totalSampleCount, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    return drwav_init_file_write__internal(pWav, filename, pFormat, totalSampleCount, DRWAV_TRUE, pAllocationCallbacks);
+}
+
+DRWAV_API drwav_bool32 drwav_init_file_write_sequential_pcm_frames(drwav* pWav, const char* filename, const drwav_data_format* pFormat, drwav_uint64 totalPCMFrameCount, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    if (pFormat == NULL) {
+        return DRWAV_FALSE;
+    }
+
+    return drwav_init_file_write_sequential(pWav, filename, pFormat, totalPCMFrameCount*pFormat->channels, pAllocationCallbacks);
+}
+
+DRWAV_API drwav_bool32 drwav_init_file_write_w(drwav* pWav, const wchar_t* filename, const drwav_data_format* pFormat, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    return drwav_init_file_write_w__internal(pWav, filename, pFormat, 0, DRWAV_FALSE, pAllocationCallbacks);
+}
+
+DRWAV_API drwav_bool32 drwav_init_file_write_sequential_w(drwav* pWav, const wchar_t* filename, const drwav_data_format* pFormat, drwav_uint64 totalSampleCount, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    return drwav_init_file_write_w__internal(pWav, filename, pFormat, totalSampleCount, DRWAV_TRUE, pAllocationCallbacks);
+}
+
+DRWAV_API drwav_bool32 drwav_init_file_write_sequential_pcm_frames_w(drwav* pWav, const wchar_t* filename, const drwav_data_format* pFormat, drwav_uint64 totalPCMFrameCount, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    if (pFormat == NULL) {
+        return DRWAV_FALSE;
+    }
+
+    return drwav_init_file_write_sequential_w(pWav, filename, pFormat, totalPCMFrameCount*pFormat->channels, pAllocationCallbacks);
+}
+#endif  /* DR_WAV_NO_STDIO */
+
+
+static size_t drwav__on_read_memory(void* pUserData, void* pBufferOut, size_t bytesToRead)
+{
+    drwav* pWav = (drwav*)pUserData;
+    size_t bytesRemaining;
+
+    DRWAV_ASSERT(pWav != NULL);
+    DRWAV_ASSERT(pWav->memoryStream.dataSize >= pWav->memoryStream.currentReadPos);
+
+    bytesRemaining = pWav->memoryStream.dataSize - pWav->memoryStream.currentReadPos;
+    if (bytesToRead > bytesRemaining) {
+        bytesToRead = bytesRemaining;
+    }
+
+    if (bytesToRead > 0) {
+        DRWAV_COPY_MEMORY(pBufferOut, pWav->memoryStream.data + pWav->memoryStream.currentReadPos, bytesToRead);
+        pWav->memoryStream.currentReadPos += bytesToRead;
+    }
+
+    return bytesToRead;
+}
+
+static drwav_bool32 drwav__on_seek_memory(void* pUserData, int offset, drwav_seek_origin origin)
+{
+    drwav* pWav = (drwav*)pUserData;
+    DRWAV_ASSERT(pWav != NULL);
+
+    if (origin == drwav_seek_origin_current) {
+        if (offset > 0) {
+            if (pWav->memoryStream.currentReadPos + offset > pWav->memoryStream.dataSize) {
+                return DRWAV_FALSE; /* Trying to seek too far forward. */
+            }
+        } else {
+            if (pWav->memoryStream.currentReadPos < (size_t)-offset) {
+                return DRWAV_FALSE; /* Trying to seek too far backwards. */
+            }
+        }
+
+        /* This will never underflow thanks to the clamps above. */
+        pWav->memoryStream.currentReadPos += offset;
+    } else {
+        if ((drwav_uint32)offset <= pWav->memoryStream.dataSize) {
+            pWav->memoryStream.currentReadPos = offset;
+        } else {
+            return DRWAV_FALSE; /* Trying to seek too far forward. */
+        }
+    }
+    
+    return DRWAV_TRUE;
+}
+
+static size_t drwav__on_write_memory(void* pUserData, const void* pDataIn, size_t bytesToWrite)
+{
+    drwav* pWav = (drwav*)pUserData;
+    size_t bytesRemaining;
+
+    DRWAV_ASSERT(pWav != NULL);
+    DRWAV_ASSERT(pWav->memoryStreamWrite.dataCapacity >= pWav->memoryStreamWrite.currentWritePos);
+
+    bytesRemaining = pWav->memoryStreamWrite.dataCapacity - pWav->memoryStreamWrite.currentWritePos;
+    if (bytesRemaining < bytesToWrite) {
+        /* Need to reallocate. */
+        void* pNewData;
+        size_t newDataCapacity = (pWav->memoryStreamWrite.dataCapacity == 0) ? 256 : pWav->memoryStreamWrite.dataCapacity * 2;
+
+        /* If doubling wasn't enough, just make it the minimum required size to write the data. */
+        if ((newDataCapacity - pWav->memoryStreamWrite.currentWritePos) < bytesToWrite) {
+            newDataCapacity = pWav->memoryStreamWrite.currentWritePos + bytesToWrite;
+        }
+
+        pNewData = drwav__realloc_from_callbacks(*pWav->memoryStreamWrite.ppData, newDataCapacity, pWav->memoryStreamWrite.dataCapacity, &pWav->allocationCallbacks);
+        if (pNewData == NULL) {
+            return 0;
+        }
+
+        *pWav->memoryStreamWrite.ppData = pNewData;
+        pWav->memoryStreamWrite.dataCapacity = newDataCapacity;
+    }
+
+    DRWAV_COPY_MEMORY(((drwav_uint8*)(*pWav->memoryStreamWrite.ppData)) + pWav->memoryStreamWrite.currentWritePos, pDataIn, bytesToWrite);
+
+    pWav->memoryStreamWrite.currentWritePos += bytesToWrite;
+    if (pWav->memoryStreamWrite.dataSize < pWav->memoryStreamWrite.currentWritePos) {
+        pWav->memoryStreamWrite.dataSize = pWav->memoryStreamWrite.currentWritePos;
+    }
+
+    *pWav->memoryStreamWrite.pDataSize = pWav->memoryStreamWrite.dataSize;
+
+    return bytesToWrite;
+}
+
+static drwav_bool32 drwav__on_seek_memory_write(void* pUserData, int offset, drwav_seek_origin origin)
+{
+    drwav* pWav = (drwav*)pUserData;
+    DRWAV_ASSERT(pWav != NULL);
+
+    if (origin == drwav_seek_origin_current) {
+        if (offset > 0) {
+            if (pWav->memoryStreamWrite.currentWritePos + offset > pWav->memoryStreamWrite.dataSize) {
+                offset = (int)(pWav->memoryStreamWrite.dataSize - pWav->memoryStreamWrite.currentWritePos);  /* Trying to seek too far forward. */
+            }
+        } else {
+            if (pWav->memoryStreamWrite.currentWritePos < (size_t)-offset) {
+                offset = -(int)pWav->memoryStreamWrite.currentWritePos;  /* Trying to seek too far backwards. */
+            }
+        }
+
+        /* This will never underflow thanks to the clamps above. */
+        pWav->memoryStreamWrite.currentWritePos += offset;
+    } else {
+        if ((drwav_uint32)offset <= pWav->memoryStreamWrite.dataSize) {
+            pWav->memoryStreamWrite.currentWritePos = offset;
+        } else {
+            pWav->memoryStreamWrite.currentWritePos = pWav->memoryStreamWrite.dataSize;  /* Trying to seek too far forward. */
+        }
+    }
+    
+    return DRWAV_TRUE;
+}
+
+DRWAV_API drwav_bool32 drwav_init_memory(drwav* pWav, const void* data, size_t dataSize, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    return drwav_init_memory_ex(pWav, data, dataSize, NULL, NULL, 0, pAllocationCallbacks);
+}
+
+DRWAV_API drwav_bool32 drwav_init_memory_ex(drwav* pWav, const void* data, size_t dataSize, drwav_chunk_proc onChunk, void* pChunkUserData, drwav_uint32 flags, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    if (data == NULL || dataSize == 0) {
+        return DRWAV_FALSE;
+    }
+
+    if (!drwav_preinit(pWav, drwav__on_read_memory, drwav__on_seek_memory, pWav, pAllocationCallbacks)) {
+        return DRWAV_FALSE;
+    }
+
+    pWav->memoryStream.data = (const drwav_uint8*)data;
+    pWav->memoryStream.dataSize = dataSize;
+    pWav->memoryStream.currentReadPos = 0;
+
+    return drwav_init__internal(pWav, onChunk, pChunkUserData, flags);
+}
+
+
+static drwav_bool32 drwav_init_memory_write__internal(drwav* pWav, void** ppData, size_t* pDataSize, const drwav_data_format* pFormat, drwav_uint64 totalSampleCount, drwav_bool32 isSequential, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    if (ppData == NULL || pDataSize == NULL) {
+        return DRWAV_FALSE;
+    }
+
+    *ppData = NULL; /* Important because we're using realloc()! */
+    *pDataSize = 0;
+
+    if (!drwav_preinit_write(pWav, pFormat, isSequential, drwav__on_write_memory, drwav__on_seek_memory_write, pWav, pAllocationCallbacks)) {
+        return DRWAV_FALSE;
+    }
+
+    pWav->memoryStreamWrite.ppData = ppData;
+    pWav->memoryStreamWrite.pDataSize = pDataSize;
+    pWav->memoryStreamWrite.dataSize = 0;
+    pWav->memoryStreamWrite.dataCapacity = 0;
+    pWav->memoryStreamWrite.currentWritePos = 0;
+
+    return drwav_init_write__internal(pWav, pFormat, totalSampleCount);
+}
+
+DRWAV_API drwav_bool32 drwav_init_memory_write(drwav* pWav, void** ppData, size_t* pDataSize, const drwav_data_format* pFormat, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    return drwav_init_memory_write__internal(pWav, ppData, pDataSize, pFormat, 0, DRWAV_FALSE, pAllocationCallbacks);
+}
+
+DRWAV_API drwav_bool32 drwav_init_memory_write_sequential(drwav* pWav, void** ppData, size_t* pDataSize, const drwav_data_format* pFormat, drwav_uint64 totalSampleCount, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    return drwav_init_memory_write__internal(pWav, ppData, pDataSize, pFormat, totalSampleCount, DRWAV_TRUE, pAllocationCallbacks);
+}
+
+DRWAV_API drwav_bool32 drwav_init_memory_write_sequential_pcm_frames(drwav* pWav, void** ppData, size_t* pDataSize, const drwav_data_format* pFormat, drwav_uint64 totalPCMFrameCount, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    if (pFormat == NULL) {
+        return DRWAV_FALSE;
+    }
+
+    return drwav_init_memory_write_sequential(pWav, ppData, pDataSize, pFormat, totalPCMFrameCount*pFormat->channels, pAllocationCallbacks);
+}
+
+
+
+DRWAV_API drwav_result drwav_uninit(drwav* pWav)
+{
+    drwav_result result = DRWAV_SUCCESS;
+
+    if (pWav == NULL) {
+        return DRWAV_INVALID_ARGS;
+    }
+
+    /*
+    If the drwav object was opened in write mode we'll need to finalize a few things:
+      - Make sure the "data" chunk is aligned to 16-bits for RIFF containers, or 64 bits for W64 containers.
+      - Set the size of the "data" chunk.
+    */
+    if (pWav->onWrite != NULL) {
+        drwav_uint32 paddingSize = 0;
+
+        /* Padding. Do not adjust pWav->dataChunkDataSize - this should not include the padding. */
+        if (pWav->container == drwav_container_riff || pWav->container == drwav_container_rf64) {
+            paddingSize = drwav__chunk_padding_size_riff(pWav->dataChunkDataSize);
+        } else {
+            paddingSize = drwav__chunk_padding_size_w64(pWav->dataChunkDataSize);
+        }
+        
+        if (paddingSize > 0) {
+            drwav_uint64 paddingData = 0;
+            drwav__write(pWav, &paddingData, paddingSize);  /* Byte order does not matter for this. */
+        }
+
+        /*
+        Chunk sizes. When using sequential mode, these will have been filled in at initialization time. We only need
+        to do this when using non-sequential mode.
+        */
+        if (pWav->onSeek && !pWav->isSequentialWrite) {
+            if (pWav->container == drwav_container_riff) {
+                /* The "RIFF" chunk size. */
+                if (pWav->onSeek(pWav->pUserData, 4, drwav_seek_origin_start)) {
+                    drwav_uint32 riffChunkSize = drwav__riff_chunk_size_riff(pWav->dataChunkDataSize);
+                    drwav__write_u32ne_to_le(pWav, riffChunkSize);
+                }
+
+                /* the "data" chunk size. */
+                if (pWav->onSeek(pWav->pUserData, (int)pWav->dataChunkDataPos + 4, drwav_seek_origin_start)) {
+                    drwav_uint32 dataChunkSize = drwav__data_chunk_size_riff(pWav->dataChunkDataSize);
+                    drwav__write_u32ne_to_le(pWav, dataChunkSize);
+                }
+            } else if (pWav->container == drwav_container_w64) {
+                /* The "RIFF" chunk size. */
+                if (pWav->onSeek(pWav->pUserData, 16, drwav_seek_origin_start)) {
+                    drwav_uint64 riffChunkSize = drwav__riff_chunk_size_w64(pWav->dataChunkDataSize);
+                    drwav__write_u64ne_to_le(pWav, riffChunkSize);
+                }
+
+                /* The "data" chunk size. */
+                if (pWav->onSeek(pWav->pUserData, (int)pWav->dataChunkDataPos + 16, drwav_seek_origin_start)) {
+                    drwav_uint64 dataChunkSize = drwav__data_chunk_size_w64(pWav->dataChunkDataSize);
+                    drwav__write_u64ne_to_le(pWav, dataChunkSize);
+                }
+            } else if (pWav->container == drwav_container_rf64) {
+                /* We only need to update the ds64 chunk. The "RIFF" and "data" chunks always have their sizes set to 0xFFFFFFFF for RF64. */
+                int ds64BodyPos = 12 + 8;
+
+                /* The "RIFF" chunk size. */
+                if (pWav->onSeek(pWav->pUserData, ds64BodyPos + 0, drwav_seek_origin_start)) {
+                    drwav_uint64 riffChunkSize = drwav__riff_chunk_size_rf64(pWav->dataChunkDataSize);
+                    drwav__write_u64ne_to_le(pWav, riffChunkSize);
+                }
+
+                /* The "data" chunk size. */
+                if (pWav->onSeek(pWav->pUserData, ds64BodyPos + 8, drwav_seek_origin_start)) {
+                    drwav_uint64 dataChunkSize = drwav__data_chunk_size_rf64(pWav->dataChunkDataSize);
+                    drwav__write_u64ne_to_le(pWav, dataChunkSize);
+                }
+            }
+        }
+
+        /* Validation for sequential mode. */
+        if (pWav->isSequentialWrite) {
+            if (pWav->dataChunkDataSize != pWav->dataChunkDataSizeTargetWrite) {
+                result = DRWAV_INVALID_FILE;
+            }
+        }
+    }
+
+#ifndef DR_WAV_NO_STDIO
+    /*
+    If we opened the file with drwav_open_file() we will want to close the file handle. We can know whether or not drwav_open_file()
+    was used by looking at the onRead and onSeek callbacks.
+    */
+    if (pWav->onRead == drwav__on_read_stdio || pWav->onWrite == drwav__on_write_stdio) {
+        fclose((FILE*)pWav->pUserData);
+    }
+#endif
+
+    return result;
+}
+
+
+
+DRWAV_API size_t drwav_read_raw(drwav* pWav, size_t bytesToRead, void* pBufferOut)
+{
+    size_t bytesRead;
+
+    if (pWav == NULL || bytesToRead == 0) {
+        return 0;
+    }
+
+    if (bytesToRead > pWav->bytesRemaining) {
+        bytesToRead = (size_t)pWav->bytesRemaining;
+    }
+
+    if (pBufferOut != NULL) {
+        bytesRead = pWav->onRead(pWav->pUserData, pBufferOut, bytesToRead);
+    } else {
+        /* We need to seek. If we fail, we need to read-and-discard to make sure we get a good byte count. */
+        bytesRead = 0;
+        while (bytesRead < bytesToRead) {
+            size_t bytesToSeek = (bytesToRead - bytesRead);
+            if (bytesToSeek > 0x7FFFFFFF) {
+                bytesToSeek = 0x7FFFFFFF;
+            }
+
+            if (pWav->onSeek(pWav->pUserData, (int)bytesToSeek, drwav_seek_origin_current) == DRWAV_FALSE) {
+                break;
+            }
+
+            bytesRead += bytesToSeek;
+        }
+
+        /* When we get here we may need to read-and-discard some data. */
+        while (bytesRead < bytesToRead) {
+            drwav_uint8 buffer[4096];
+            size_t bytesSeeked;
+            size_t bytesToSeek = (bytesToRead - bytesRead);
+            if (bytesToSeek > sizeof(buffer)) {
+                bytesToSeek = sizeof(buffer);
+            }
+
+            bytesSeeked = pWav->onRead(pWav->pUserData, buffer, bytesToSeek);
+            bytesRead += bytesSeeked;
+
+            if (bytesSeeked < bytesToSeek) {
+                break;  /* Reached the end. */
+            }
+        }
+    }
+
+    pWav->bytesRemaining -= bytesRead;
+    return bytesRead;
+}
+
+
+
+DRWAV_API drwav_uint64 drwav_read_pcm_frames_le(drwav* pWav, drwav_uint64 framesToRead, void* pBufferOut)
+{
+    drwav_uint32 bytesPerFrame;
+    drwav_uint64 bytesToRead;   /* Intentionally uint64 instead of size_t so we can do a check that we're not reading too much on 32-bit builds. */
+
+    if (pWav == NULL || framesToRead == 0) {
+        return 0;
+    }
+
+    /* Cannot use this function for compressed formats. */
+    if (drwav__is_compressed_format_tag(pWav->translatedFormatTag)) {
+        return 0;
+    }
+
+    bytesPerFrame = drwav_get_bytes_per_pcm_frame(pWav);
+    if (bytesPerFrame == 0) {
+        return 0;
+    }
+
+    /* Don't try to read more samples than can potentially fit in the output buffer. */
+    bytesToRead = framesToRead * bytesPerFrame;
+    if (bytesToRead > DRWAV_SIZE_MAX) {
+        bytesToRead = (DRWAV_SIZE_MAX / bytesPerFrame) * bytesPerFrame; /* Round the number of bytes to read to a clean frame boundary. */
+    }
+
+    /*
+    Doing an explicit check here just to make it clear that we don't want to be attempt to read anything if there's no bytes to read. There
+    *could* be a time where it evaluates to 0 due to overflowing.
+    */
+    if (bytesToRead == 0) {
+        return 0;
+    }
+
+    return drwav_read_raw(pWav, (size_t)bytesToRead, pBufferOut) / bytesPerFrame;
+}
+
+DRWAV_API drwav_uint64 drwav_read_pcm_frames_be(drwav* pWav, drwav_uint64 framesToRead, void* pBufferOut)
+{
+    drwav_uint64 framesRead = drwav_read_pcm_frames_le(pWav, framesToRead, pBufferOut);
+
+    if (pBufferOut != NULL) {
+        drwav__bswap_samples(pBufferOut, framesRead*pWav->channels, drwav_get_bytes_per_pcm_frame(pWav)/pWav->channels, pWav->translatedFormatTag);
+    }
+
+    return framesRead;
+}
+
+DRWAV_API drwav_uint64 drwav_read_pcm_frames(drwav* pWav, drwav_uint64 framesToRead, void* pBufferOut)
+{
+    if (drwav__is_little_endian()) {
+        return drwav_read_pcm_frames_le(pWav, framesToRead, pBufferOut);
+    } else {
+        return drwav_read_pcm_frames_be(pWav, framesToRead, pBufferOut);
+    }
+}
+
+
+
+DRWAV_API drwav_bool32 drwav_seek_to_first_pcm_frame(drwav* pWav)
+{
+    if (pWav->onWrite != NULL) {
+        return DRWAV_FALSE; /* No seeking in write mode. */
+    }
+
+    if (!pWav->onSeek(pWav->pUserData, (int)pWav->dataChunkDataPos, drwav_seek_origin_start)) {
+        return DRWAV_FALSE;
+    }
+
+    if (drwav__is_compressed_format_tag(pWav->translatedFormatTag)) {
+        pWav->compressed.iCurrentPCMFrame = 0;
+
+        /* Cached data needs to be cleared for compressed formats. */
+        if (pWav->translatedFormatTag == DR_WAVE_FORMAT_ADPCM) {
+            DRWAV_ZERO_OBJECT(&pWav->msadpcm);
+        } else if (pWav->translatedFormatTag == DR_WAVE_FORMAT_DVI_ADPCM) {
+            DRWAV_ZERO_OBJECT(&pWav->ima);
+        } else {
+            DRWAV_ASSERT(DRWAV_FALSE);  /* If this assertion is triggered it means I've implemented a new compressed format but forgot to add a branch for it here. */
+        }
+    }
+    
+    pWav->bytesRemaining = pWav->dataChunkDataSize;
+    return DRWAV_TRUE;
+}
+
+DRWAV_API drwav_bool32 drwav_seek_to_pcm_frame(drwav* pWav, drwav_uint64 targetFrameIndex)
+{
+    /* Seeking should be compatible with wave files > 2GB. */
+
+    if (pWav == NULL || pWav->onSeek == NULL) {
+        return DRWAV_FALSE;
+    }
+
+    /* No seeking in write mode. */
+    if (pWav->onWrite != NULL) {
+        return DRWAV_FALSE;
+    }
+
+    /* If there are no samples, just return DRWAV_TRUE without doing anything. */
+    if (pWav->totalPCMFrameCount == 0) {
+        return DRWAV_TRUE;
+    }
+
+    /* Make sure the sample is clamped. */
+    if (targetFrameIndex >= pWav->totalPCMFrameCount) {
+        targetFrameIndex  = pWav->totalPCMFrameCount - 1;
+    }
+
+    /*
+    For compressed formats we just use a slow generic seek. If we are seeking forward we just seek forward. If we are going backwards we need
+    to seek back to the start.
+    */
+    if (drwav__is_compressed_format_tag(pWav->translatedFormatTag)) {
+        /* TODO: This can be optimized. */
+        
+        /*
+        If we're seeking forward it's simple - just keep reading samples until we hit the sample we're requesting. If we're seeking backwards,
+        we first need to seek back to the start and then just do the same thing as a forward seek.
+        */
+        if (targetFrameIndex < pWav->compressed.iCurrentPCMFrame) {
+            if (!drwav_seek_to_first_pcm_frame(pWav)) {
+                return DRWAV_FALSE;
+            }
+        }
+
+        if (targetFrameIndex > pWav->compressed.iCurrentPCMFrame) {
+            drwav_uint64 offsetInFrames = targetFrameIndex - pWav->compressed.iCurrentPCMFrame;
+
+            drwav_int16 devnull[2048];
+            while (offsetInFrames > 0) {
+                drwav_uint64 framesRead = 0;
+                drwav_uint64 framesToRead = offsetInFrames;
+                if (framesToRead > drwav_countof(devnull)/pWav->channels) {
+                    framesToRead = drwav_countof(devnull)/pWav->channels;
+                }
+
+                if (pWav->translatedFormatTag == DR_WAVE_FORMAT_ADPCM) {
+                    framesRead = drwav_read_pcm_frames_s16__msadpcm(pWav, framesToRead, devnull);
+                } else if (pWav->translatedFormatTag == DR_WAVE_FORMAT_DVI_ADPCM) {
+                    framesRead = drwav_read_pcm_frames_s16__ima(pWav, framesToRead, devnull);
+                } else {
+                    DRWAV_ASSERT(DRWAV_FALSE);  /* If this assertion is triggered it means I've implemented a new compressed format but forgot to add a branch for it here. */
+                }
+
+                if (framesRead != framesToRead) {
+                    return DRWAV_FALSE;
+                }
+
+                offsetInFrames -= framesRead;
+            }
+        }
+    } else {
+        drwav_uint64 totalSizeInBytes;
+        drwav_uint64 currentBytePos;
+        drwav_uint64 targetBytePos;
+        drwav_uint64 offset;
+
+        totalSizeInBytes = pWav->totalPCMFrameCount * drwav_get_bytes_per_pcm_frame(pWav);
+        DRWAV_ASSERT(totalSizeInBytes >= pWav->bytesRemaining);
+
+        currentBytePos = totalSizeInBytes - pWav->bytesRemaining;
+        targetBytePos  = targetFrameIndex * drwav_get_bytes_per_pcm_frame(pWav);
+
+        if (currentBytePos < targetBytePos) {
+            /* Offset forwards. */
+            offset = (targetBytePos - currentBytePos);
+        } else {
+            /* Offset backwards. */
+            if (!drwav_seek_to_first_pcm_frame(pWav)) {
+                return DRWAV_FALSE;
+            }
+            offset = targetBytePos;
+        }
+
+        while (offset > 0) {
+            int offset32 = ((offset > INT_MAX) ? INT_MAX : (int)offset);
+            if (!pWav->onSeek(pWav->pUserData, offset32, drwav_seek_origin_current)) {
+                return DRWAV_FALSE;
+            }
+
+            pWav->bytesRemaining -= offset32;
+            offset -= offset32;
+        }
+    }
+
+    return DRWAV_TRUE;
+}
+
+
+DRWAV_API size_t drwav_write_raw(drwav* pWav, size_t bytesToWrite, const void* pData)
+{
+    size_t bytesWritten;
+
+    if (pWav == NULL || bytesToWrite == 0 || pData == NULL) {
+        return 0;
+    }
+
+    bytesWritten = pWav->onWrite(pWav->pUserData, pData, bytesToWrite);
+    pWav->dataChunkDataSize += bytesWritten;
+
+    return bytesWritten;
+}
+
+
+DRWAV_API drwav_uint64 drwav_write_pcm_frames_le(drwav* pWav, drwav_uint64 framesToWrite, const void* pData)
+{
+    drwav_uint64 bytesToWrite;
+    drwav_uint64 bytesWritten;
+    const drwav_uint8* pRunningData;
+
+    if (pWav == NULL || framesToWrite == 0 || pData == NULL) {
+        return 0;
+    }
+
+    bytesToWrite = ((framesToWrite * pWav->channels * pWav->bitsPerSample) / 8);
+    if (bytesToWrite > DRWAV_SIZE_MAX) {
+        return 0;
+    }
+
+    bytesWritten = 0;
+    pRunningData = (const drwav_uint8*)pData;
+
+    while (bytesToWrite > 0) {
+        size_t bytesJustWritten;
+        drwav_uint64 bytesToWriteThisIteration;
+
+        bytesToWriteThisIteration = bytesToWrite;
+        DRWAV_ASSERT(bytesToWriteThisIteration <= DRWAV_SIZE_MAX);  /* <-- This is checked above. */
+
+        bytesJustWritten = drwav_write_raw(pWav, (size_t)bytesToWriteThisIteration, pRunningData);
+        if (bytesJustWritten == 0) {
+            break;
+        }
+
+        bytesToWrite -= bytesJustWritten;
+        bytesWritten += bytesJustWritten;
+        pRunningData += bytesJustWritten;
+    }
+
+    return (bytesWritten * 8) / pWav->bitsPerSample / pWav->channels;
+}
+
+DRWAV_API drwav_uint64 drwav_write_pcm_frames_be(drwav* pWav, drwav_uint64 framesToWrite, const void* pData)
+{
+    drwav_uint64 bytesToWrite;
+    drwav_uint64 bytesWritten;
+    drwav_uint32 bytesPerSample;
+    const drwav_uint8* pRunningData;
+
+    if (pWav == NULL || framesToWrite == 0 || pData == NULL) {
+        return 0;
+    }
+
+    bytesToWrite = ((framesToWrite * pWav->channels * pWav->bitsPerSample) / 8);
+    if (bytesToWrite > DRWAV_SIZE_MAX) {
+        return 0;
+    }
+
+    bytesWritten = 0;
+    pRunningData = (const drwav_uint8*)pData;
+
+    bytesPerSample = drwav_get_bytes_per_pcm_frame(pWav) / pWav->channels;
+    
+    while (bytesToWrite > 0) {
+        drwav_uint8 temp[4096];
+        drwav_uint32 sampleCount;
+        size_t bytesJustWritten;
+        drwav_uint64 bytesToWriteThisIteration;
+
+        bytesToWriteThisIteration = bytesToWrite;
+        DRWAV_ASSERT(bytesToWriteThisIteration <= DRWAV_SIZE_MAX);  /* <-- This is checked above. */
+
+        /*
+        WAV files are always little-endian. We need to byte swap on big-endian architectures. Since our input buffer is read-only we need
+        to use an intermediary buffer for the conversion.
+        */
+        sampleCount = sizeof(temp)/bytesPerSample;
+
+        if (bytesToWriteThisIteration > ((drwav_uint64)sampleCount)*bytesPerSample) {
+            bytesToWriteThisIteration = ((drwav_uint64)sampleCount)*bytesPerSample;
+        }
+
+        DRWAV_COPY_MEMORY(temp, pRunningData, (size_t)bytesToWriteThisIteration);
+        drwav__bswap_samples(temp, sampleCount, bytesPerSample, pWav->translatedFormatTag);
+
+        bytesJustWritten = drwav_write_raw(pWav, (size_t)bytesToWriteThisIteration, temp);
+        if (bytesJustWritten == 0) {
+            break;
+        }
+
+        bytesToWrite -= bytesJustWritten;
+        bytesWritten += bytesJustWritten;
+        pRunningData += bytesJustWritten;
+    }
+
+    return (bytesWritten * 8) / pWav->bitsPerSample / pWav->channels;
+}
+
+DRWAV_API drwav_uint64 drwav_write_pcm_frames(drwav* pWav, drwav_uint64 framesToWrite, const void* pData)
+{
+    if (drwav__is_little_endian()) {
+        return drwav_write_pcm_frames_le(pWav, framesToWrite, pData);
+    } else {
+        return drwav_write_pcm_frames_be(pWav, framesToWrite, pData);
+    }
+}
+
+
+static drwav_uint64 drwav_read_pcm_frames_s16__msadpcm(drwav* pWav, drwav_uint64 framesToRead, drwav_int16* pBufferOut)
+{
+    drwav_uint64 totalFramesRead = 0;
+
+    DRWAV_ASSERT(pWav != NULL);
+    DRWAV_ASSERT(framesToRead > 0);
+
+    /* TODO: Lots of room for optimization here. */
+
+    while (framesToRead > 0 && pWav->compressed.iCurrentPCMFrame < pWav->totalPCMFrameCount) {
+        /* If there are no cached frames we need to load a new block. */
+        if (pWav->msadpcm.cachedFrameCount == 0 && pWav->msadpcm.bytesRemainingInBlock == 0) {
+            if (pWav->channels == 1) {
+                /* Mono. */
+                drwav_uint8 header[7];
+                if (pWav->onRead(pWav->pUserData, header, sizeof(header)) != sizeof(header)) {
+                    return totalFramesRead;
+                }
+                pWav->msadpcm.bytesRemainingInBlock = pWav->fmt.blockAlign - sizeof(header);
+
+                pWav->msadpcm.predictor[0]     = header[0];
+                pWav->msadpcm.delta[0]         = drwav__bytes_to_s16(header + 1);
+                pWav->msadpcm.prevFrames[0][1] = (drwav_int32)drwav__bytes_to_s16(header + 3);
+                pWav->msadpcm.prevFrames[0][0] = (drwav_int32)drwav__bytes_to_s16(header + 5);
+                pWav->msadpcm.cachedFrames[2]  = pWav->msadpcm.prevFrames[0][0];
+                pWav->msadpcm.cachedFrames[3]  = pWav->msadpcm.prevFrames[0][1];
+                pWav->msadpcm.cachedFrameCount = 2;
+            } else {
+                /* Stereo. */
+                drwav_uint8 header[14];
+                if (pWav->onRead(pWav->pUserData, header, sizeof(header)) != sizeof(header)) {
+                    return totalFramesRead;
+                }
+                pWav->msadpcm.bytesRemainingInBlock = pWav->fmt.blockAlign - sizeof(header);
+
+                pWav->msadpcm.predictor[0] = header[0];
+                pWav->msadpcm.predictor[1] = header[1];
+                pWav->msadpcm.delta[0] = drwav__bytes_to_s16(header + 2);
+                pWav->msadpcm.delta[1] = drwav__bytes_to_s16(header + 4);
+                pWav->msadpcm.prevFrames[0][1] = (drwav_int32)drwav__bytes_to_s16(header + 6);
+                pWav->msadpcm.prevFrames[1][1] = (drwav_int32)drwav__bytes_to_s16(header + 8);
+                pWav->msadpcm.prevFrames[0][0] = (drwav_int32)drwav__bytes_to_s16(header + 10);
+                pWav->msadpcm.prevFrames[1][0] = (drwav_int32)drwav__bytes_to_s16(header + 12);
+
+                pWav->msadpcm.cachedFrames[0] = pWav->msadpcm.prevFrames[0][0];
+                pWav->msadpcm.cachedFrames[1] = pWav->msadpcm.prevFrames[1][0];
+                pWav->msadpcm.cachedFrames[2] = pWav->msadpcm.prevFrames[0][1];
+                pWav->msadpcm.cachedFrames[3] = pWav->msadpcm.prevFrames[1][1];
+                pWav->msadpcm.cachedFrameCount = 2;
+            }
+        }
+
+        /* Output anything that's cached. */
+        while (framesToRead > 0 && pWav->msadpcm.cachedFrameCount > 0 && pWav->compressed.iCurrentPCMFrame < pWav->totalPCMFrameCount) {
+            if (pBufferOut != NULL) {
+                drwav_uint32 iSample = 0;
+                for (iSample = 0; iSample < pWav->channels; iSample += 1) {
+                    pBufferOut[iSample] = (drwav_int16)pWav->msadpcm.cachedFrames[(drwav_countof(pWav->msadpcm.cachedFrames) - (pWav->msadpcm.cachedFrameCount*pWav->channels)) + iSample];
+                }
+
+                pBufferOut += pWav->channels;
+            }
+
+            framesToRead    -= 1;
+            totalFramesRead += 1;
+            pWav->compressed.iCurrentPCMFrame += 1;
+            pWav->msadpcm.cachedFrameCount -= 1;
+        }
+
+        if (framesToRead == 0) {
+            return totalFramesRead;
+        }
+
+
+        /*
+        If there's nothing left in the cache, just go ahead and load more. If there's nothing left to load in the current block we just continue to the next
+        loop iteration which will trigger the loading of a new block.
+        */
+        if (pWav->msadpcm.cachedFrameCount == 0) {
+            if (pWav->msadpcm.bytesRemainingInBlock == 0) {
+                continue;
+            } else {
+                static drwav_int32 adaptationTable[] = { 
+                    230, 230, 230, 230, 307, 409, 512, 614, 
+                    768, 614, 512, 409, 307, 230, 230, 230 
+                };
+                static drwav_int32 coeff1Table[] = { 256, 512, 0, 192, 240, 460,  392 };
+                static drwav_int32 coeff2Table[] = { 0,  -256, 0, 64,  0,  -208, -232 };
+
+                drwav_uint8 nibbles;
+                drwav_int32 nibble0;
+                drwav_int32 nibble1;
+
+                if (pWav->onRead(pWav->pUserData, &nibbles, 1) != 1) {
+                    return totalFramesRead;
+                }
+                pWav->msadpcm.bytesRemainingInBlock -= 1;
+
+                /* TODO: Optimize away these if statements. */
+                nibble0 = ((nibbles & 0xF0) >> 4); if ((nibbles & 0x80)) { nibble0 |= 0xFFFFFFF0UL; }
+                nibble1 = ((nibbles & 0x0F) >> 0); if ((nibbles & 0x08)) { nibble1 |= 0xFFFFFFF0UL; }
+
+                if (pWav->channels == 1) {
+                    /* Mono. */
+                    drwav_int32 newSample0;
+                    drwav_int32 newSample1;
+
+                    newSample0  = ((pWav->msadpcm.prevFrames[0][1] * coeff1Table[pWav->msadpcm.predictor[0]]) + (pWav->msadpcm.prevFrames[0][0] * coeff2Table[pWav->msadpcm.predictor[0]])) >> 8;
+                    newSample0 += nibble0 * pWav->msadpcm.delta[0];
+                    newSample0  = drwav_clamp(newSample0, -32768, 32767);
+
+                    pWav->msadpcm.delta[0] = (adaptationTable[((nibbles & 0xF0) >> 4)] * pWav->msadpcm.delta[0]) >> 8;
+                    if (pWav->msadpcm.delta[0] < 16) {
+                        pWav->msadpcm.delta[0] = 16;
+                    }
+
+                    pWav->msadpcm.prevFrames[0][0] = pWav->msadpcm.prevFrames[0][1];
+                    pWav->msadpcm.prevFrames[0][1] = newSample0;
+
+
+                    newSample1  = ((pWav->msadpcm.prevFrames[0][1] * coeff1Table[pWav->msadpcm.predictor[0]]) + (pWav->msadpcm.prevFrames[0][0] * coeff2Table[pWav->msadpcm.predictor[0]])) >> 8;
+                    newSample1 += nibble1 * pWav->msadpcm.delta[0];
+                    newSample1  = drwav_clamp(newSample1, -32768, 32767);
+
+                    pWav->msadpcm.delta[0] = (adaptationTable[((nibbles & 0x0F) >> 0)] * pWav->msadpcm.delta[0]) >> 8;
+                    if (pWav->msadpcm.delta[0] < 16) {
+                        pWav->msadpcm.delta[0] = 16;
+                    }
+
+                    pWav->msadpcm.prevFrames[0][0] = pWav->msadpcm.prevFrames[0][1];
+                    pWav->msadpcm.prevFrames[0][1] = newSample1;
+
+
+                    pWav->msadpcm.cachedFrames[2] = newSample0;
+                    pWav->msadpcm.cachedFrames[3] = newSample1;
+                    pWav->msadpcm.cachedFrameCount = 2;
+                } else {
+                    /* Stereo. */
+                    drwav_int32 newSample0;
+                    drwav_int32 newSample1;
+
+                    /* Left. */
+                    newSample0  = ((pWav->msadpcm.prevFrames[0][1] * coeff1Table[pWav->msadpcm.predictor[0]]) + (pWav->msadpcm.prevFrames[0][0] * coeff2Table[pWav->msadpcm.predictor[0]])) >> 8;
+                    newSample0 += nibble0 * pWav->msadpcm.delta[0];
+                    newSample0  = drwav_clamp(newSample0, -32768, 32767);
+
+                    pWav->msadpcm.delta[0] = (adaptationTable[((nibbles & 0xF0) >> 4)] * pWav->msadpcm.delta[0]) >> 8;
+                    if (pWav->msadpcm.delta[0] < 16) {
+                        pWav->msadpcm.delta[0] = 16;
+                    }
+
+                    pWav->msadpcm.prevFrames[0][0] = pWav->msadpcm.prevFrames[0][1];
+                    pWav->msadpcm.prevFrames[0][1] = newSample0;
+
+
+                    /* Right. */
+                    newSample1  = ((pWav->msadpcm.prevFrames[1][1] * coeff1Table[pWav->msadpcm.predictor[1]]) + (pWav->msadpcm.prevFrames[1][0] * coeff2Table[pWav->msadpcm.predictor[1]])) >> 8;
+                    newSample1 += nibble1 * pWav->msadpcm.delta[1];
+                    newSample1  = drwav_clamp(newSample1, -32768, 32767);
+
+                    pWav->msadpcm.delta[1] = (adaptationTable[((nibbles & 0x0F) >> 0)] * pWav->msadpcm.delta[1]) >> 8;
+                    if (pWav->msadpcm.delta[1] < 16) {
+                        pWav->msadpcm.delta[1] = 16;
+                    }
+
+                    pWav->msadpcm.prevFrames[1][0] = pWav->msadpcm.prevFrames[1][1];
+                    pWav->msadpcm.prevFrames[1][1] = newSample1;
+
+                    pWav->msadpcm.cachedFrames[2] = newSample0;
+                    pWav->msadpcm.cachedFrames[3] = newSample1;
+                    pWav->msadpcm.cachedFrameCount = 1;
+                }
+            }
+        }
+    }
+
+    return totalFramesRead;
+}
+
+
+static drwav_uint64 drwav_read_pcm_frames_s16__ima(drwav* pWav, drwav_uint64 framesToRead, drwav_int16* pBufferOut)
+{
+    drwav_uint64 totalFramesRead = 0;
+    drwav_uint32 iChannel;
+
+    static drwav_int32 indexTable[16] = {
+        -1, -1, -1, -1, 2, 4, 6, 8,
+        -1, -1, -1, -1, 2, 4, 6, 8
+    };
+
+    static drwav_int32 stepTable[89] = {
+        7,     8,     9,     10,    11,    12,    13,    14,    16,    17, 
+        19,    21,    23,    25,    28,    31,    34,    37,    41,    45, 
+        50,    55,    60,    66,    73,    80,    88,    97,    107,   118, 
+        130,   143,   157,   173,   190,   209,   230,   253,   279,   307,
+        337,   371,   408,   449,   494,   544,   598,   658,   724,   796,
+        876,   963,   1060,  1166,  1282,  1411,  1552,  1707,  1878,  2066, 
+        2272,  2499,  2749,  3024,  3327,  3660,  4026,  4428,  4871,  5358,
+        5894,  6484,  7132,  7845,  8630,  9493,  10442, 11487, 12635, 13899, 
+        15289, 16818, 18500, 20350, 22385, 24623, 27086, 29794, 32767 
+    };
+
+    DRWAV_ASSERT(pWav != NULL);
+    DRWAV_ASSERT(framesToRead > 0);
+
+    /* TODO: Lots of room for optimization here. */
+
+    while (framesToRead > 0 && pWav->compressed.iCurrentPCMFrame < pWav->totalPCMFrameCount) {
+        /* If there are no cached samples we need to load a new block. */
+        if (pWav->ima.cachedFrameCount == 0 && pWav->ima.bytesRemainingInBlock == 0) {
+            if (pWav->channels == 1) {
+                /* Mono. */
+                drwav_uint8 header[4];
+                if (pWav->onRead(pWav->pUserData, header, sizeof(header)) != sizeof(header)) {
+                    return totalFramesRead;
+                }
+                pWav->ima.bytesRemainingInBlock = pWav->fmt.blockAlign - sizeof(header);
+
+                if (header[2] >= drwav_countof(stepTable)) {
+                    pWav->onSeek(pWav->pUserData, pWav->ima.bytesRemainingInBlock, drwav_seek_origin_current);
+                    pWav->ima.bytesRemainingInBlock = 0;
+                    return totalFramesRead; /* Invalid data. */
+                }
+
+                pWav->ima.predictor[0] = drwav__bytes_to_s16(header + 0);
+                pWav->ima.stepIndex[0] = header[2];
+                pWav->ima.cachedFrames[drwav_countof(pWav->ima.cachedFrames) - 1] = pWav->ima.predictor[0];
+                pWav->ima.cachedFrameCount = 1;
+            } else {
+                /* Stereo. */
+                drwav_uint8 header[8];
+                if (pWav->onRead(pWav->pUserData, header, sizeof(header)) != sizeof(header)) {
+                    return totalFramesRead;
+                }
+                pWav->ima.bytesRemainingInBlock = pWav->fmt.blockAlign - sizeof(header);
+
+                if (header[2] >= drwav_countof(stepTable) || header[6] >= drwav_countof(stepTable)) {
+                    pWav->onSeek(pWav->pUserData, pWav->ima.bytesRemainingInBlock, drwav_seek_origin_current);
+                    pWav->ima.bytesRemainingInBlock = 0;
+                    return totalFramesRead; /* Invalid data. */
+                }
+
+                pWav->ima.predictor[0] = drwav__bytes_to_s16(header + 0);
+                pWav->ima.stepIndex[0] = header[2];
+                pWav->ima.predictor[1] = drwav__bytes_to_s16(header + 4);
+                pWav->ima.stepIndex[1] = header[6];
+
+                pWav->ima.cachedFrames[drwav_countof(pWav->ima.cachedFrames) - 2] = pWav->ima.predictor[0];
+                pWav->ima.cachedFrames[drwav_countof(pWav->ima.cachedFrames) - 1] = pWav->ima.predictor[1];
+                pWav->ima.cachedFrameCount = 1;
+            }
+        }
+
+        /* Output anything that's cached. */
+        while (framesToRead > 0 && pWav->ima.cachedFrameCount > 0 && pWav->compressed.iCurrentPCMFrame < pWav->totalPCMFrameCount) {
+            if (pBufferOut != NULL) {
+                drwav_uint32 iSample;
+                for (iSample = 0; iSample < pWav->channels; iSample += 1) {
+                    pBufferOut[iSample] = (drwav_int16)pWav->ima.cachedFrames[(drwav_countof(pWav->ima.cachedFrames) - (pWav->ima.cachedFrameCount*pWav->channels)) + iSample];
+                }
+                pBufferOut += pWav->channels;
+            }
+
+            framesToRead    -= 1;
+            totalFramesRead += 1;
+            pWav->compressed.iCurrentPCMFrame += 1;
+            pWav->ima.cachedFrameCount -= 1;
+        }
+
+        if (framesToRead == 0) {
+            return totalFramesRead;
+        }
+
+        /*
+        If there's nothing left in the cache, just go ahead and load more. If there's nothing left to load in the current block we just continue to the next
+        loop iteration which will trigger the loading of a new block.
+        */
+        if (pWav->ima.cachedFrameCount == 0) {
+            if (pWav->ima.bytesRemainingInBlock == 0) {
+                continue;
+            } else {
+                /*
+                From what I can tell with stereo streams, it looks like every 4 bytes (8 samples) is for one channel. So it goes 4 bytes for the
+                left channel, 4 bytes for the right channel.
+                */
+                pWav->ima.cachedFrameCount = 8;
+                for (iChannel = 0; iChannel < pWav->channels; ++iChannel) {
+                    drwav_uint32 iByte;
+                    drwav_uint8 nibbles[4];
+                    if (pWav->onRead(pWav->pUserData, &nibbles, 4) != 4) {
+                        pWav->ima.cachedFrameCount = 0;
+                        return totalFramesRead;
+                    }
+                    pWav->ima.bytesRemainingInBlock -= 4;
+
+                    for (iByte = 0; iByte < 4; ++iByte) {
+                        drwav_uint8 nibble0 = ((nibbles[iByte] & 0x0F) >> 0);
+                        drwav_uint8 nibble1 = ((nibbles[iByte] & 0xF0) >> 4);
+
+                        drwav_int32 step      = stepTable[pWav->ima.stepIndex[iChannel]];
+                        drwav_int32 predictor = pWav->ima.predictor[iChannel];
+
+                        drwav_int32      diff  = step >> 3;
+                        if (nibble0 & 1) diff += step >> 2;
+                        if (nibble0 & 2) diff += step >> 1;
+                        if (nibble0 & 4) diff += step;
+                        if (nibble0 & 8) diff  = -diff;
+
+                        predictor = drwav_clamp(predictor + diff, -32768, 32767);
+                        pWav->ima.predictor[iChannel] = predictor;
+                        pWav->ima.stepIndex[iChannel] = drwav_clamp(pWav->ima.stepIndex[iChannel] + indexTable[nibble0], 0, (drwav_int32)drwav_countof(stepTable)-1);
+                        pWav->ima.cachedFrames[(drwav_countof(pWav->ima.cachedFrames) - (pWav->ima.cachedFrameCount*pWav->channels)) + (iByte*2+0)*pWav->channels + iChannel] = predictor;
+
+
+                        step      = stepTable[pWav->ima.stepIndex[iChannel]];
+                        predictor = pWav->ima.predictor[iChannel];
+
+                                         diff  = step >> 3;
+                        if (nibble1 & 1) diff += step >> 2;
+                        if (nibble1 & 2) diff += step >> 1;
+                        if (nibble1 & 4) diff += step;
+                        if (nibble1 & 8) diff  = -diff;
+
+                        predictor = drwav_clamp(predictor + diff, -32768, 32767);
+                        pWav->ima.predictor[iChannel] = predictor;
+                        pWav->ima.stepIndex[iChannel] = drwav_clamp(pWav->ima.stepIndex[iChannel] + indexTable[nibble1], 0, (drwav_int32)drwav_countof(stepTable)-1);
+                        pWav->ima.cachedFrames[(drwav_countof(pWav->ima.cachedFrames) - (pWav->ima.cachedFrameCount*pWav->channels)) + (iByte*2+1)*pWav->channels + iChannel] = predictor;
+                    }
+                }
+            }
+        }
+    }
+
+    return totalFramesRead;
+}
+
+
+#ifndef DR_WAV_NO_CONVERSION_API
+static unsigned short g_drwavAlawTable[256] = {
+    0xEA80, 0xEB80, 0xE880, 0xE980, 0xEE80, 0xEF80, 0xEC80, 0xED80, 0xE280, 0xE380, 0xE080, 0xE180, 0xE680, 0xE780, 0xE480, 0xE580, 
+    0xF540, 0xF5C0, 0xF440, 0xF4C0, 0xF740, 0xF7C0, 0xF640, 0xF6C0, 0xF140, 0xF1C0, 0xF040, 0xF0C0, 0xF340, 0xF3C0, 0xF240, 0xF2C0, 
+    0xAA00, 0xAE00, 0xA200, 0xA600, 0xBA00, 0xBE00, 0xB200, 0xB600, 0x8A00, 0x8E00, 0x8200, 0x8600, 0x9A00, 0x9E00, 0x9200, 0x9600, 
+    0xD500, 0xD700, 0xD100, 0xD300, 0xDD00, 0xDF00, 0xD900, 0xDB00, 0xC500, 0xC700, 0xC100, 0xC300, 0xCD00, 0xCF00, 0xC900, 0xCB00, 
+    0xFEA8, 0xFEB8, 0xFE88, 0xFE98, 0xFEE8, 0xFEF8, 0xFEC8, 0xFED8, 0xFE28, 0xFE38, 0xFE08, 0xFE18, 0xFE68, 0xFE78, 0xFE48, 0xFE58, 
+    0xFFA8, 0xFFB8, 0xFF88, 0xFF98, 0xFFE8, 0xFFF8, 0xFFC8, 0xFFD8, 0xFF28, 0xFF38, 0xFF08, 0xFF18, 0xFF68, 0xFF78, 0xFF48, 0xFF58, 
+    0xFAA0, 0xFAE0, 0xFA20, 0xFA60, 0xFBA0, 0xFBE0, 0xFB20, 0xFB60, 0xF8A0, 0xF8E0, 0xF820, 0xF860, 0xF9A0, 0xF9E0, 0xF920, 0xF960, 
+    0xFD50, 0xFD70, 0xFD10, 0xFD30, 0xFDD0, 0xFDF0, 0xFD90, 0xFDB0, 0xFC50, 0xFC70, 0xFC10, 0xFC30, 0xFCD0, 0xFCF0, 0xFC90, 0xFCB0, 
+    0x1580, 0x1480, 0x1780, 0x1680, 0x1180, 0x1080, 0x1380, 0x1280, 0x1D80, 0x1C80, 0x1F80, 0x1E80, 0x1980, 0x1880, 0x1B80, 0x1A80, 
+    0x0AC0, 0x0A40, 0x0BC0, 0x0B40, 0x08C0, 0x0840, 0x09C0, 0x0940, 0x0EC0, 0x0E40, 0x0FC0, 0x0F40, 0x0CC0, 0x0C40, 0x0DC0, 0x0D40, 
+    0x5600, 0x5200, 0x5E00, 0x5A00, 0x4600, 0x4200, 0x4E00, 0x4A00, 0x7600, 0x7200, 0x7E00, 0x7A00, 0x6600, 0x6200, 0x6E00, 0x6A00, 
+    0x2B00, 0x2900, 0x2F00, 0x2D00, 0x2300, 0x2100, 0x2700, 0x2500, 0x3B00, 0x3900, 0x3F00, 0x3D00, 0x3300, 0x3100, 0x3700, 0x3500, 
+    0x0158, 0x0148, 0x0178, 0x0168, 0x0118, 0x0108, 0x0138, 0x0128, 0x01D8, 0x01C8, 0x01F8, 0x01E8, 0x0198, 0x0188, 0x01B8, 0x01A8, 
+    0x0058, 0x0048, 0x0078, 0x0068, 0x0018, 0x0008, 0x0038, 0x0028, 0x00D8, 0x00C8, 0x00F8, 0x00E8, 0x0098, 0x0088, 0x00B8, 0x00A8, 
+    0x0560, 0x0520, 0x05E0, 0x05A0, 0x0460, 0x0420, 0x04E0, 0x04A0, 0x0760, 0x0720, 0x07E0, 0x07A0, 0x0660, 0x0620, 0x06E0, 0x06A0, 
+    0x02B0, 0x0290, 0x02F0, 0x02D0, 0x0230, 0x0210, 0x0270, 0x0250, 0x03B0, 0x0390, 0x03F0, 0x03D0, 0x0330, 0x0310, 0x0370, 0x0350
+};
+
+static unsigned short g_drwavMulawTable[256] = {
+    0x8284, 0x8684, 0x8A84, 0x8E84, 0x9284, 0x9684, 0x9A84, 0x9E84, 0xA284, 0xA684, 0xAA84, 0xAE84, 0xB284, 0xB684, 0xBA84, 0xBE84, 
+    0xC184, 0xC384, 0xC584, 0xC784, 0xC984, 0xCB84, 0xCD84, 0xCF84, 0xD184, 0xD384, 0xD584, 0xD784, 0xD984, 0xDB84, 0xDD84, 0xDF84, 
+    0xE104, 0xE204, 0xE304, 0xE404, 0xE504, 0xE604, 0xE704, 0xE804, 0xE904, 0xEA04, 0xEB04, 0xEC04, 0xED04, 0xEE04, 0xEF04, 0xF004, 
+    0xF0C4, 0xF144, 0xF1C4, 0xF244, 0xF2C4, 0xF344, 0xF3C4, 0xF444, 0xF4C4, 0xF544, 0xF5C4, 0xF644, 0xF6C4, 0xF744, 0xF7C4, 0xF844, 
+    0xF8A4, 0xF8E4, 0xF924, 0xF964, 0xF9A4, 0xF9E4, 0xFA24, 0xFA64, 0xFAA4, 0xFAE4, 0xFB24, 0xFB64, 0xFBA4, 0xFBE4, 0xFC24, 0xFC64, 
+    0xFC94, 0xFCB4, 0xFCD4, 0xFCF4, 0xFD14, 0xFD34, 0xFD54, 0xFD74, 0xFD94, 0xFDB4, 0xFDD4, 0xFDF4, 0xFE14, 0xFE34, 0xFE54, 0xFE74, 
+    0xFE8C, 0xFE9C, 0xFEAC, 0xFEBC, 0xFECC, 0xFEDC, 0xFEEC, 0xFEFC, 0xFF0C, 0xFF1C, 0xFF2C, 0xFF3C, 0xFF4C, 0xFF5C, 0xFF6C, 0xFF7C, 
+    0xFF88, 0xFF90, 0xFF98, 0xFFA0, 0xFFA8, 0xFFB0, 0xFFB8, 0xFFC0, 0xFFC8, 0xFFD0, 0xFFD8, 0xFFE0, 0xFFE8, 0xFFF0, 0xFFF8, 0x0000, 
+    0x7D7C, 0x797C, 0x757C, 0x717C, 0x6D7C, 0x697C, 0x657C, 0x617C, 0x5D7C, 0x597C, 0x557C, 0x517C, 0x4D7C, 0x497C, 0x457C, 0x417C, 
+    0x3E7C, 0x3C7C, 0x3A7C, 0x387C, 0x367C, 0x347C, 0x327C, 0x307C, 0x2E7C, 0x2C7C, 0x2A7C, 0x287C, 0x267C, 0x247C, 0x227C, 0x207C, 
+    0x1EFC, 0x1DFC, 0x1CFC, 0x1BFC, 0x1AFC, 0x19FC, 0x18FC, 0x17FC, 0x16FC, 0x15FC, 0x14FC, 0x13FC, 0x12FC, 0x11FC, 0x10FC, 0x0FFC, 
+    0x0F3C, 0x0EBC, 0x0E3C, 0x0DBC, 0x0D3C, 0x0CBC, 0x0C3C, 0x0BBC, 0x0B3C, 0x0ABC, 0x0A3C, 0x09BC, 0x093C, 0x08BC, 0x083C, 0x07BC, 
+    0x075C, 0x071C, 0x06DC, 0x069C, 0x065C, 0x061C, 0x05DC, 0x059C, 0x055C, 0x051C, 0x04DC, 0x049C, 0x045C, 0x041C, 0x03DC, 0x039C, 
+    0x036C, 0x034C, 0x032C, 0x030C, 0x02EC, 0x02CC, 0x02AC, 0x028C, 0x026C, 0x024C, 0x022C, 0x020C, 0x01EC, 0x01CC, 0x01AC, 0x018C, 
+    0x0174, 0x0164, 0x0154, 0x0144, 0x0134, 0x0124, 0x0114, 0x0104, 0x00F4, 0x00E4, 0x00D4, 0x00C4, 0x00B4, 0x00A4, 0x0094, 0x0084, 
+    0x0078, 0x0070, 0x0068, 0x0060, 0x0058, 0x0050, 0x0048, 0x0040, 0x0038, 0x0030, 0x0028, 0x0020, 0x0018, 0x0010, 0x0008, 0x0000
+};
+
+static DRWAV_INLINE drwav_int16 drwav__alaw_to_s16(drwav_uint8 sampleIn)
+{
+    return (short)g_drwavAlawTable[sampleIn];
+}
+
+static DRWAV_INLINE drwav_int16 drwav__mulaw_to_s16(drwav_uint8 sampleIn)
+{
+    return (short)g_drwavMulawTable[sampleIn];
+}
+
+
+
+static void drwav__pcm_to_s16(drwav_int16* pOut, const drwav_uint8* pIn, size_t totalSampleCount, unsigned int bytesPerSample)
+{
+    unsigned int i;
+
+    /* Special case for 8-bit sample data because it's treated as unsigned. */
+    if (bytesPerSample == 1) {
+        drwav_u8_to_s16(pOut, pIn, totalSampleCount);
+        return;
+    }
+
+
+    /* Slightly more optimal implementation for common formats. */
+    if (bytesPerSample == 2) {
+        for (i = 0; i < totalSampleCount; ++i) {
+           *pOut++ = ((const drwav_int16*)pIn)[i];
+        }
+        return;
+    }
+    if (bytesPerSample == 3) {
+        drwav_s24_to_s16(pOut, pIn, totalSampleCount);
+        return;
+    }
+    if (bytesPerSample == 4) {
+        drwav_s32_to_s16(pOut, (const drwav_int32*)pIn, totalSampleCount);
+        return;
+    }
+
+
+    /* Anything more than 64 bits per sample is not supported. */
+    if (bytesPerSample > 8) {
+        DRWAV_ZERO_MEMORY(pOut, totalSampleCount * sizeof(*pOut));
+        return;
+    }
+
+
+    /* Generic, slow converter. */
+    for (i = 0; i < totalSampleCount; ++i) {
+        drwav_uint64 sample = 0;
+        unsigned int shift  = (8 - bytesPerSample) * 8;
+
+        unsigned int j;
+        for (j = 0; j < bytesPerSample; j += 1) {
+            DRWAV_ASSERT(j < 8);
+            sample |= (drwav_uint64)(pIn[j]) << shift;
+            shift  += 8;
+        }
+
+        pIn += j;
+        *pOut++ = (drwav_int16)((drwav_int64)sample >> 48);
+    }
+}
+
+static void drwav__ieee_to_s16(drwav_int16* pOut, const drwav_uint8* pIn, size_t totalSampleCount, unsigned int bytesPerSample)
+{
+    if (bytesPerSample == 4) {
+        drwav_f32_to_s16(pOut, (const float*)pIn, totalSampleCount);
+        return;
+    } else if (bytesPerSample == 8) {
+        drwav_f64_to_s16(pOut, (const double*)pIn, totalSampleCount);
+        return;
+    } else {
+        /* Only supporting 32- and 64-bit float. Output silence in all other cases. Contributions welcome for 16-bit float. */
+        DRWAV_ZERO_MEMORY(pOut, totalSampleCount * sizeof(*pOut));
+        return;
+    }
+}
+
+static drwav_uint64 drwav_read_pcm_frames_s16__pcm(drwav* pWav, drwav_uint64 framesToRead, drwav_int16* pBufferOut)
+{
+    drwav_uint32 bytesPerFrame;
+    drwav_uint64 totalFramesRead;
+    drwav_uint8 sampleData[4096];
+
+    /* Fast path. */
+    if ((pWav->translatedFormatTag == DR_WAVE_FORMAT_PCM && pWav->bitsPerSample == 16) || pBufferOut == NULL) {
+        return drwav_read_pcm_frames(pWav, framesToRead, pBufferOut);
+    }
+    
+    bytesPerFrame = drwav_get_bytes_per_pcm_frame(pWav);
+    if (bytesPerFrame == 0) {
+        return 0;
+    }
+
+    totalFramesRead = 0;
+    
+    while (framesToRead > 0) {
+        drwav_uint64 framesRead = drwav_read_pcm_frames(pWav, drwav_min(framesToRead, sizeof(sampleData)/bytesPerFrame), sampleData);
+        if (framesRead == 0) {
+            break;
+        }
+
+        drwav__pcm_to_s16(pBufferOut, sampleData, (size_t)(framesRead*pWav->channels), bytesPerFrame/pWav->channels);
+
+        pBufferOut      += framesRead*pWav->channels;
+        framesToRead    -= framesRead;
+        totalFramesRead += framesRead;
+    }
+
+    return totalFramesRead;
+}
+
+static drwav_uint64 drwav_read_pcm_frames_s16__ieee(drwav* pWav, drwav_uint64 framesToRead, drwav_int16* pBufferOut)
+{
+    drwav_uint64 totalFramesRead;
+    drwav_uint8 sampleData[4096];
+    drwav_uint32 bytesPerFrame;
+
+    if (pBufferOut == NULL) {
+        return drwav_read_pcm_frames(pWav, framesToRead, NULL);
+    }
+
+    bytesPerFrame = drwav_get_bytes_per_pcm_frame(pWav);
+    if (bytesPerFrame == 0) {
+        return 0;
+    }
+
+    totalFramesRead = 0;
+    
+    while (framesToRead > 0) {
+        drwav_uint64 framesRead = drwav_read_pcm_frames(pWav, drwav_min(framesToRead, sizeof(sampleData)/bytesPerFrame), sampleData);
+        if (framesRead == 0) {
+            break;
+        }
+
+        drwav__ieee_to_s16(pBufferOut, sampleData, (size_t)(framesRead*pWav->channels), bytesPerFrame/pWav->channels);
+
+        pBufferOut      += framesRead*pWav->channels;
+        framesToRead    -= framesRead;
+        totalFramesRead += framesRead;
+    }
+
+    return totalFramesRead;
+}
+
+static drwav_uint64 drwav_read_pcm_frames_s16__alaw(drwav* pWav, drwav_uint64 framesToRead, drwav_int16* pBufferOut)
+{
+    drwav_uint64 totalFramesRead;
+    drwav_uint8 sampleData[4096];
+    drwav_uint32 bytesPerFrame;
+
+    if (pBufferOut == NULL) {
+        return drwav_read_pcm_frames(pWav, framesToRead, NULL);
+    }
+
+    bytesPerFrame = drwav_get_bytes_per_pcm_frame(pWav);
+    if (bytesPerFrame == 0) {
+        return 0;
+    }
+
+    totalFramesRead = 0;
+    
+    while (framesToRead > 0) {
+        drwav_uint64 framesRead = drwav_read_pcm_frames(pWav, drwav_min(framesToRead, sizeof(sampleData)/bytesPerFrame), sampleData);
+        if (framesRead == 0) {
+            break;
+        }
+
+        drwav_alaw_to_s16(pBufferOut, sampleData, (size_t)(framesRead*pWav->channels));
+
+        pBufferOut      += framesRead*pWav->channels;
+        framesToRead    -= framesRead;
+        totalFramesRead += framesRead;
+    }
+
+    return totalFramesRead;
+}
+
+static drwav_uint64 drwav_read_pcm_frames_s16__mulaw(drwav* pWav, drwav_uint64 framesToRead, drwav_int16* pBufferOut)
+{
+    drwav_uint64 totalFramesRead;
+    drwav_uint8 sampleData[4096];
+    drwav_uint32 bytesPerFrame;
+
+    if (pBufferOut == NULL) {
+        return drwav_read_pcm_frames(pWav, framesToRead, NULL);
+    }
+
+    bytesPerFrame = drwav_get_bytes_per_pcm_frame(pWav);
+    if (bytesPerFrame == 0) {
+        return 0;
+    }
+
+    totalFramesRead = 0;
+
+    while (framesToRead > 0) {
+        drwav_uint64 framesRead = drwav_read_pcm_frames(pWav, drwav_min(framesToRead, sizeof(sampleData)/bytesPerFrame), sampleData);
+        if (framesRead == 0) {
+            break;
+        }
+
+        drwav_mulaw_to_s16(pBufferOut, sampleData, (size_t)(framesRead*pWav->channels));
+
+        pBufferOut      += framesRead*pWav->channels;
+        framesToRead    -= framesRead;
+        totalFramesRead += framesRead;
+    }
+
+    return totalFramesRead;
+}
+
+DRWAV_API drwav_uint64 drwav_read_pcm_frames_s16(drwav* pWav, drwav_uint64 framesToRead, drwav_int16* pBufferOut)
+{
+    if (pWav == NULL || framesToRead == 0) {
+        return 0;
+    }
+
+    if (pBufferOut == NULL) {
+        return drwav_read_pcm_frames(pWav, framesToRead, NULL);
+    }
+
+    /* Don't try to read more samples than can potentially fit in the output buffer. */
+    if (framesToRead * pWav->channels * sizeof(drwav_int16) > DRWAV_SIZE_MAX) {
+        framesToRead = DRWAV_SIZE_MAX / sizeof(drwav_int16) / pWav->channels;
+    }
+
+    if (pWav->translatedFormatTag == DR_WAVE_FORMAT_PCM) {
+        return drwav_read_pcm_frames_s16__pcm(pWav, framesToRead, pBufferOut);
+    }
+
+    if (pWav->translatedFormatTag == DR_WAVE_FORMAT_IEEE_FLOAT) {
+        return drwav_read_pcm_frames_s16__ieee(pWav, framesToRead, pBufferOut);
+    }
+
+    if (pWav->translatedFormatTag == DR_WAVE_FORMAT_ALAW) {
+        return drwav_read_pcm_frames_s16__alaw(pWav, framesToRead, pBufferOut);
+    }
+
+    if (pWav->translatedFormatTag == DR_WAVE_FORMAT_MULAW) {
+        return drwav_read_pcm_frames_s16__mulaw(pWav, framesToRead, pBufferOut);
+    }
+
+    if (pWav->translatedFormatTag == DR_WAVE_FORMAT_ADPCM) {
+        return drwav_read_pcm_frames_s16__msadpcm(pWav, framesToRead, pBufferOut);
+    }
+
+    if (pWav->translatedFormatTag == DR_WAVE_FORMAT_DVI_ADPCM) {
+        return drwav_read_pcm_frames_s16__ima(pWav, framesToRead, pBufferOut);
+    }
+
+    return 0;
+}
+
+DRWAV_API drwav_uint64 drwav_read_pcm_frames_s16le(drwav* pWav, drwav_uint64 framesToRead, drwav_int16* pBufferOut)
+{
+    drwav_uint64 framesRead = drwav_read_pcm_frames_s16(pWav, framesToRead, pBufferOut);
+    if (pBufferOut != NULL && drwav__is_little_endian() == DRWAV_FALSE) {
+        drwav__bswap_samples_s16(pBufferOut, framesRead*pWav->channels);
+    }
+
+    return framesRead;
+}
+
+DRWAV_API drwav_uint64 drwav_read_pcm_frames_s16be(drwav* pWav, drwav_uint64 framesToRead, drwav_int16* pBufferOut)
+{
+    drwav_uint64 framesRead = drwav_read_pcm_frames_s16(pWav, framesToRead, pBufferOut);
+    if (pBufferOut != NULL && drwav__is_little_endian() == DRWAV_TRUE) {
+        drwav__bswap_samples_s16(pBufferOut, framesRead*pWav->channels);
+    }
+
+    return framesRead;
+}
+
+
+DRWAV_API void drwav_u8_to_s16(drwav_int16* pOut, const drwav_uint8* pIn, size_t sampleCount)
+{
+    int r;
+    size_t i;
+    for (i = 0; i < sampleCount; ++i) {
+        int x = pIn[i];
+        r = x << 8;
+        r = r - 32768;
+        pOut[i] = (short)r;
+    }
+}
+
+DRWAV_API void drwav_s24_to_s16(drwav_int16* pOut, const drwav_uint8* pIn, size_t sampleCount)
+{
+    int r;
+    size_t i;
+    for (i = 0; i < sampleCount; ++i) {
+        int x = ((int)(((unsigned int)(((const drwav_uint8*)pIn)[i*3+0]) << 8) | ((unsigned int)(((const drwav_uint8*)pIn)[i*3+1]) << 16) | ((unsigned int)(((const drwav_uint8*)pIn)[i*3+2])) << 24)) >> 8;
+        r = x >> 8;
+        pOut[i] = (short)r;
+    }
+}
+
+DRWAV_API void drwav_s32_to_s16(drwav_int16* pOut, const drwav_int32* pIn, size_t sampleCount)
+{
+    int r;
+    size_t i;
+    for (i = 0; i < sampleCount; ++i) {
+        int x = pIn[i];
+        r = x >> 16;
+        pOut[i] = (short)r;
+    }
+}
+
+DRWAV_API void drwav_f32_to_s16(drwav_int16* pOut, const float* pIn, size_t sampleCount)
+{
+    int r;
+    size_t i;
+    for (i = 0; i < sampleCount; ++i) {
+        float x = pIn[i];
+        float c;
+        c = ((x < -1) ? -1 : ((x > 1) ? 1 : x));
+        c = c + 1;
+        r = (int)(c * 32767.5f);
+        r = r - 32768;
+        pOut[i] = (short)r;
+    }
+}
+
+DRWAV_API void drwav_f64_to_s16(drwav_int16* pOut, const double* pIn, size_t sampleCount)
+{
+    int r;
+    size_t i;
+    for (i = 0; i < sampleCount; ++i) {
+        double x = pIn[i];
+        double c;
+        c = ((x < -1) ? -1 : ((x > 1) ? 1 : x));
+        c = c + 1;
+        r = (int)(c * 32767.5);
+        r = r - 32768;
+        pOut[i] = (short)r;
+    }
+}
+
+DRWAV_API void drwav_alaw_to_s16(drwav_int16* pOut, const drwav_uint8* pIn, size_t sampleCount)
+{
+    size_t i;
+    for (i = 0; i < sampleCount; ++i) {
+        pOut[i] = drwav__alaw_to_s16(pIn[i]);
+    }
+}
+
+DRWAV_API void drwav_mulaw_to_s16(drwav_int16* pOut, const drwav_uint8* pIn, size_t sampleCount)
+{
+    size_t i;
+    for (i = 0; i < sampleCount; ++i) {
+        pOut[i] = drwav__mulaw_to_s16(pIn[i]);
+    }
+}
+
+
+
+static void drwav__pcm_to_f32(float* pOut, const drwav_uint8* pIn, size_t sampleCount, unsigned int bytesPerSample)
+{
+    unsigned int i;
+
+    /* Special case for 8-bit sample data because it's treated as unsigned. */
+    if (bytesPerSample == 1) {
+        drwav_u8_to_f32(pOut, pIn, sampleCount);
+        return;
+    }
+
+    /* Slightly more optimal implementation for common formats. */
+    if (bytesPerSample == 2) {
+        drwav_s16_to_f32(pOut, (const drwav_int16*)pIn, sampleCount);
+        return;
+    }
+    if (bytesPerSample == 3) {
+        drwav_s24_to_f32(pOut, pIn, sampleCount);
+        return;
+    }
+    if (bytesPerSample == 4) {
+        drwav_s32_to_f32(pOut, (const drwav_int32*)pIn, sampleCount);
+        return;
+    }
+
+
+    /* Anything more than 64 bits per sample is not supported. */
+    if (bytesPerSample > 8) {
+        DRWAV_ZERO_MEMORY(pOut, sampleCount * sizeof(*pOut));
+        return;
+    }
+
+
+    /* Generic, slow converter. */
+    for (i = 0; i < sampleCount; ++i) {
+        drwav_uint64 sample = 0;
+        unsigned int shift  = (8 - bytesPerSample) * 8;
+
+        unsigned int j;
+        for (j = 0; j < bytesPerSample; j += 1) {
+            DRWAV_ASSERT(j < 8);
+            sample |= (drwav_uint64)(pIn[j]) << shift;
+            shift  += 8;
+        }
+
+        pIn += j;
+        *pOut++ = (float)((drwav_int64)sample / 9223372036854775807.0);
+    }
+}
+
+static void drwav__ieee_to_f32(float* pOut, const drwav_uint8* pIn, size_t sampleCount, unsigned int bytesPerSample)
+{
+    if (bytesPerSample == 4) {
+        unsigned int i;
+        for (i = 0; i < sampleCount; ++i) {
+            *pOut++ = ((const float*)pIn)[i];
+        }
+        return;
+    } else if (bytesPerSample == 8) {
+        drwav_f64_to_f32(pOut, (const double*)pIn, sampleCount);
+        return;
+    } else {
+        /* Only supporting 32- and 64-bit float. Output silence in all other cases. Contributions welcome for 16-bit float. */
+        DRWAV_ZERO_MEMORY(pOut, sampleCount * sizeof(*pOut));
+        return;
+    }
+}
+
+
+static drwav_uint64 drwav_read_pcm_frames_f32__pcm(drwav* pWav, drwav_uint64 framesToRead, float* pBufferOut)
+{
+    drwav_uint64 totalFramesRead;
+    drwav_uint8 sampleData[4096];
+
+    drwav_uint32 bytesPerFrame = drwav_get_bytes_per_pcm_frame(pWav);
+    if (bytesPerFrame == 0) {
+        return 0;
+    }
+
+    totalFramesRead = 0;
+
+    while (framesToRead > 0) {
+        drwav_uint64 framesRead = drwav_read_pcm_frames(pWav, drwav_min(framesToRead, sizeof(sampleData)/bytesPerFrame), sampleData);
+        if (framesRead == 0) {
+            break;
+        }
+
+        drwav__pcm_to_f32(pBufferOut, sampleData, (size_t)framesRead*pWav->channels, bytesPerFrame/pWav->channels);
+
+        pBufferOut      += framesRead*pWav->channels;
+        framesToRead    -= framesRead;
+        totalFramesRead += framesRead;
+    }
+
+    return totalFramesRead;
+}
+
+static drwav_uint64 drwav_read_pcm_frames_f32__msadpcm(drwav* pWav, drwav_uint64 framesToRead, float* pBufferOut)
+{
+    /*
+    We're just going to borrow the implementation from the drwav_read_s16() since ADPCM is a little bit more complicated than other formats and I don't
+    want to duplicate that code.
+    */
+    drwav_uint64 totalFramesRead = 0;
+    drwav_int16 samples16[2048];
+    while (framesToRead > 0) {
+        drwav_uint64 framesRead = drwav_read_pcm_frames_s16(pWav, drwav_min(framesToRead, drwav_countof(samples16)/pWav->channels), samples16);
+        if (framesRead == 0) {
+            break;
+        }
+
+        drwav_s16_to_f32(pBufferOut, samples16, (size_t)(framesRead*pWav->channels));   /* <-- Safe cast because we're clamping to 2048. */
+
+        pBufferOut      += framesRead*pWav->channels;
+        framesToRead    -= framesRead;
+        totalFramesRead += framesRead;
+    }
+
+    return totalFramesRead;
+}
+
+static drwav_uint64 drwav_read_pcm_frames_f32__ima(drwav* pWav, drwav_uint64 framesToRead, float* pBufferOut)
+{
+    /*
+    We're just going to borrow the implementation from the drwav_read_s16() since IMA-ADPCM is a little bit more complicated than other formats and I don't
+    want to duplicate that code.
+    */
+    drwav_uint64 totalFramesRead = 0;
+    drwav_int16 samples16[2048];
+    while (framesToRead > 0) {
+        drwav_uint64 framesRead = drwav_read_pcm_frames_s16(pWav, drwav_min(framesToRead, drwav_countof(samples16)/pWav->channels), samples16);
+        if (framesRead == 0) {
+            break;
+        }
+
+        drwav_s16_to_f32(pBufferOut, samples16, (size_t)(framesRead*pWav->channels));   /* <-- Safe cast because we're clamping to 2048. */
+
+        pBufferOut      += framesRead*pWav->channels;
+        framesToRead    -= framesRead;
+        totalFramesRead += framesRead;
+    }
+
+    return totalFramesRead;
+}
+
+static drwav_uint64 drwav_read_pcm_frames_f32__ieee(drwav* pWav, drwav_uint64 framesToRead, float* pBufferOut)
+{
+    drwav_uint64 totalFramesRead;
+    drwav_uint8 sampleData[4096];
+    drwav_uint32 bytesPerFrame;
+
+    /* Fast path. */
+    if (pWav->translatedFormatTag == DR_WAVE_FORMAT_IEEE_FLOAT && pWav->bitsPerSample == 32) {
+        return drwav_read_pcm_frames(pWav, framesToRead, pBufferOut);
+    }
+    
+    bytesPerFrame = drwav_get_bytes_per_pcm_frame(pWav);
+    if (bytesPerFrame == 0) {
+        return 0;
+    }
+
+    totalFramesRead = 0;
+
+    while (framesToRead > 0) {
+        drwav_uint64 framesRead = drwav_read_pcm_frames(pWav, drwav_min(framesToRead, sizeof(sampleData)/bytesPerFrame), sampleData);
+        if (framesRead == 0) {
+            break;
+        }
+
+        drwav__ieee_to_f32(pBufferOut, sampleData, (size_t)(framesRead*pWav->channels), bytesPerFrame/pWav->channels);
+
+        pBufferOut      += framesRead*pWav->channels;
+        framesToRead    -= framesRead;
+        totalFramesRead += framesRead;
+    }
+
+    return totalFramesRead;
+}
+
+static drwav_uint64 drwav_read_pcm_frames_f32__alaw(drwav* pWav, drwav_uint64 framesToRead, float* pBufferOut)
+{
+    drwav_uint64 totalFramesRead;
+    drwav_uint8 sampleData[4096];
+    drwav_uint32 bytesPerFrame = drwav_get_bytes_per_pcm_frame(pWav);
+    if (bytesPerFrame == 0) {
+        return 0;
+    }
+
+    totalFramesRead = 0;
+
+    while (framesToRead > 0) {
+        drwav_uint64 framesRead = drwav_read_pcm_frames(pWav, drwav_min(framesToRead, sizeof(sampleData)/bytesPerFrame), sampleData);
+        if (framesRead == 0) {
+            break;
+        }
+
+        drwav_alaw_to_f32(pBufferOut, sampleData, (size_t)(framesRead*pWav->channels));
+
+        pBufferOut      += framesRead*pWav->channels;
+        framesToRead    -= framesRead;
+        totalFramesRead += framesRead;
+    }
+
+    return totalFramesRead;
+}
+
+static drwav_uint64 drwav_read_pcm_frames_f32__mulaw(drwav* pWav, drwav_uint64 framesToRead, float* pBufferOut)
+{
+    drwav_uint64 totalFramesRead;
+    drwav_uint8 sampleData[4096];
+
+    drwav_uint32 bytesPerFrame = drwav_get_bytes_per_pcm_frame(pWav);
+    if (bytesPerFrame == 0) {
+        return 0;
+    }
+
+    totalFramesRead = 0;
+
+    while (framesToRead > 0) {
+        drwav_uint64 framesRead = drwav_read_pcm_frames(pWav, drwav_min(framesToRead, sizeof(sampleData)/bytesPerFrame), sampleData);
+        if (framesRead == 0) {
+            break;
+        }
+
+        drwav_mulaw_to_f32(pBufferOut, sampleData, (size_t)(framesRead*pWav->channels));
+
+        pBufferOut      += framesRead*pWav->channels;
+        framesToRead    -= framesRead;
+        totalFramesRead += framesRead;
+    }
+
+    return totalFramesRead;
+}
+
+DRWAV_API drwav_uint64 drwav_read_pcm_frames_f32(drwav* pWav, drwav_uint64 framesToRead, float* pBufferOut)
+{
+    if (pWav == NULL || framesToRead == 0) {
+        return 0;
+    }
+
+    if (pBufferOut == NULL) {
+        return drwav_read_pcm_frames(pWav, framesToRead, NULL);
+    }
+
+    /* Don't try to read more samples than can potentially fit in the output buffer. */
+    if (framesToRead * pWav->channels * sizeof(float) > DRWAV_SIZE_MAX) {
+        framesToRead = DRWAV_SIZE_MAX / sizeof(float) / pWav->channels;
+    }
+
+    if (pWav->translatedFormatTag == DR_WAVE_FORMAT_PCM) {
+        return drwav_read_pcm_frames_f32__pcm(pWav, framesToRead, pBufferOut);
+    }
+
+    if (pWav->translatedFormatTag == DR_WAVE_FORMAT_ADPCM) {
+        return drwav_read_pcm_frames_f32__msadpcm(pWav, framesToRead, pBufferOut);
+    }
+
+    if (pWav->translatedFormatTag == DR_WAVE_FORMAT_IEEE_FLOAT) {
+        return drwav_read_pcm_frames_f32__ieee(pWav, framesToRead, pBufferOut);
+    }
+
+    if (pWav->translatedFormatTag == DR_WAVE_FORMAT_ALAW) {
+        return drwav_read_pcm_frames_f32__alaw(pWav, framesToRead, pBufferOut);
+    }
+
+    if (pWav->translatedFormatTag == DR_WAVE_FORMAT_MULAW) {
+        return drwav_read_pcm_frames_f32__mulaw(pWav, framesToRead, pBufferOut);
+    }
+
+    if (pWav->translatedFormatTag == DR_WAVE_FORMAT_DVI_ADPCM) {
+        return drwav_read_pcm_frames_f32__ima(pWav, framesToRead, pBufferOut);
+    }
+
+    return 0;
+}
+
+DRWAV_API drwav_uint64 drwav_read_pcm_frames_f32le(drwav* pWav, drwav_uint64 framesToRead, float* pBufferOut)
+{
+    drwav_uint64 framesRead = drwav_read_pcm_frames_f32(pWav, framesToRead, pBufferOut);
+    if (pBufferOut != NULL && drwav__is_little_endian() == DRWAV_FALSE) {
+        drwav__bswap_samples_f32(pBufferOut, framesRead*pWav->channels);
+    }
+
+    return framesRead;
+}
+
+DRWAV_API drwav_uint64 drwav_read_pcm_frames_f32be(drwav* pWav, drwav_uint64 framesToRead, float* pBufferOut)
+{
+    drwav_uint64 framesRead = drwav_read_pcm_frames_f32(pWav, framesToRead, pBufferOut);
+    if (pBufferOut != NULL && drwav__is_little_endian() == DRWAV_TRUE) {
+        drwav__bswap_samples_f32(pBufferOut, framesRead*pWav->channels);
+    }
+
+    return framesRead;
+}
+
+
+DRWAV_API void drwav_u8_to_f32(float* pOut, const drwav_uint8* pIn, size_t sampleCount)
+{
+    size_t i;
+
+    if (pOut == NULL || pIn == NULL) {
+        return;
+    }
+
+#ifdef DR_WAV_LIBSNDFILE_COMPAT
+    /*
+    It appears libsndfile uses slightly different logic for the u8 -> f32 conversion to dr_wav, which in my opinion is incorrect. It appears
+    libsndfile performs the conversion something like "f32 = (u8 / 256) * 2 - 1", however I think it should be "f32 = (u8 / 255) * 2 - 1" (note
+    the divisor of 256 vs 255). I use libsndfile as a benchmark for testing, so I'm therefore leaving this block here just for my automated
+    correctness testing. This is disabled by default.
+    */
+    for (i = 0; i < sampleCount; ++i) {
+        *pOut++ = (pIn[i] / 256.0f) * 2 - 1;
+    }
+#else
+    for (i = 0; i < sampleCount; ++i) {
+        float x = pIn[i];
+        x = x * 0.00784313725490196078f;    /* 0..255 to 0..2 */
+        x = x - 1;                          /* 0..2 to -1..1 */
+
+        *pOut++ = x;
+    }
+#endif
+}
+
+DRWAV_API void drwav_s16_to_f32(float* pOut, const drwav_int16* pIn, size_t sampleCount)
+{
+    size_t i;
+
+    if (pOut == NULL || pIn == NULL) {
+        return;
+    }
+
+    for (i = 0; i < sampleCount; ++i) {
+        *pOut++ = pIn[i] * 0.000030517578125f;
+    }
+}
+
+DRWAV_API void drwav_s24_to_f32(float* pOut, const drwav_uint8* pIn, size_t sampleCount)
+{
+    size_t i;
+
+    if (pOut == NULL || pIn == NULL) {
+        return;
+    }
+
+    for (i = 0; i < sampleCount; ++i) {
+        double x;
+        drwav_uint32 a = ((drwav_uint32)(pIn[i*3+0]) <<  8);
+        drwav_uint32 b = ((drwav_uint32)(pIn[i*3+1]) << 16);
+        drwav_uint32 c = ((drwav_uint32)(pIn[i*3+2]) << 24);
+
+        x = (double)((drwav_int32)(a | b | c) >> 8);
+        *pOut++ = (float)(x * 0.00000011920928955078125);
+    }
+}
+
+DRWAV_API void drwav_s32_to_f32(float* pOut, const drwav_int32* pIn, size_t sampleCount)
+{
+    size_t i;
+    if (pOut == NULL || pIn == NULL) {
+        return;
+    }
+
+    for (i = 0; i < sampleCount; ++i) {
+        *pOut++ = (float)(pIn[i] / 2147483648.0);
+    }
+}
+
+DRWAV_API void drwav_f64_to_f32(float* pOut, const double* pIn, size_t sampleCount)
+{
+    size_t i;
+
+    if (pOut == NULL || pIn == NULL) {
+        return;
+    }
+
+    for (i = 0; i < sampleCount; ++i) {
+        *pOut++ = (float)pIn[i];
+    }
+}
+
+DRWAV_API void drwav_alaw_to_f32(float* pOut, const drwav_uint8* pIn, size_t sampleCount)
+{
+    size_t i;
+
+    if (pOut == NULL || pIn == NULL) {
+        return;
+    }
+
+    for (i = 0; i < sampleCount; ++i) {
+        *pOut++ = drwav__alaw_to_s16(pIn[i]) / 32768.0f;
+    }
+}
+
+DRWAV_API void drwav_mulaw_to_f32(float* pOut, const drwav_uint8* pIn, size_t sampleCount)
+{
+    size_t i;
+
+    if (pOut == NULL || pIn == NULL) {
+        return;
+    }
+
+    for (i = 0; i < sampleCount; ++i) {
+        *pOut++ = drwav__mulaw_to_s16(pIn[i]) / 32768.0f;
+    }
+}
+
+
+
+static void drwav__pcm_to_s32(drwav_int32* pOut, const drwav_uint8* pIn, size_t totalSampleCount, unsigned int bytesPerSample)
+{
+    unsigned int i;
+
+    /* Special case for 8-bit sample data because it's treated as unsigned. */
+    if (bytesPerSample == 1) {
+        drwav_u8_to_s32(pOut, pIn, totalSampleCount);
+        return;
+    }
+
+    /* Slightly more optimal implementation for common formats. */
+    if (bytesPerSample == 2) {
+        drwav_s16_to_s32(pOut, (const drwav_int16*)pIn, totalSampleCount);
+        return;
+    }
+    if (bytesPerSample == 3) {
+        drwav_s24_to_s32(pOut, pIn, totalSampleCount);
+        return;
+    }
+    if (bytesPerSample == 4) {
+        for (i = 0; i < totalSampleCount; ++i) {
+           *pOut++ = ((const drwav_int32*)pIn)[i];
+        }
+        return;
+    }
+
+
+    /* Anything more than 64 bits per sample is not supported. */
+    if (bytesPerSample > 8) {
+        DRWAV_ZERO_MEMORY(pOut, totalSampleCount * sizeof(*pOut));
+        return;
+    }
+
+
+    /* Generic, slow converter. */
+    for (i = 0; i < totalSampleCount; ++i) {
+        drwav_uint64 sample = 0;
+        unsigned int shift  = (8 - bytesPerSample) * 8;
+
+        unsigned int j;
+        for (j = 0; j < bytesPerSample; j += 1) {
+            DRWAV_ASSERT(j < 8);
+            sample |= (drwav_uint64)(pIn[j]) << shift;
+            shift  += 8;
+        }
+
+        pIn += j;
+        *pOut++ = (drwav_int32)((drwav_int64)sample >> 32);
+    }
+}
+
+static void drwav__ieee_to_s32(drwav_int32* pOut, const drwav_uint8* pIn, size_t totalSampleCount, unsigned int bytesPerSample)
+{
+    if (bytesPerSample == 4) {
+        drwav_f32_to_s32(pOut, (const float*)pIn, totalSampleCount);
+        return;
+    } else if (bytesPerSample == 8) {
+        drwav_f64_to_s32(pOut, (const double*)pIn, totalSampleCount);
+        return;
+    } else {
+        /* Only supporting 32- and 64-bit float. Output silence in all other cases. Contributions welcome for 16-bit float. */
+        DRWAV_ZERO_MEMORY(pOut, totalSampleCount * sizeof(*pOut));
+        return;
+    }
+}
+
+
+static drwav_uint64 drwav_read_pcm_frames_s32__pcm(drwav* pWav, drwav_uint64 framesToRead, drwav_int32* pBufferOut)
+{
+    drwav_uint64 totalFramesRead;
+    drwav_uint8 sampleData[4096];
+    drwav_uint32 bytesPerFrame;
+
+    /* Fast path. */
+    if (pWav->translatedFormatTag == DR_WAVE_FORMAT_PCM && pWav->bitsPerSample == 32) {
+        return drwav_read_pcm_frames(pWav, framesToRead, pBufferOut);
+    }
+    
+    bytesPerFrame = drwav_get_bytes_per_pcm_frame(pWav);
+    if (bytesPerFrame == 0) {
+        return 0;
+    }
+
+    totalFramesRead = 0;
+
+    while (framesToRead > 0) {
+        drwav_uint64 framesRead = drwav_read_pcm_frames(pWav, drwav_min(framesToRead, sizeof(sampleData)/bytesPerFrame), sampleData);
+        if (framesRead == 0) {
+            break;
+        }
+
+        drwav__pcm_to_s32(pBufferOut, sampleData, (size_t)(framesRead*pWav->channels), bytesPerFrame/pWav->channels);
+
+        pBufferOut      += framesRead*pWav->channels;
+        framesToRead    -= framesRead;
+        totalFramesRead += framesRead;
+    }
+
+    return totalFramesRead;
+}
+
+static drwav_uint64 drwav_read_pcm_frames_s32__msadpcm(drwav* pWav, drwav_uint64 framesToRead, drwav_int32* pBufferOut)
+{
+    /*
+    We're just going to borrow the implementation from the drwav_read_s16() since ADPCM is a little bit more complicated than other formats and I don't
+    want to duplicate that code.
+    */
+    drwav_uint64 totalFramesRead = 0;
+    drwav_int16 samples16[2048];
+    while (framesToRead > 0) {
+        drwav_uint64 framesRead = drwav_read_pcm_frames_s16(pWav, drwav_min(framesToRead, drwav_countof(samples16)/pWav->channels), samples16);
+        if (framesRead == 0) {
+            break;
+        }
+
+        drwav_s16_to_s32(pBufferOut, samples16, (size_t)(framesRead*pWav->channels));   /* <-- Safe cast because we're clamping to 2048. */
+
+        pBufferOut      += framesRead*pWav->channels;
+        framesToRead    -= framesRead;
+        totalFramesRead += framesRead;
+    }
+
+    return totalFramesRead;
+}
+
+static drwav_uint64 drwav_read_pcm_frames_s32__ima(drwav* pWav, drwav_uint64 framesToRead, drwav_int32* pBufferOut)
+{
+    /*
+    We're just going to borrow the implementation from the drwav_read_s16() since IMA-ADPCM is a little bit more complicated than other formats and I don't
+    want to duplicate that code.
+    */
+    drwav_uint64 totalFramesRead = 0;
+    drwav_int16 samples16[2048];
+    while (framesToRead > 0) {
+        drwav_uint64 framesRead = drwav_read_pcm_frames_s16(pWav, drwav_min(framesToRead, drwav_countof(samples16)/pWav->channels), samples16);
+        if (framesRead == 0) {
+            break;
+        }
+
+        drwav_s16_to_s32(pBufferOut, samples16, (size_t)(framesRead*pWav->channels));   /* <-- Safe cast because we're clamping to 2048. */
+
+        pBufferOut      += framesRead*pWav->channels;
+        framesToRead    -= framesRead;
+        totalFramesRead += framesRead;
+    }
+
+    return totalFramesRead;
+}
+
+static drwav_uint64 drwav_read_pcm_frames_s32__ieee(drwav* pWav, drwav_uint64 framesToRead, drwav_int32* pBufferOut)
+{
+    drwav_uint64 totalFramesRead;
+    drwav_uint8 sampleData[4096];
+
+    drwav_uint32 bytesPerFrame = drwav_get_bytes_per_pcm_frame(pWav);
+    if (bytesPerFrame == 0) {
+        return 0;
+    }
+
+    totalFramesRead = 0;
+
+    while (framesToRead > 0) {
+        drwav_uint64 framesRead = drwav_read_pcm_frames(pWav, drwav_min(framesToRead, sizeof(sampleData)/bytesPerFrame), sampleData);
+        if (framesRead == 0) {
+            break;
+        }
+
+        drwav__ieee_to_s32(pBufferOut, sampleData, (size_t)(framesRead*pWav->channels), bytesPerFrame/pWav->channels);
+
+        pBufferOut      += framesRead*pWav->channels;
+        framesToRead    -= framesRead;
+        totalFramesRead += framesRead;
+    }
+
+    return totalFramesRead;
+}
+
+static drwav_uint64 drwav_read_pcm_frames_s32__alaw(drwav* pWav, drwav_uint64 framesToRead, drwav_int32* pBufferOut)
+{
+    drwav_uint64 totalFramesRead;
+    drwav_uint8 sampleData[4096];
+
+    drwav_uint32 bytesPerFrame = drwav_get_bytes_per_pcm_frame(pWav);
+    if (bytesPerFrame == 0) {
+        return 0;
+    }
+
+    totalFramesRead = 0;
+
+    while (framesToRead > 0) {
+        drwav_uint64 framesRead = drwav_read_pcm_frames(pWav, drwav_min(framesToRead, sizeof(sampleData)/bytesPerFrame), sampleData);
+        if (framesRead == 0) {
+            break;
+        }
+
+        drwav_alaw_to_s32(pBufferOut, sampleData, (size_t)(framesRead*pWav->channels));
+
+        pBufferOut      += framesRead*pWav->channels;
+        framesToRead    -= framesRead;
+        totalFramesRead += framesRead;
+    }
+
+    return totalFramesRead;
+}
+
+static drwav_uint64 drwav_read_pcm_frames_s32__mulaw(drwav* pWav, drwav_uint64 framesToRead, drwav_int32* pBufferOut)
+{
+    drwav_uint64 totalFramesRead;
+    drwav_uint8 sampleData[4096];
+
+    drwav_uint32 bytesPerFrame = drwav_get_bytes_per_pcm_frame(pWav);
+    if (bytesPerFrame == 0) {
+        return 0;
+    }
+
+    totalFramesRead = 0;
+
+    while (framesToRead > 0) {
+        drwav_uint64 framesRead = drwav_read_pcm_frames(pWav, drwav_min(framesToRead, sizeof(sampleData)/bytesPerFrame), sampleData);
+        if (framesRead == 0) {
+            break;
+        }
+
+        drwav_mulaw_to_s32(pBufferOut, sampleData, (size_t)(framesRead*pWav->channels));
+
+        pBufferOut      += framesRead*pWav->channels;
+        framesToRead    -= framesRead;
+        totalFramesRead += framesRead;
+    }
+
+    return totalFramesRead;
+}
+
+DRWAV_API drwav_uint64 drwav_read_pcm_frames_s32(drwav* pWav, drwav_uint64 framesToRead, drwav_int32* pBufferOut)
+{
+    if (pWav == NULL || framesToRead == 0) {
+        return 0;
+    }
+
+    if (pBufferOut == NULL) {
+        return drwav_read_pcm_frames(pWav, framesToRead, NULL);
+    }
+
+    /* Don't try to read more samples than can potentially fit in the output buffer. */
+    if (framesToRead * pWav->channels * sizeof(drwav_int32) > DRWAV_SIZE_MAX) {
+        framesToRead = DRWAV_SIZE_MAX / sizeof(drwav_int32) / pWav->channels;
+    }
+
+    if (pWav->translatedFormatTag == DR_WAVE_FORMAT_PCM) {
+        return drwav_read_pcm_frames_s32__pcm(pWav, framesToRead, pBufferOut);
+    }
+
+    if (pWav->translatedFormatTag == DR_WAVE_FORMAT_ADPCM) {
+        return drwav_read_pcm_frames_s32__msadpcm(pWav, framesToRead, pBufferOut);
+    }
+
+    if (pWav->translatedFormatTag == DR_WAVE_FORMAT_IEEE_FLOAT) {
+        return drwav_read_pcm_frames_s32__ieee(pWav, framesToRead, pBufferOut);
+    }
+
+    if (pWav->translatedFormatTag == DR_WAVE_FORMAT_ALAW) {
+        return drwav_read_pcm_frames_s32__alaw(pWav, framesToRead, pBufferOut);
+    }
+
+    if (pWav->translatedFormatTag == DR_WAVE_FORMAT_MULAW) {
+        return drwav_read_pcm_frames_s32__mulaw(pWav, framesToRead, pBufferOut);
+    }
+
+    if (pWav->translatedFormatTag == DR_WAVE_FORMAT_DVI_ADPCM) {
+        return drwav_read_pcm_frames_s32__ima(pWav, framesToRead, pBufferOut);
+    }
+
+    return 0;
+}
+
+DRWAV_API drwav_uint64 drwav_read_pcm_frames_s32le(drwav* pWav, drwav_uint64 framesToRead, drwav_int32* pBufferOut)
+{
+    drwav_uint64 framesRead = drwav_read_pcm_frames_s32(pWav, framesToRead, pBufferOut);
+    if (pBufferOut != NULL && drwav__is_little_endian() == DRWAV_FALSE) {
+        drwav__bswap_samples_s32(pBufferOut, framesRead*pWav->channels);
+    }
+
+    return framesRead;
+}
+
+DRWAV_API drwav_uint64 drwav_read_pcm_frames_s32be(drwav* pWav, drwav_uint64 framesToRead, drwav_int32* pBufferOut)
+{
+    drwav_uint64 framesRead = drwav_read_pcm_frames_s32(pWav, framesToRead, pBufferOut);
+    if (pBufferOut != NULL && drwav__is_little_endian() == DRWAV_TRUE) {
+        drwav__bswap_samples_s32(pBufferOut, framesRead*pWav->channels);
+    }
+
+    return framesRead;
+}
+
+
+DRWAV_API void drwav_u8_to_s32(drwav_int32* pOut, const drwav_uint8* pIn, size_t sampleCount)
+{
+    size_t i;
+
+    if (pOut == NULL || pIn == NULL) {
+        return;
+    }
+
+    for (i = 0; i < sampleCount; ++i) {
+        *pOut++ = ((int)pIn[i] - 128) << 24;
+    }
+}
+
+DRWAV_API void drwav_s16_to_s32(drwav_int32* pOut, const drwav_int16* pIn, size_t sampleCount)
+{
+    size_t i;
+
+    if (pOut == NULL || pIn == NULL) {
+        return;
+    }
+
+    for (i = 0; i < sampleCount; ++i) {
+        *pOut++ = pIn[i] << 16;
+    }
+}
+
+DRWAV_API void drwav_s24_to_s32(drwav_int32* pOut, const drwav_uint8* pIn, size_t sampleCount)
+{
+    size_t i;
+
+    if (pOut == NULL || pIn == NULL) {
+        return;
+    }
+
+    for (i = 0; i < sampleCount; ++i) {
+        unsigned int s0 = pIn[i*3 + 0];
+        unsigned int s1 = pIn[i*3 + 1];
+        unsigned int s2 = pIn[i*3 + 2];
+
+        drwav_int32 sample32 = (drwav_int32)((s0 << 8) | (s1 << 16) | (s2 << 24));
+        *pOut++ = sample32;
+    }
+}
+
+DRWAV_API void drwav_f32_to_s32(drwav_int32* pOut, const float* pIn, size_t sampleCount)
+{
+    size_t i;
+
+    if (pOut == NULL || pIn == NULL) {
+        return;
+    }
+
+    for (i = 0; i < sampleCount; ++i) {
+        *pOut++ = (drwav_int32)(2147483648.0 * pIn[i]);
+    }
+}
+
+DRWAV_API void drwav_f64_to_s32(drwav_int32* pOut, const double* pIn, size_t sampleCount)
+{
+    size_t i;
+
+    if (pOut == NULL || pIn == NULL) {
+        return;
+    }
+
+    for (i = 0; i < sampleCount; ++i) {
+        *pOut++ = (drwav_int32)(2147483648.0 * pIn[i]);
+    }
+}
+
+DRWAV_API void drwav_alaw_to_s32(drwav_int32* pOut, const drwav_uint8* pIn, size_t sampleCount)
+{
+    size_t i;
+
+    if (pOut == NULL || pIn == NULL) {
+        return;
+    }
+
+    for (i = 0; i < sampleCount; ++i) {
+        *pOut++ = ((drwav_int32)drwav__alaw_to_s16(pIn[i])) << 16;
+    }
+}
+
+DRWAV_API void drwav_mulaw_to_s32(drwav_int32* pOut, const drwav_uint8* pIn, size_t sampleCount)
+{
+    size_t i;
+
+    if (pOut == NULL || pIn == NULL) {
+        return;
+    }
+
+    for (i= 0; i < sampleCount; ++i) {
+        *pOut++ = ((drwav_int32)drwav__mulaw_to_s16(pIn[i])) << 16;
+    }
+}
+
+
+
+static drwav_int16* drwav__read_pcm_frames_and_close_s16(drwav* pWav, unsigned int* channels, unsigned int* sampleRate, drwav_uint64* totalFrameCount)
+{
+    drwav_uint64 sampleDataSize;
+    drwav_int16* pSampleData;
+    drwav_uint64 framesRead;
+
+    DRWAV_ASSERT(pWav != NULL);
+
+    sampleDataSize = pWav->totalPCMFrameCount * pWav->channels * sizeof(drwav_int16);
+    if (sampleDataSize > DRWAV_SIZE_MAX) {
+        drwav_uninit(pWav);
+        return NULL;    /* File's too big. */
+    }
+
+    pSampleData = (drwav_int16*)drwav__malloc_from_callbacks((size_t)sampleDataSize, &pWav->allocationCallbacks); /* <-- Safe cast due to the check above. */
+    if (pSampleData == NULL) {
+        drwav_uninit(pWav);
+        return NULL;    /* Failed to allocate memory. */
+    }
+
+    framesRead = drwav_read_pcm_frames_s16(pWav, (size_t)pWav->totalPCMFrameCount, pSampleData);
+    if (framesRead != pWav->totalPCMFrameCount) {
+        drwav__free_from_callbacks(pSampleData, &pWav->allocationCallbacks);
+        drwav_uninit(pWav);
+        return NULL;    /* There was an error reading the samples. */
+    }
+
+    drwav_uninit(pWav);
+
+    if (sampleRate) {
+        *sampleRate = pWav->sampleRate;
+    }
+    if (channels) {
+        *channels = pWav->channels;
+    }
+    if (totalFrameCount) {
+        *totalFrameCount = pWav->totalPCMFrameCount;
+    }
+
+    return pSampleData;
+}
+
+static float* drwav__read_pcm_frames_and_close_f32(drwav* pWav, unsigned int* channels, unsigned int* sampleRate, drwav_uint64* totalFrameCount)
+{
+    drwav_uint64 sampleDataSize;
+    float* pSampleData;
+    drwav_uint64 framesRead;
+
+    DRWAV_ASSERT(pWav != NULL);
+
+    sampleDataSize = pWav->totalPCMFrameCount * pWav->channels * sizeof(float);
+    if (sampleDataSize > DRWAV_SIZE_MAX) {
+        drwav_uninit(pWav);
+        return NULL;    /* File's too big. */
+    }
+
+    pSampleData = (float*)drwav__malloc_from_callbacks((size_t)sampleDataSize, &pWav->allocationCallbacks); /* <-- Safe cast due to the check above. */
+    if (pSampleData == NULL) {
+        drwav_uninit(pWav);
+        return NULL;    /* Failed to allocate memory. */
+    }
+
+    framesRead = drwav_read_pcm_frames_f32(pWav, (size_t)pWav->totalPCMFrameCount, pSampleData);
+    if (framesRead != pWav->totalPCMFrameCount) {
+        drwav__free_from_callbacks(pSampleData, &pWav->allocationCallbacks);
+        drwav_uninit(pWav);
+        return NULL;    /* There was an error reading the samples. */
+    }
+
+    drwav_uninit(pWav);
+
+    if (sampleRate) {
+        *sampleRate = pWav->sampleRate;
+    }
+    if (channels) {
+        *channels = pWav->channels;
+    }
+    if (totalFrameCount) {
+        *totalFrameCount = pWav->totalPCMFrameCount;
+    }
+
+    return pSampleData;
+}
+
+static drwav_int32* drwav__read_pcm_frames_and_close_s32(drwav* pWav, unsigned int* channels, unsigned int* sampleRate, drwav_uint64* totalFrameCount)
+{
+    drwav_uint64 sampleDataSize;
+    drwav_int32* pSampleData;
+    drwav_uint64 framesRead;
+
+    DRWAV_ASSERT(pWav != NULL);
+
+    sampleDataSize = pWav->totalPCMFrameCount * pWav->channels * sizeof(drwav_int32);
+    if (sampleDataSize > DRWAV_SIZE_MAX) {
+        drwav_uninit(pWav);
+        return NULL;    /* File's too big. */
+    }
+
+    pSampleData = (drwav_int32*)drwav__malloc_from_callbacks((size_t)sampleDataSize, &pWav->allocationCallbacks); /* <-- Safe cast due to the check above. */
+    if (pSampleData == NULL) {
+        drwav_uninit(pWav);
+        return NULL;    /* Failed to allocate memory. */
+    }
+
+    framesRead = drwav_read_pcm_frames_s32(pWav, (size_t)pWav->totalPCMFrameCount, pSampleData);
+    if (framesRead != pWav->totalPCMFrameCount) {
+        drwav__free_from_callbacks(pSampleData, &pWav->allocationCallbacks);
+        drwav_uninit(pWav);
+        return NULL;    /* There was an error reading the samples. */
+    }
+
+    drwav_uninit(pWav);
+
+    if (sampleRate) {
+        *sampleRate = pWav->sampleRate;
+    }
+    if (channels) {
+        *channels = pWav->channels;
+    }
+    if (totalFrameCount) {
+        *totalFrameCount = pWav->totalPCMFrameCount;
+    }
+
+    return pSampleData;
+}
+
+
+
+DRWAV_API drwav_int16* drwav_open_and_read_pcm_frames_s16(drwav_read_proc onRead, drwav_seek_proc onSeek, void* pUserData, unsigned int* channelsOut, unsigned int* sampleRateOut, drwav_uint64* totalFrameCountOut, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    drwav wav;
+
+    if (channelsOut) {
+        *channelsOut = 0;
+    }
+    if (sampleRateOut) {
+        *sampleRateOut = 0;
+    }
+    if (totalFrameCountOut) {
+        *totalFrameCountOut = 0;
+    }
+
+    if (!drwav_init(&wav, onRead, onSeek, pUserData, pAllocationCallbacks)) {
+        return NULL;
+    }
+
+    return drwav__read_pcm_frames_and_close_s16(&wav, channelsOut, sampleRateOut, totalFrameCountOut);
+}
+
+DRWAV_API float* drwav_open_and_read_pcm_frames_f32(drwav_read_proc onRead, drwav_seek_proc onSeek, void* pUserData, unsigned int* channelsOut, unsigned int* sampleRateOut, drwav_uint64* totalFrameCountOut, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    drwav wav;
+
+    if (channelsOut) {
+        *channelsOut = 0;
+    }
+    if (sampleRateOut) {
+        *sampleRateOut = 0;
+    }
+    if (totalFrameCountOut) {
+        *totalFrameCountOut = 0;
+    }
+
+    if (!drwav_init(&wav, onRead, onSeek, pUserData, pAllocationCallbacks)) {
+        return NULL;
+    }
+
+    return drwav__read_pcm_frames_and_close_f32(&wav, channelsOut, sampleRateOut, totalFrameCountOut);
+}
+
+DRWAV_API drwav_int32* drwav_open_and_read_pcm_frames_s32(drwav_read_proc onRead, drwav_seek_proc onSeek, void* pUserData, unsigned int* channelsOut, unsigned int* sampleRateOut, drwav_uint64* totalFrameCountOut, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    drwav wav;
+
+    if (channelsOut) {
+        *channelsOut = 0;
+    }
+    if (sampleRateOut) {
+        *sampleRateOut = 0;
+    }
+    if (totalFrameCountOut) {
+        *totalFrameCountOut = 0;
+    }
+
+    if (!drwav_init(&wav, onRead, onSeek, pUserData, pAllocationCallbacks)) {
+        return NULL;
+    }
+
+    return drwav__read_pcm_frames_and_close_s32(&wav, channelsOut, sampleRateOut, totalFrameCountOut);
+}
+
+#ifndef DR_WAV_NO_STDIO
+DRWAV_API drwav_int16* drwav_open_file_and_read_pcm_frames_s16(const char* filename, unsigned int* channelsOut, unsigned int* sampleRateOut, drwav_uint64* totalFrameCountOut, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    drwav wav;
+
+    if (channelsOut) {
+        *channelsOut = 0;
+    }
+    if (sampleRateOut) {
+        *sampleRateOut = 0;
+    }
+    if (totalFrameCountOut) {
+        *totalFrameCountOut = 0;
+    }
+
+    if (!drwav_init_file(&wav, filename, pAllocationCallbacks)) {
+        return NULL;
+    }
+
+    return drwav__read_pcm_frames_and_close_s16(&wav, channelsOut, sampleRateOut, totalFrameCountOut);
+}
+
+DRWAV_API float* drwav_open_file_and_read_pcm_frames_f32(const char* filename, unsigned int* channelsOut, unsigned int* sampleRateOut, drwav_uint64* totalFrameCountOut, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    drwav wav;
+
+    if (channelsOut) {
+        *channelsOut = 0;
+    }
+    if (sampleRateOut) {
+        *sampleRateOut = 0;
+    }
+    if (totalFrameCountOut) {
+        *totalFrameCountOut = 0;
+    }
+
+    if (!drwav_init_file(&wav, filename, pAllocationCallbacks)) {
+        return NULL;
+    }
+
+    return drwav__read_pcm_frames_and_close_f32(&wav, channelsOut, sampleRateOut, totalFrameCountOut);
+}
+
+DRWAV_API drwav_int32* drwav_open_file_and_read_pcm_frames_s32(const char* filename, unsigned int* channelsOut, unsigned int* sampleRateOut, drwav_uint64* totalFrameCountOut, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    drwav wav;
+
+    if (channelsOut) {
+        *channelsOut = 0;
+    }
+    if (sampleRateOut) {
+        *sampleRateOut = 0;
+    }
+    if (totalFrameCountOut) {
+        *totalFrameCountOut = 0;
+    }
+
+    if (!drwav_init_file(&wav, filename, pAllocationCallbacks)) {
+        return NULL;
+    }
+
+    return drwav__read_pcm_frames_and_close_s32(&wav, channelsOut, sampleRateOut, totalFrameCountOut);
+}
+
+
+DRWAV_API drwav_int16* drwav_open_file_and_read_pcm_frames_s16_w(const wchar_t* filename, unsigned int* channelsOut, unsigned int* sampleRateOut, drwav_uint64* totalFrameCountOut, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    drwav wav;
+
+    if (sampleRateOut) {
+        *sampleRateOut = 0;
+    }
+    if (channelsOut) {
+        *channelsOut = 0;
+    }
+    if (totalFrameCountOut) {
+        *totalFrameCountOut = 0;
+    }
+
+    if (!drwav_init_file_w(&wav, filename, pAllocationCallbacks)) {
+        return NULL;
+    }
+
+    return drwav__read_pcm_frames_and_close_s16(&wav, channelsOut, sampleRateOut, totalFrameCountOut);
+}
+
+DRWAV_API float* drwav_open_file_and_read_pcm_frames_f32_w(const wchar_t* filename, unsigned int* channelsOut, unsigned int* sampleRateOut, drwav_uint64* totalFrameCountOut, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    drwav wav;
+
+    if (sampleRateOut) {
+        *sampleRateOut = 0;
+    }
+    if (channelsOut) {
+        *channelsOut = 0;
+    }
+    if (totalFrameCountOut) {
+        *totalFrameCountOut = 0;
+    }
+
+    if (!drwav_init_file_w(&wav, filename, pAllocationCallbacks)) {
+        return NULL;
+    }
+
+    return drwav__read_pcm_frames_and_close_f32(&wav, channelsOut, sampleRateOut, totalFrameCountOut);
+}
+
+DRWAV_API drwav_int32* drwav_open_file_and_read_pcm_frames_s32_w(const wchar_t* filename, unsigned int* channelsOut, unsigned int* sampleRateOut, drwav_uint64* totalFrameCountOut, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    drwav wav;
+
+    if (sampleRateOut) {
+        *sampleRateOut = 0;
+    }
+    if (channelsOut) {
+        *channelsOut = 0;
+    }
+    if (totalFrameCountOut) {
+        *totalFrameCountOut = 0;
+    }
+
+    if (!drwav_init_file_w(&wav, filename, pAllocationCallbacks)) {
+        return NULL;
+    }
+
+    return drwav__read_pcm_frames_and_close_s32(&wav, channelsOut, sampleRateOut, totalFrameCountOut);
+}
+#endif
+
+DRWAV_API drwav_int16* drwav_open_memory_and_read_pcm_frames_s16(const void* data, size_t dataSize, unsigned int* channelsOut, unsigned int* sampleRateOut, drwav_uint64* totalFrameCountOut, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    drwav wav;
+
+    if (channelsOut) {
+        *channelsOut = 0;
+    }
+    if (sampleRateOut) {
+        *sampleRateOut = 0;
+    }
+    if (totalFrameCountOut) {
+        *totalFrameCountOut = 0;
+    }
+
+    if (!drwav_init_memory(&wav, data, dataSize, pAllocationCallbacks)) {
+        return NULL;
+    }
+
+    return drwav__read_pcm_frames_and_close_s16(&wav, channelsOut, sampleRateOut, totalFrameCountOut);
+}
+
+DRWAV_API float* drwav_open_memory_and_read_pcm_frames_f32(const void* data, size_t dataSize, unsigned int* channelsOut, unsigned int* sampleRateOut, drwav_uint64* totalFrameCountOut, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    drwav wav;
+
+    if (channelsOut) {
+        *channelsOut = 0;
+    }
+    if (sampleRateOut) {
+        *sampleRateOut = 0;
+    }
+    if (totalFrameCountOut) {
+        *totalFrameCountOut = 0;
+    }
+
+    if (!drwav_init_memory(&wav, data, dataSize, pAllocationCallbacks)) {
+        return NULL;
+    }
+
+    return drwav__read_pcm_frames_and_close_f32(&wav, channelsOut, sampleRateOut, totalFrameCountOut);
+}
+
+DRWAV_API drwav_int32* drwav_open_memory_and_read_pcm_frames_s32(const void* data, size_t dataSize, unsigned int* channelsOut, unsigned int* sampleRateOut, drwav_uint64* totalFrameCountOut, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    drwav wav;
+
+    if (channelsOut) {
+        *channelsOut = 0;
+    }
+    if (sampleRateOut) {
+        *sampleRateOut = 0;
+    }
+    if (totalFrameCountOut) {
+        *totalFrameCountOut = 0;
+    }
+
+    if (!drwav_init_memory(&wav, data, dataSize, pAllocationCallbacks)) {
+        return NULL;
+    }
+
+    return drwav__read_pcm_frames_and_close_s32(&wav, channelsOut, sampleRateOut, totalFrameCountOut);
+}
+#endif  /* DR_WAV_NO_CONVERSION_API */
+
+
+DRWAV_API void drwav_free(void* p, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    if (pAllocationCallbacks != NULL) {
+        drwav__free_from_callbacks(p, pAllocationCallbacks);
+    } else {
+        drwav__free_default(p, NULL);
+    }
+}
+
+DRWAV_API drwav_uint16 drwav_bytes_to_u16(const drwav_uint8* data)
+{
+    return drwav__bytes_to_u16(data);
+}
+
+DRWAV_API drwav_int16 drwav_bytes_to_s16(const drwav_uint8* data)
+{
+    return drwav__bytes_to_s16(data);
+}
+
+DRWAV_API drwav_uint32 drwav_bytes_to_u32(const drwav_uint8* data)
+{
+    return drwav__bytes_to_u32(data);
+}
+
+DRWAV_API drwav_int32 drwav_bytes_to_s32(const drwav_uint8* data)
+{
+    return drwav__bytes_to_s32(data);
+}
+
+DRWAV_API drwav_uint64 drwav_bytes_to_u64(const drwav_uint8* data)
+{
+    return drwav__bytes_to_u64(data);
+}
+
+DRWAV_API drwav_int64 drwav_bytes_to_s64(const drwav_uint8* data)
+{
+    return drwav__bytes_to_s64(data);
+}
+
+
+DRWAV_API drwav_bool32 drwav_guid_equal(const drwav_uint8 a[16], const drwav_uint8 b[16])
+{
+    return drwav__guid_equal(a, b);
+}
+
+DRWAV_API drwav_bool32 drwav_fourcc_equal(const drwav_uint8* a, const char* b)
+{
+    return drwav__fourcc_equal(a, b);
+}
+
+#endif  /* dr_wav_c */
+#endif  /* DR_WAV_IMPLEMENTATION */
+
+/*
+RELEASE NOTES - v0.11.0
+=======================
+Version 0.11.0 has breaking API changes.
+
+Improved Client-Defined Memory Allocation
+-----------------------------------------
+The main change with this release is the addition of a more flexible way of implementing custom memory allocation routines. The
+existing system of DRWAV_MALLOC, DRWAV_REALLOC and DRWAV_FREE are still in place and will be used by default when no custom
+allocation callbacks are specified.
+
+To use the new system, you pass in a pointer to a drwav_allocation_callbacks object to drwav_init() and family, like this:
+
+    void* my_malloc(size_t sz, void* pUserData)
+    {
+        return malloc(sz);
+    }
+    void* my_realloc(void* p, size_t sz, void* pUserData)
+    {
+        return realloc(p, sz);
+    }
+    void my_free(void* p, void* pUserData)
+    {
+        free(p);
+    }
+
+    ...
+
+    drwav_allocation_callbacks allocationCallbacks;
+    allocationCallbacks.pUserData = &myData;
+    allocationCallbacks.onMalloc  = my_malloc;
+    allocationCallbacks.onRealloc = my_realloc;
+    allocationCallbacks.onFree    = my_free;
+    drwav_init_file(&wav, "my_file.wav", &allocationCallbacks);
+
+The advantage of this new system is that it allows you to specify user data which will be passed in to the allocation routines.
+
+Passing in null for the allocation callbacks object will cause dr_wav to use defaults which is the same as DRWAV_MALLOC,
+DRWAV_REALLOC and DRWAV_FREE and the equivalent of how it worked in previous versions.
+
+Every API that opens a drwav object now takes this extra parameter. These include the following:
+
+    drwav_init()
+    drwav_init_ex()
+    drwav_init_file()
+    drwav_init_file_ex()
+    drwav_init_file_w()
+    drwav_init_file_w_ex()
+    drwav_init_memory()
+    drwav_init_memory_ex()
+    drwav_init_write()
+    drwav_init_write_sequential()
+    drwav_init_write_sequential_pcm_frames()
+    drwav_init_file_write()
+    drwav_init_file_write_sequential()
+    drwav_init_file_write_sequential_pcm_frames()
+    drwav_init_file_write_w()
+    drwav_init_file_write_sequential_w()
+    drwav_init_file_write_sequential_pcm_frames_w()
+    drwav_init_memory_write()
+    drwav_init_memory_write_sequential()
+    drwav_init_memory_write_sequential_pcm_frames()
+    drwav_open_and_read_pcm_frames_s16()
+    drwav_open_and_read_pcm_frames_f32()
+    drwav_open_and_read_pcm_frames_s32()
+    drwav_open_file_and_read_pcm_frames_s16()
+    drwav_open_file_and_read_pcm_frames_f32()
+    drwav_open_file_and_read_pcm_frames_s32()
+    drwav_open_file_and_read_pcm_frames_s16_w()
+    drwav_open_file_and_read_pcm_frames_f32_w()
+    drwav_open_file_and_read_pcm_frames_s32_w()
+    drwav_open_memory_and_read_pcm_frames_s16()
+    drwav_open_memory_and_read_pcm_frames_f32()
+    drwav_open_memory_and_read_pcm_frames_s32()
+
+Endian Improvements
+-------------------
+Previously, the following APIs returned little-endian audio data. These now return native-endian data. This improves compatibility
+on big-endian architectures.
+
+    drwav_read_pcm_frames()
+    drwav_read_pcm_frames_s16()
+    drwav_read_pcm_frames_s32()
+    drwav_read_pcm_frames_f32()
+    drwav_open_and_read_pcm_frames_s16()
+    drwav_open_and_read_pcm_frames_s32()
+    drwav_open_and_read_pcm_frames_f32()
+    drwav_open_file_and_read_pcm_frames_s16()
+    drwav_open_file_and_read_pcm_frames_s32()
+    drwav_open_file_and_read_pcm_frames_f32()
+    drwav_open_file_and_read_pcm_frames_s16_w()
+    drwav_open_file_and_read_pcm_frames_s32_w()
+    drwav_open_file_and_read_pcm_frames_f32_w()
+    drwav_open_memory_and_read_pcm_frames_s16()
+    drwav_open_memory_and_read_pcm_frames_s32()
+    drwav_open_memory_and_read_pcm_frames_f32()
+
+APIs have been added to give you explicit control over whether or not audio data is read or written in big- or little-endian byte
+order:
+
+    drwav_read_pcm_frames_le()
+    drwav_read_pcm_frames_be()
+    drwav_read_pcm_frames_s16le()
+    drwav_read_pcm_frames_s16be()
+    drwav_read_pcm_frames_f32le()
+    drwav_read_pcm_frames_f32be()
+    drwav_read_pcm_frames_s32le()
+    drwav_read_pcm_frames_s32be()
+    drwav_write_pcm_frames_le()
+    drwav_write_pcm_frames_be()
+
+Removed APIs
+------------
+The following APIs were deprecated in version 0.10.0 and have now been removed:
+
+    drwav_open()
+    drwav_open_ex()
+    drwav_open_write()
+    drwav_open_write_sequential()
+    drwav_open_file()
+    drwav_open_file_ex()
+    drwav_open_file_write()
+    drwav_open_file_write_sequential()
+    drwav_open_memory()
+    drwav_open_memory_ex()
+    drwav_open_memory_write()
+    drwav_open_memory_write_sequential()
+    drwav_close()
+
+
+
+RELEASE NOTES - v0.10.0
+=======================
+Version 0.10.0 has breaking API changes. There are no significant bug fixes in this release, so if you are affected you do
+not need to upgrade.
+
+Removed APIs
+------------
+The following APIs were deprecated in version 0.9.0 and have been completely removed in version 0.10.0:
+
+    drwav_read()
+    drwav_read_s16()
+    drwav_read_f32()
+    drwav_read_s32()
+    drwav_seek_to_sample()
+    drwav_write()
+    drwav_open_and_read_s16()
+    drwav_open_and_read_f32()
+    drwav_open_and_read_s32()
+    drwav_open_file_and_read_s16()
+    drwav_open_file_and_read_f32()
+    drwav_open_file_and_read_s32()
+    drwav_open_memory_and_read_s16()
+    drwav_open_memory_and_read_f32()
+    drwav_open_memory_and_read_s32()
+    drwav::totalSampleCount
+
+See release notes for version 0.9.0 at the bottom of this file for replacement APIs.
+
+Deprecated APIs
+---------------
+The following APIs have been deprecated. There is a confusing and completely arbitrary difference between drwav_init*() and
+drwav_open*(), where drwav_init*() initializes a pre-allocated drwav object, whereas drwav_open*() will first allocated a
+drwav object on the heap and then initialize it. drwav_open*() has been deprecated which means you must now use a pre-
+allocated drwav object with drwav_init*(). If you need the previous functionality, you can just do a malloc() followed by
+a called to one of the drwav_init*() APIs.
+
+    drwav_open()
+    drwav_open_ex()
+    drwav_open_write()
+    drwav_open_write_sequential()
+    drwav_open_file()
+    drwav_open_file_ex()
+    drwav_open_file_write()
+    drwav_open_file_write_sequential()
+    drwav_open_memory()
+    drwav_open_memory_ex()
+    drwav_open_memory_write()
+    drwav_open_memory_write_sequential()
+    drwav_close()
+
+These APIs will be removed completely in a future version. The rationale for this change is to remove confusion between the
+two different ways to initialize a drwav object.
+*/
+
+/*
+REVISION HISTORY
+================
+v0.12.16 - 2020-12-02
+  - Fix a bug when trying to read more bytes than can fit in a size_t.
+
+v0.12.15 - 2020-11-21
+  - Fix compilation with OpenWatcom.
+
+v0.12.14 - 2020-11-13
+  - Minor code clean up.
+
+v0.12.13 - 2020-11-01
+  - Improve compiler support for older versions of GCC.
+
+v0.12.12 - 2020-09-28
+  - Add support for RF64.
+  - Fix a bug in writing mode where the size of the RIFF chunk incorrectly includes the header section.
+
+v0.12.11 - 2020-09-08
+  - Fix a compilation error on older compilers.
+
+v0.12.10 - 2020-08-24
+  - Fix a bug when seeking with ADPCM formats.
+
+v0.12.9 - 2020-08-02
+  - Simplify sized types.
+
+v0.12.8 - 2020-07-25
+  - Fix a compilation warning.
+
+v0.12.7 - 2020-07-15
+  - Fix some bugs on big-endian architectures.
+  - Fix an error in s24 to f32 conversion.
+
+v0.12.6 - 2020-06-23
+  - Change drwav_read_*() to allow NULL to be passed in as the output buffer which is equivalent to a forward seek.
+  - Fix a buffer overflow when trying to decode invalid IMA-ADPCM files.
+  - Add include guard for the implementation section.
+
+v0.12.5 - 2020-05-27
+  - Minor documentation fix.
+
+v0.12.4 - 2020-05-16
+  - Replace assert() with DRWAV_ASSERT().
+  - Add compile-time and run-time version querying.
+    - DRWAV_VERSION_MINOR
+    - DRWAV_VERSION_MAJOR
+    - DRWAV_VERSION_REVISION
+    - DRWAV_VERSION_STRING
+    - drwav_version()
+    - drwav_version_string()
+
+v0.12.3 - 2020-04-30
+  - Fix compilation errors with VC6.
+
+v0.12.2 - 2020-04-21
+  - Fix a bug where drwav_init_file() does not close the file handle after attempting to load an erroneous file.
+
+v0.12.1 - 2020-04-13
+  - Fix some pedantic warnings.
+
+v0.12.0 - 2020-04-04
+  - API CHANGE: Add container and format parameters to the chunk callback.
+  - Minor documentation updates.
+
+v0.11.5 - 2020-03-07
+  - Fix compilation error with Visual Studio .NET 2003.
+
+v0.11.4 - 2020-01-29
+  - Fix some static analysis warnings.
+  - Fix a bug when reading f32 samples from an A-law encoded stream.
+
+v0.11.3 - 2020-01-12
+  - Minor changes to some f32 format conversion routines.
+  - Minor bug fix for ADPCM conversion when end of file is reached.
+
+v0.11.2 - 2019-12-02
+  - Fix a possible crash when using custom memory allocators without a custom realloc() implementation.
+  - Fix an integer overflow bug.
+  - Fix a null pointer dereference bug.
+  - Add limits to sample rate, channels and bits per sample to tighten up some validation.
+
+v0.11.1 - 2019-10-07
+  - Internal code clean up.
+
+v0.11.0 - 2019-10-06
+  - API CHANGE: Add support for user defined memory allocation routines. This system allows the program to specify their own memory allocation
+    routines with a user data pointer for client-specific contextual data. This adds an extra parameter to the end of the following APIs:
+    - drwav_init()
+    - drwav_init_ex()
+    - drwav_init_file()
+    - drwav_init_file_ex()
+    - drwav_init_file_w()
+    - drwav_init_file_w_ex()
+    - drwav_init_memory()
+    - drwav_init_memory_ex()
+    - drwav_init_write()
+    - drwav_init_write_sequential()
+    - drwav_init_write_sequential_pcm_frames()
+    - drwav_init_file_write()
+    - drwav_init_file_write_sequential()
+    - drwav_init_file_write_sequential_pcm_frames()
+    - drwav_init_file_write_w()
+    - drwav_init_file_write_sequential_w()
+    - drwav_init_file_write_sequential_pcm_frames_w()
+    - drwav_init_memory_write()
+    - drwav_init_memory_write_sequential()
+    - drwav_init_memory_write_sequential_pcm_frames()
+    - drwav_open_and_read_pcm_frames_s16()
+    - drwav_open_and_read_pcm_frames_f32()
+    - drwav_open_and_read_pcm_frames_s32()
+    - drwav_open_file_and_read_pcm_frames_s16()
+    - drwav_open_file_and_read_pcm_frames_f32()
+    - drwav_open_file_and_read_pcm_frames_s32()
+    - drwav_open_file_and_read_pcm_frames_s16_w()
+    - drwav_open_file_and_read_pcm_frames_f32_w()
+    - drwav_open_file_and_read_pcm_frames_s32_w()
+    - drwav_open_memory_and_read_pcm_frames_s16()
+    - drwav_open_memory_and_read_pcm_frames_f32()
+    - drwav_open_memory_and_read_pcm_frames_s32()
+    Set this extra parameter to NULL to use defaults which is the same as the previous behaviour. Setting this NULL will use
+    DRWAV_MALLOC, DRWAV_REALLOC and DRWAV_FREE.
+  - Add support for reading and writing PCM frames in an explicit endianness. New APIs:
+    - drwav_read_pcm_frames_le()
+    - drwav_read_pcm_frames_be()
+    - drwav_read_pcm_frames_s16le()
+    - drwav_read_pcm_frames_s16be()
+    - drwav_read_pcm_frames_f32le()
+    - drwav_read_pcm_frames_f32be()
+    - drwav_read_pcm_frames_s32le()
+    - drwav_read_pcm_frames_s32be()
+    - drwav_write_pcm_frames_le()
+    - drwav_write_pcm_frames_be()
+  - Remove deprecated APIs.
+  - API CHANGE: The following APIs now return native-endian data. Previously they returned little-endian data.
+    - drwav_read_pcm_frames()
+    - drwav_read_pcm_frames_s16()
+    - drwav_read_pcm_frames_s32()
+    - drwav_read_pcm_frames_f32()
+    - drwav_open_and_read_pcm_frames_s16()
+    - drwav_open_and_read_pcm_frames_s32()
+    - drwav_open_and_read_pcm_frames_f32()
+    - drwav_open_file_and_read_pcm_frames_s16()
+    - drwav_open_file_and_read_pcm_frames_s32()
+    - drwav_open_file_and_read_pcm_frames_f32()
+    - drwav_open_file_and_read_pcm_frames_s16_w()
+    - drwav_open_file_and_read_pcm_frames_s32_w()
+    - drwav_open_file_and_read_pcm_frames_f32_w()
+    - drwav_open_memory_and_read_pcm_frames_s16()
+    - drwav_open_memory_and_read_pcm_frames_s32()
+    - drwav_open_memory_and_read_pcm_frames_f32()
+
+v0.10.1 - 2019-08-31
+  - Correctly handle partial trailing ADPCM blocks.
+
+v0.10.0 - 2019-08-04
+  - Remove deprecated APIs.
+  - Add wchar_t variants for file loading APIs:
+      drwav_init_file_w()
+      drwav_init_file_ex_w()
+      drwav_init_file_write_w()
+      drwav_init_file_write_sequential_w()
+  - Add drwav_target_write_size_bytes() which calculates the total size in bytes of a WAV file given a format and sample count.
+  - Add APIs for specifying the PCM frame count instead of the sample count when opening in sequential write mode:
+      drwav_init_write_sequential_pcm_frames()
+      drwav_init_file_write_sequential_pcm_frames()
+      drwav_init_file_write_sequential_pcm_frames_w()
+      drwav_init_memory_write_sequential_pcm_frames()
+  - Deprecate drwav_open*() and drwav_close():
+      drwav_open()
+      drwav_open_ex()
+      drwav_open_write()
+      drwav_open_write_sequential()
+      drwav_open_file()
+      drwav_open_file_ex()
+      drwav_open_file_write()
+      drwav_open_file_write_sequential()
+      drwav_open_memory()
+      drwav_open_memory_ex()
+      drwav_open_memory_write()
+      drwav_open_memory_write_sequential()
+      drwav_close()
+  - Minor documentation updates.
+
+v0.9.2 - 2019-05-21
+  - Fix warnings.
+
+v0.9.1 - 2019-05-05
+  - Add support for C89.
+  - Change license to choice of public domain or MIT-0.
+
+v0.9.0 - 2018-12-16
+  - API CHANGE: Add new reading APIs for reading by PCM frames instead of samples. Old APIs have been deprecated and
+    will be removed in v0.10.0. Deprecated APIs and their replacements:
+      drwav_read()                     -> drwav_read_pcm_frames()
+      drwav_read_s16()                 -> drwav_read_pcm_frames_s16()
+      drwav_read_f32()                 -> drwav_read_pcm_frames_f32()
+      drwav_read_s32()                 -> drwav_read_pcm_frames_s32()
+      drwav_seek_to_sample()           -> drwav_seek_to_pcm_frame()
+      drwav_write()                    -> drwav_write_pcm_frames()
+      drwav_open_and_read_s16()        -> drwav_open_and_read_pcm_frames_s16()
+      drwav_open_and_read_f32()        -> drwav_open_and_read_pcm_frames_f32()
+      drwav_open_and_read_s32()        -> drwav_open_and_read_pcm_frames_s32()
+      drwav_open_file_and_read_s16()   -> drwav_open_file_and_read_pcm_frames_s16()
+      drwav_open_file_and_read_f32()   -> drwav_open_file_and_read_pcm_frames_f32()
+      drwav_open_file_and_read_s32()   -> drwav_open_file_and_read_pcm_frames_s32()
+      drwav_open_memory_and_read_s16() -> drwav_open_memory_and_read_pcm_frames_s16()
+      drwav_open_memory_and_read_f32() -> drwav_open_memory_and_read_pcm_frames_f32()
+      drwav_open_memory_and_read_s32() -> drwav_open_memory_and_read_pcm_frames_s32()
+      drwav::totalSampleCount          -> drwav::totalPCMFrameCount
+  - API CHANGE: Rename drwav_open_and_read_file_*() to drwav_open_file_and_read_*().
+  - API CHANGE: Rename drwav_open_and_read_memory_*() to drwav_open_memory_and_read_*().
+  - Add built-in support for smpl chunks.
+  - Add support for firing a callback for each chunk in the file at initialization time.
+    - This is enabled through the drwav_init_ex(), etc. family of APIs.
+  - Handle invalid FMT chunks more robustly.
+
+v0.8.5 - 2018-09-11
+  - Const correctness.
+  - Fix a potential stack overflow.
+
+v0.8.4 - 2018-08-07
+  - Improve 64-bit detection.
+
+v0.8.3 - 2018-08-05
+  - Fix C++ build on older versions of GCC.
+
+v0.8.2 - 2018-08-02
+  - Fix some big-endian bugs.
+
+v0.8.1 - 2018-06-29
+  - Add support for sequential writing APIs.
+  - Disable seeking in write mode.
+  - Fix bugs with Wave64.
+  - Fix typos.
+
+v0.8 - 2018-04-27
+  - Bug fix.
+  - Start using major.minor.revision versioning.
+
+v0.7f - 2018-02-05
+  - Restrict ADPCM formats to a maximum of 2 channels.
+
+v0.7e - 2018-02-02
+  - Fix a crash.
+
+v0.7d - 2018-02-01
+  - Fix a crash.
+
+v0.7c - 2018-02-01
+  - Set drwav.bytesPerSample to 0 for all compressed formats.
+  - Fix a crash when reading 16-bit floating point WAV files. In this case dr_wav will output silence for
+    all format conversion reading APIs (*_s16, *_s32, *_f32 APIs).
+  - Fix some divide-by-zero errors.
+
+v0.7b - 2018-01-22
+  - Fix errors with seeking of compressed formats.
+  - Fix compilation error when DR_WAV_NO_CONVERSION_API
+
+v0.7a - 2017-11-17
+  - Fix some GCC warnings.
+
+v0.7 - 2017-11-04
+  - Add writing APIs.
+
+v0.6 - 2017-08-16
+  - API CHANGE: Rename dr_* types to drwav_*.
+  - Add support for custom implementations of malloc(), realloc(), etc.
+  - Add support for Microsoft ADPCM.
+  - Add support for IMA ADPCM (DVI, format code 0x11).
+  - Optimizations to drwav_read_s16().
+  - Bug fixes.
+
+v0.5g - 2017-07-16
+  - Change underlying type for booleans to unsigned.
+
+v0.5f - 2017-04-04
+  - Fix a minor bug with drwav_open_and_read_s16() and family.
+
+v0.5e - 2016-12-29
+  - Added support for reading samples as signed 16-bit integers. Use the _s16() family of APIs for this.
+  - Minor fixes to documentation.
+
+v0.5d - 2016-12-28
+  - Use drwav_int* and drwav_uint* sized types to improve compiler support.
+
+v0.5c - 2016-11-11
+  - Properly handle JUNK chunks that come before the FMT chunk.
+
+v0.5b - 2016-10-23
+  - A minor change to drwav_bool8 and drwav_bool32 types.
+
+v0.5a - 2016-10-11
+  - Fixed a bug with drwav_open_and_read() and family due to incorrect argument ordering.
+  - Improve A-law and mu-law efficiency.
+
+v0.5 - 2016-09-29
+  - API CHANGE. Swap the order of "channels" and "sampleRate" parameters in drwav_open_and_read*(). Rationale for this is to
+    keep it consistent with dr_audio and dr_flac.
+
+v0.4b - 2016-09-18
+  - Fixed a typo in documentation.
+
+v0.4a - 2016-09-18
+  - Fixed a typo.
+  - Change date format to ISO 8601 (YYYY-MM-DD)
+
+v0.4 - 2016-07-13
+  - API CHANGE. Make onSeek consistent with dr_flac.
+  - API CHANGE. Rename drwav_seek() to drwav_seek_to_sample() for clarity and consistency with dr_flac.
+  - Added support for Sony Wave64.
+
+v0.3a - 2016-05-28
+  - API CHANGE. Return drwav_bool32 instead of int in onSeek callback.
+  - Fixed a memory leak.
+
+v0.3 - 2016-05-22
+  - Lots of API changes for consistency.
+
+v0.2a - 2016-05-16
+  - Fixed Linux/GCC build.
+
+v0.2 - 2016-05-11
+  - Added support for reading data as signed 32-bit PCM for consistency with dr_flac.
+
+v0.1a - 2016-05-07
+  - Fixed a bug in drwav_open_file() where the file handle would not be closed if the loader failed to initialize.
+
+v0.1 - 2016-05-04
+  - Initial versioned release.
+*/
+
+/*
+This software is available as a choice of the following licenses. Choose
+whichever you prefer.
+
+===============================================================================
+ALTERNATIVE 1 - Public Domain (www.unlicense.org)
+===============================================================================
+This is free and unencumbered software released into the public domain.
+
+Anyone is free to copy, modify, publish, use, compile, sell, or distribute this
+software, either in source code form or as a compiled binary, for any purpose,
+commercial or non-commercial, and by any means.
+
+In jurisdictions that recognize copyright laws, the author or authors of this
+software dedicate any and all copyright interest in the software to the public
+domain. We make this dedication for the benefit of the public at large and to
+the detriment of our heirs and successors. We intend this dedication to be an
+overt act of relinquishment in perpetuity of all present and future rights to
+this software under copyright law.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
+WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+
+For more information, please refer to <http://unlicense.org/>
+
+===============================================================================
+ALTERNATIVE 2 - MIT No Attribution
+===============================================================================
+Copyright 2020 David Reid
+
+Permission is hereby granted, free of charge, to any person obtaining a copy of
+this software and associated documentation files (the "Software"), to deal in
+the Software without restriction, including without limitation the rights to
+use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies
+of the Software, and to permit persons to whom the Software is furnished to do
+so.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
+*/

ggml/examples/gpt-2/CMakeLists.txt

@@ -0,0 +1,13 @@
+#
+# gpt-2
+
+set(TEST_TARGET gpt-2)
+add_executable(${TEST_TARGET} main.cpp)
+target_link_libraries(${TEST_TARGET} PRIVATE ggml common common-ggml)
+
+#
+# gpt-2-quantize
+
+set(TEST_TARGET gpt-2-quantize)
+add_executable(${TEST_TARGET} quantize.cpp)
+target_link_libraries(${TEST_TARGET} PRIVATE ggml common common-ggml)

ggml/examples/gpt-2/README.md

@@ -0,0 +1,158 @@
+# gpt-2
+
+This is a C++ example running GPT-2 inference using the [ggml](https://github.com/ggerganov/ggml) library.
+
+The program runs on the CPU - no video card is required.
+
+The [Cerebras-GPT](https://huggingface.co/cerebras) models are also supported.
+
+The example supports the following GPT-2 models:
+
+| Model | Description  | Disk Size |
+| ---   | ---          | ---       |
+| 117M  | Small model  | 240 MB    |
+| 345M  | Medium model | 680 MB    |
+| 774M  | Large model  | 1.5 GB    |
+| 1558M | XL model     | 3.0 GB    |
+
+Sample performance on MacBook M1 Pro:
+
+| Model | Size  | Time / Token |
+| ---   | ---   | ---    |
+| GPT-2 |  117M |   5 ms |
+| GPT-2 |  345M |  12 ms |
+| GPT-2 |  774M |  23 ms |
+| GPT-2 | 1558M |  42 ms |
+
+*TODO: add tables for Cerebras-GPT models*
+
+Sample output:
+
+```
+$ ./bin/gpt-2 -h
+usage: ./bin/gpt-2 [options]
+
+options:
+  -h, --help            show this help message and exit
+  -s SEED, --seed SEED  RNG seed (default: -1)
+  -t N, --threads N     number of threads to use during computation (default: 8)
+  -p PROMPT, --prompt PROMPT
+                        prompt to start generation with (default: random)
+  -n N, --n_predict N   number of tokens to predict (default: 200)
+  --top_k N             top-k sampling (default: 40)
+  --top_p N             top-p sampling (default: 0.9)
+  --temp N              temperature (default: 1.0)
+  -b N, --batch_size N  batch size for prompt processing (default: 8)
+  -m FNAME, --model FNAME
+                        model path (default: models/gpt-2-117M/ggml-model.bin)
+
+$ ./bin/gpt-2
+gpt2_model_load: loading model from 'models/gpt-2-117M/ggml-model.bin'
+gpt2_model_load: n_vocab = 50257
+gpt2_model_load: n_ctx   = 1024
+gpt2_model_load: n_embd  = 768
+gpt2_model_load: n_head  = 12
+gpt2_model_load: n_layer = 12
+gpt2_model_load: f16     = 1
+gpt2_model_load: ggml ctx size = 311.12 MB
+gpt2_model_load: memory size =    72.00 MB, n_mem = 12288
+gpt2_model_load: model size  =   239.08 MB
+main: number of tokens in prompt = 1
+
+So this is going to be the end of the line for us.
+
+If the Dolphins continue to do their business, it's possible that the team could make a bid to bring in new defensive coordinator Scott Linehan.
+
+Linehan's job is a little daunting, but he's a great coach and an excellent coach. I don't believe we're going to make the playoffs.
+
+We're going to have to work hard to keep our heads down and get ready to go.<|endoftext|>
+
+main: mem per token =  2048612 bytes
+main:     load time =   106.32 ms
+main:   sample time =     7.10 ms
+main:  predict time =   506.40 ms / 5.06 ms per token
+main:    total time =   629.84 ms
+```
+
+## Downloading and converting the original models (GPT-2)
+
+You can download the original model files using the [download-model.sh](download-model.sh) Bash script. The models are
+in Tensorflow format, so in order to use them with ggml, you need to convert them to appropriate format. This is done
+via the [convert-ckpt-to-ggml.py](convert-ckpt-to-ggml.py) python script.
+
+Here is the entire process for the GPT-2 117M model (download from official site + conversion):
+
+```
+cd ggml/build
+../examples/gpt-2/download-model.sh 117M
+
+Downloading model 117M ...
+models/gpt-2-117M/checkpoint                      100%[=============================>]      77  --.-KB/s    in 0s
+models/gpt-2-117M/encoder.json                    100%[=============================>]   1018K  1.20MB/s    in 0.8s
+models/gpt-2-117M/hparams.json                    100%[=============================>]      90  --.-KB/s    in 0s
+models/gpt-2-117M/model.ckpt.data-00000-of-00001  100%[=============================>] 474.70M  1.21MB/s    in 8m 39s
+models/gpt-2-117M/model.ckpt.index                100%[=============================>]   5.09K  --.-KB/s    in 0s
+models/gpt-2-117M/model.ckpt.meta                 100%[=============================>] 460.11K   806KB/s    in 0.6s
+models/gpt-2-117M/vocab.bpe                       100%[=============================>] 445.62K   799KB/s    in 0.6s
+Done! Model '117M' saved in 'models/gpt-2-117M/'
+
+Run the convert-ckpt-to-ggml.py script to convert the model to ggml format.
+
+  python /Users/john/ggml/examples/gpt-2/convert-ckpt-to-ggml.py models/gpt-2-117M/ 1
+
+```
+
+This conversion requires that you have python and Tensorflow installed on your computer. Still, if you want to avoid
+this, you can download the already converted ggml models as described below.
+
+## Downloading and converting the original models (Cerebras-GPT)
+
+Clone the respective repository from here: https://huggingface.co/cerebras
+
+Use the [convert-cerebras-to-ggml.py](convert-cerebras-to-ggml.py) script to convert the model to `ggml` format:
+
+```
+cd ggml/build
+git clone https://huggingface.co/cerebras/Cerebras-GPT-111M models/
+python ../examples/gpt-2/convert-cerebras-to-ggml.py models/Cerebras-GPT-111M/
+
+```
+
+## Downloading the ggml model directly (GPT-2)
+
+For convenience, I will be hosting the converted ggml model files in order to make it easier to run the examples. This
+way, you can directly download a single binary file and start using it. No python or Tensorflow is required.
+
+Here is how to get the 117M ggml model:
+
+```
+cd ggml/build
+../examples/gpt-2/download-ggml-model.sh 117M
+
+Downloading ggml model 117M ...
+models/gpt-2-117M/ggml-model.bin         100%[===============================>] 239.58M  8.52MB/s    in 28s
+Done! Model '117M' saved in 'models/gpt-2-117M/ggml-model.bin'
+You can now use it like this:
+
+  $ ./bin/gpt-2 -m models/gpt-2-117M/ggml-model.bin -p "This is an example"
+
+```
+
+At some point, I might decide to stop hosting these models. So in that case, simply revert to the manual process above.
+
+## Quantizing the models
+
+You can also try to quantize the `ggml` models via 4-bit integer quantization.
+Keep in mind that for smaller models, this will render them completely useless.
+You generally want to quantize larger models.
+
+```
+# quantize GPT-2 F16 to Q4_0 (faster but less precise)
+./bin/gpt-2-quantize models/gpt-2-1558M/ggml-model-f16.bin models/gpt-2-1558M/ggml-model-q4_0.bin 2
+./bin/gpt-2 -m models/gpt-2-1558M/ggml-model-q4_0.bin -p "This is an example"
+
+# quantize Cerebras F16 to Q4_1 (slower but more precise)
+./bin/gpt-2-quantize models/Cerebras-GPT-6.7B/ggml-model-f16.bin models/Cerebras-GPT-6.7B/ggml-model-q4_1.bin 3
+./bin/gpt-2 -m models/Cerebras-GPT-6.7B/ggml-model-q4_1.bin -p "This is an example"
+
+```

ggml/examples/gpt-2/convert-cerebras-to-ggml.py

@@ -0,0 +1,183 @@
+# Convert Cerebras models to ggml format
+#
+# ref: https://www.cerebras.net/blog/cerebras-gpt-a-family-of-open-compute-efficient-large-language-models/
+#
+
+import sys
+import struct
+import json
+import torch
+import numpy as np
+import re
+
+from transformers import AutoModelForCausalLM
+
+# ref: https://github.com/openai/gpt-2/blob/master/src/encoder.py
+def bytes_to_unicode():
+    """
+    Returns list of utf-8 byte and a corresponding list of unicode strings.
+    The reversible bpe codes work on unicode strings.
+    This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.
+    When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.
+    This is a signficant percentage of your normal, say, 32K bpe vocab.
+    To avoid that, we want lookup tables between utf-8 bytes and unicode strings.
+    And avoids mapping to whitespace/control characters the bpe code barfs on.
+    """
+    bs = list(range(ord("!"), ord("~")+1))+list(range(ord("¡"), ord("¬")+1))+list(range(ord("®"), ord("ÿ")+1))
+    cs = bs[:]
+    n = 0
+    for b in range(2**8):
+        if b not in bs:
+            bs.append(b)
+            cs.append(2**8+n)
+            n += 1
+    cs = [chr(n) for n in cs]
+    return dict(zip(bs, cs))
+
+if len(sys.argv) < 2:
+    print("Usage: convert-cerebras-to-ggml.py dir-model [use-f32]\n")
+    sys.exit(1)
+
+# output in the same directory as the model
+dir_model = sys.argv[1]
+fname_out = sys.argv[1] + "/ggml-model-f16.bin"
+
+with open(dir_model + "/vocab.json", "r", encoding="utf-8") as f:
+    encoder = json.load(f)
+
+with open(dir_model + "/config.json", "r", encoding="utf-8") as f:
+    hparams = json.load(f)
+
+# use 16-bit or 32-bit floats
+use_f16 = True
+if len(sys.argv) > 2:
+    use_f16 = False
+    fname_out = sys.argv[1] + "/ggml-model-f32.bin"
+
+model = AutoModelForCausalLM.from_pretrained(dir_model, low_cpu_mem_usage=True)
+#print (model)
+
+list_vars = model.state_dict()
+#print (list_vars)
+
+print(hparams)
+
+fout = open(fname_out, "wb")
+
+fout.write(struct.pack("i", 0x67676d6c)) # magic: ggml in hex
+fout.write(struct.pack("i", hparams["vocab_size"]))
+fout.write(struct.pack("i", hparams["n_positions"]))
+fout.write(struct.pack("i", hparams["n_embd"]))
+fout.write(struct.pack("i", hparams["n_head"]))
+fout.write(struct.pack("i", hparams["n_layer"]))
+fout.write(struct.pack("i", use_f16))
+
+byte_encoder = bytes_to_unicode()
+byte_decoder = {v:k for k, v in byte_encoder.items()}
+
+fout.write(struct.pack("i", len(encoder)))
+
+for key in encoder:
+    text = bytearray([byte_decoder[c] for c in key])
+    fout.write(struct.pack("i", len(text)))
+    fout.write(text)
+
+for name in list_vars.keys():
+    data = list_vars[name].squeeze().numpy()
+    print("Processing variable: " + name + " with shape: ", data.shape)
+
+    # rename headers to keep compatibility
+    if name == "transformer.ln_f.weight":
+        name = "model/ln_f/g"
+    elif name == "transformer.ln_f.bias":
+        name = "model/ln_f/b"
+    elif name == "transformer.wte.weight":
+        name = "model/wte"
+    elif name == "transformer.wpe.weight":
+        name = "model/wpe"
+    elif name == "lm_head.weight":
+        name = "model/lm_head"
+    elif re.match(r"transformer.h\.\d+\.ln_1\.weight", name):
+        i = re.findall("\d+", name)[0]
+        name = f"model/h{i}/ln_1/g"
+    elif re.match(r"transformer.h\.\d+\.ln_1\.bias", name):
+        i = re.findall("\d+", name)[0]
+        name = f"model/h{i}/ln_1/b"
+    elif re.match(r"transformer.h\.\d+\.attn\.c_attn\.weight", name):
+        i = re.findall("\d+", name)[0]
+        name = f"model/h{i}/attn/c_attn/w"
+    elif re.match(r"transformer.h\.\d+\.attn\.c_attn\.bias", name):
+        i = re.findall("\d+", name)[0]
+        name = f"model/h{i}/attn/c_attn/b"
+    elif re.match(r"transformer.h\.\d+\.attn\.c_proj\.weight", name):
+        i = re.findall("\d+", name)[0]
+        name = f"model/h{i}/attn/c_proj/w"
+    elif re.match(r"transformer.h.\d+.attn.c_proj.bias", name):
+        i = re.findall("\d+", name)[0]
+        name = f"model/h{i}/attn/c_proj/b"
+    elif re.match(r"transformer.h.\d+.ln_2.weight", name):
+        i = re.findall("\d+", name)[0]
+        name = f"model/h{i}/ln_2/g"
+    elif re.match(r"transformer.h.\d+.ln_2.bias", name):
+        i = re.findall("\d+", name)[0]
+        name = f"model/h{i}/ln_2/b"
+    elif re.match(r"transformer.h.\d+.mlp.c_fc.weight", name):
+        i = re.findall("\d+", name)[0]
+        name = f"model/h{i}/mlp/c_fc/w"
+    elif re.match(r"transformer.h.\d+.mlp.c_fc.bias", name):
+        i = re.findall("\d+", name)[0]
+        name = f"model/h{i}/mlp/c_fc/b"
+    elif re.match(r"transformer.h.\d+.mlp.c_proj.weight", name):
+        i = re.findall("\d+", name)[0]
+        name = f"model/h{i}/mlp/c_proj/w"
+    elif re.match(r"transformer.h.\d+.mlp.c_proj.bias", name):
+        i = re.findall("\d+", name)[0]
+        name = f"model/h{i}/mlp/c_proj/b"
+    else:
+        print("Unrecognized variable name. %s", name)
+
+    # we don't need these
+    if name.endswith("attn.masked_bias") or name.endswith(".attn.bias"):
+        print("  Skipping variable: " + name)
+        continue
+
+    n_dims = len(data.shape);
+
+    # ftype == 0 -> float32, ftype == 1 -> float16
+    ftype = 0;
+    if use_f16:
+        if (name == "model/wte" or name == "model/lm_head" or name[-2:] == "/g" or name[-2:] == "/w") and n_dims == 2:
+            print("  Converting to float16")
+            data = data.astype(np.float16)
+            ftype = 1
+        else:
+            print("  Converting to float32")
+            data = data.astype(np.float32)
+            ftype = 0
+
+    # for efficiency - transpose the projection matrices
+    # "model/h.*/attn/c_attn/w"
+    # "model/h.*/attn/c_proj/w"
+    # "model/h.*/mlp/c_fc/w"
+    # "model/h.*/mlp/c_proj/w"
+    if name[-14:] == "/attn/c_attn/w" or \
+       name[-14:] == "/attn/c_proj/w" or \
+       name[-11:] == "/mlp/c_fc/w" or \
+       name[-13:] == "/mlp/c_proj/w":
+        print("  Transposing")
+        data = data.transpose()
+
+    # header
+    str = name.encode('utf-8')
+    fout.write(struct.pack("iii", n_dims, len(str), ftype))
+    for i in range(n_dims):
+        fout.write(struct.pack("i", data.shape[n_dims - 1 - i]))
+    fout.write(str);
+
+    # data
+    data.tofile(fout)
+
+fout.close()
+
+print("Done. Output file: " + fname_out)
+print("")

ggml/examples/gpt-2/convert-ckpt-to-ggml.py

@@ -0,0 +1,159 @@
+# Convert a model checkpoint to a ggml compatible file
+#
+# Load the model using TensorFlow.
+# Iterate over all variables and write them to a binary file.
+#
+# For each variable, write the following:
+#   - Number of dimensions (int)
+#   - Name length (int)
+#   - Dimensions (int[n_dims])
+#   - Name (char[name_length])
+#   - Data (float[n_dims])
+#
+# By default, the bigger matrices are converted to 16-bit floats.
+# This can be disabled by adding the "use-f32" CLI argument.
+#
+# At the start of the ggml file we write the model parameters
+# and vocabulary.
+#
+
+import sys
+import json
+import struct
+import numpy as np
+import tensorflow as tf
+
+# ref: https://github.com/openai/gpt-2/blob/master/src/encoder.py
+def bytes_to_unicode():
+    """
+    Returns list of utf-8 byte and a corresponding list of unicode strings.
+    The reversible bpe codes work on unicode strings.
+    This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.
+    When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.
+    This is a signficant percentage of your normal, say, 32K bpe vocab.
+    To avoid that, we want lookup tables between utf-8 bytes and unicode strings.
+    And avoids mapping to whitespace/control characters the bpe code barfs on.
+    """
+    bs = list(range(ord("!"), ord("~")+1))+list(range(ord("¡"), ord("¬")+1))+list(range(ord("®"), ord("ÿ")+1))
+    cs = bs[:]
+    n = 0
+    for b in range(2**8):
+        if b not in bs:
+            bs.append(b)
+            cs.append(2**8+n)
+            n += 1
+    cs = [chr(n) for n in cs]
+    return dict(zip(bs, cs))
+
+# helper method to convert a numpy array to different float types
+def convert_to_ftype(data, ftype):
+    # fp16
+    if ftype == 1:
+        return data.astype(np.float16)
+
+    assert False, "Invalid ftype: " + str(ftype)
+
+if len(sys.argv) < 3:
+    print("Usage: convert-ckpt-to-ggml.py dir-model ftype\n")
+    print("  ftype == 0 -> float32")
+    print("  ftype == 1 -> float16")
+    sys.exit(1)
+
+# output in the same directory as the model
+dir_model = sys.argv[1]
+fname_out = sys.argv[1] + "/ggml-model.bin"
+
+with open(dir_model + "/encoder.json", "r", encoding="utf-8") as f:
+    encoder = json.load(f)
+
+with open(dir_model + "/hparams.json", "r", encoding="utf-8") as f:
+    hparams = json.load(f)
+
+# possible data types
+#   ftype == 0 -> float32
+#   ftype == 1 -> float16
+#
+# map from ftype to string
+ftype_str = ["f32", "f16"]
+
+ftype = 1
+if len(sys.argv) > 2:
+    ftype = int(sys.argv[2])
+    if ftype < 0 or ftype > 1:
+        print("Invalid ftype: " + str(ftype))
+        sys.exit(1)
+    fname_out = sys.argv[1] + "/ggml-model-" + ftype_str[ftype] + ".bin"
+
+list_vars = tf.train.list_variables(dir_model)
+
+fout = open(fname_out, "wb")
+
+fout.write(struct.pack("i", 0x67676d6c)) # magic: ggml in hex
+fout.write(struct.pack("i", hparams["n_vocab"]))
+fout.write(struct.pack("i", hparams["n_ctx"]))
+fout.write(struct.pack("i", hparams["n_embd"]))
+fout.write(struct.pack("i", hparams["n_head"]))
+fout.write(struct.pack("i", hparams["n_layer"]))
+fout.write(struct.pack("i", ftype))
+
+byte_encoder = bytes_to_unicode()
+byte_decoder = {v:k for k, v in byte_encoder.items()}
+
+fout.write(struct.pack("i", len(encoder)))
+
+for key in encoder:
+    text = bytearray([byte_decoder[c] for c in key])
+    fout.write(struct.pack("i", len(text)))
+    fout.write(text)
+
+for name, shape in list_vars:
+    print("Processing variable: " + name + " with shape: ", shape)
+
+    data = tf.train.load_variable(dir_model, name).squeeze()
+    n_dims = len(data.shape);
+
+    # for efficiency - transpose the projection matrices
+    # "model/h.*/attn/c_attn/w"
+    # "model/h.*/attn/c_proj/w"
+    # "model/h.*/mlp/c_fc/w"
+    # "model/h.*/mlp/c_proj/w"
+    if name[-14:] == "/attn/c_attn/w" or \
+       name[-14:] == "/attn/c_proj/w" or \
+       name[-11:] == "/mlp/c_fc/w" or \
+       name[-13:] == "/mlp/c_proj/w":
+        print("  Transposing")
+        data = data.transpose()
+
+    dshape = data.shape
+
+    ftype_cur = 0
+    if ftype != 0:
+        # match name:
+        #  "model/wte"
+        #  "model/h.*/attn/c_attn/w"
+        #  "model/h.*/attn/c_proj/w"
+        #  "model/h.*/mlp/c_fc/w"
+        #  "model/h.*/mlp/c_proj/w"
+        if name == "model/wte" or name[-2:] == "/w":
+            print("  Converting to " + ftype_str[ftype])
+            data = convert_to_ftype(data, ftype)
+            ftype_cur = ftype
+        else:
+            print("  Converting to float32")
+            data = data.astype(np.float32)
+            ftype_cur = 0
+
+    # header
+    str = name.encode('utf-8')
+    fout.write(struct.pack("iii", n_dims, len(str), ftype_cur))
+    for i in range(n_dims):
+        fout.write(struct.pack("i", dshape[n_dims - 1 - i]))
+    fout.write(str);
+
+    # data
+    data.tofile(fout)
+
+fout.close()
+
+print("Done. Output file: " + fname_out)
+print("")

ggml/examples/gpt-2/convert-h5-to-ggml.py

@@ -0,0 +1,195 @@
+# Convert GPT-2 h5 transformer model to ggml format
+#
+# Load the model using GPT2Model.
+# Iterate over all variables and write them to a binary file.
+#
+# For each variable, write the following:
+#   - Number of dimensions (int)
+#   - Name length (int)
+#   - Dimensions (int[n_dims])
+#   - Name (char[name_length])
+#   - Data (float[n_dims])
+#
+# By default, the bigger matrices are converted to 16-bit floats.
+# This can be disabled by adding the "use-f32" CLI argument.
+#
+# At the start of the ggml file we write the model parameters
+# and vocabulary.
+#
+
+import sys
+import struct
+import json
+import numpy as np
+import re
+
+from transformers import GPT2Model
+
+# ref: https://github.com/openai/gpt-2/blob/master/src/encoder.py
+def bytes_to_unicode():
+    """
+    Returns list of utf-8 byte and a corresponding list of unicode strings.
+    The reversible bpe codes work on unicode strings.
+    This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.
+    When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.
+    This is a signficant percentage of your normal, say, 32K bpe vocab.
+    To avoid that, we want lookup tables between utf-8 bytes and unicode strings.
+    And avoids mapping to whitespace/control characters the bpe code barfs on.
+    """
+    bs = list(range(ord("!"), ord("~")+1))+list(range(ord("¡"), ord("¬")+1))+list(range(ord("®"), ord("ÿ")+1))
+    cs = bs[:]
+    n = 0
+    for b in range(2**8):
+        if b not in bs:
+            bs.append(b)
+            cs.append(2**8+n)
+            n += 1
+    cs = [chr(n) for n in cs]
+    return dict(zip(bs, cs))
+
+if len(sys.argv) < 2:
+    print("Usage: convert-h5-to-ggml.py dir-model [use-f32]\n")
+    sys.exit(1)
+
+# output in the same directory as the model
+dir_model = sys.argv[1]
+fname_out = sys.argv[1] + "/ggml-model.bin"
+
+with open(dir_model + "/vocab.json", "r", encoding="utf-8") as f:
+    encoder = json.load(f)
+
+with open(dir_model + "/added_tokens.json", "r", encoding="utf-8") as f:
+    encoder_added = json.load(f)
+
+with open(dir_model + "/config.json", "r", encoding="utf-8") as f:
+    hparams = json.load(f)
+
+# use 16-bit or 32-bit floats
+use_f16 = True
+if len(sys.argv) > 2:
+    use_f16 = False
+    fname_out = sys.argv[1] + "/ggml-model-f32.bin"
+
+model = GPT2Model.from_pretrained(dir_model, low_cpu_mem_usage=True)
+#print (model)
+
+list_vars = model.state_dict()
+#print (list_vars)
+
+fout = open(fname_out, "wb")
+
+fout.write(struct.pack("i", 0x67676d6c)) # magic: ggml in hex
+fout.write(struct.pack("i", hparams["vocab_size"]))
+fout.write(struct.pack("i", hparams["n_positions"]))
+fout.write(struct.pack("i", hparams["n_embd"]))
+fout.write(struct.pack("i", hparams["n_head"]))
+fout.write(struct.pack("i", hparams["n_layer"]))
+#fout.write(struct.pack("i", hparams["rotary_dim"]))
+fout.write(struct.pack("i", use_f16))
+
+byte_encoder = bytes_to_unicode()
+byte_decoder = {v:k for k, v in byte_encoder.items()}
+
+fout.write(struct.pack("i", len(encoder) + len(encoder_added)))
+
+for key in encoder:
+    text = bytearray([byte_decoder[c] for c in key])
+    fout.write(struct.pack("i", len(text)))
+    fout.write(text)
+
+for key in encoder_added:
+    text = bytearray([byte_decoder[c] for c in key])
+    fout.write(struct.pack("i", len(text)))
+    fout.write(text)
+
+for name in list_vars.keys():
+    data = list_vars[name].squeeze().numpy()
+    print("Processing variable: " + name + " with shape: ", data.shape)
+
+    # we don't need these
+    if name.endswith("attn.masked_bias") or name.endswith(".attn.bias"):
+        print("  Skipping variable: " + name)
+        continue
+
+    n_dims = len(data.shape);
+
+    # ftype == 0 -> float32, ftype == 1 -> float16
+    ftype = 0;
+    if use_f16:
+        if name[-7:] == ".weight" and n_dims == 2:
+            print("  Converting to float16")
+            data = data.astype(np.float16)
+            ftype = 1
+        else:
+            print("  Converting to float32")
+            data = data.astype(np.float32)
+            ftype = 0
+
+    # for efficiency - transpose these matrices:
+    #  "transformer.h.*.mlp.c_proj.weight
+    if name.endswith(".mlp.c_proj.weight"):
+        print("  Transposing")
+        data = data.transpose()
+
+    # rename headers to keep compatibility
+    if name == "ln_f.weight":
+        name = "model/ln_f/g"
+    elif name == "ln_f.bias":
+        name = "model/ln_f/b"
+    elif name == "wte.weight":
+        name = "model/wte"
+    elif name == "wpe.weight":
+        name = "model/wpe"
+    elif re.match(r"h\.\d+\.ln_1\.weight", name):
+        i = re.findall("\d+", name)[0]
+        name = f"model/h{i}/ln_1/g"
+    elif re.match(r"h\.\d+\.ln_1\.bias", name):
+        i = re.findall("\d+", name)[0]
+        name = f"model/h{i}/ln_1/b"
+    elif re.match(r"h\.\d+\.attn\.c_attn\.weight", name):
+        i = re.findall("\d+", name)[0]
+        name = f"model/h{i}/attn/c_attn/w"
+    elif re.match(r"h\.\d+\.attn\.c_attn\.bias", name):
+        i = re.findall("\d+", name)[0]
+        name = f"model/h{i}/attn/c_attn/b"
+    elif re.match(r"h\.\d+\.attn\.c_proj\.weight", name):
+        i = re.findall("\d+", name)[0]
+        name = f"model/h{i}/attn/c_proj/w"
+    elif re.match(r"h.\d+.attn.c_proj.bias", name):
+        i = re.findall("\d+", name)[0]
+        name = f"model/h{i}/attn/c_proj/b"
+    elif re.match(r"h.\d+.ln_2.weight", name):
+        i = re.findall("\d+", name)[0]
+        name = f"model/h{i}/ln_2/g"
+    elif re.match(r"h.\d+.ln_2.bias", name):
+        i = re.findall("\d+", name)[0]
+        name = f"model/h{i}/ln_2/b"
+    elif re.match(r"h.\d+.mlp.c_fc.weight", name):
+        i = re.findall("\d+", name)[0]
+        name = f"model/h{i}/mlp/c_fc/w"
+    elif re.match(r"h.\d+.mlp.c_fc.bias", name):
+        i = re.findall("\d+", name)[0]
+        name = f"model/h{i}/mlp/c_fc/b"
+    elif re.match(r"h.\d+.mlp.c_proj.weight", name):
+        i = re.findall("\d+", name)[0]
+        name = f"model/h{i}/mlp/c_proj/w"
+    elif re.match(r"h.\d+.mlp.c_proj.bias", name):
+        i = re.findall("\d+", name)[0]
+        name = f"model/h{i}/mlp/c_proj/b"
+    else:
+        print("Unrecognized variable name. %s", name)
+
+    str = name.encode('utf-8')
+
+    fout.write(struct.pack("iii", n_dims, len(str), ftype))
+    for i in range(n_dims):
+        fout.write(struct.pack("i", data.shape[n_dims - 1 - i]))
+    fout.write(str);
+
+    # data
+    data.tofile(fout)
+
+fout.close()
+
+print("Done. Output file: " + fname_out)
+print("")

ggml/examples/gpt-2/download-ggml-model.sh

@@ -0,0 +1,69 @@
+#!/bin/bash
+
+# This script downloads GPT-2 model files that have already been converted to ggml format.
+# This way you don't have to convert them yourself.
+#
+# If you want to download the original GPT-2 model files, use the "download-model.sh" script instead.
+
+#src="https://ggml.ggerganov.com"
+#pfx="ggml-model-gpt-2"
+
+src="https://huggingface.co/ggerganov/ggml"
+pfx="resolve/main/ggml-model-gpt-2"
+
+ggml_path=$(dirname $(realpath $0))
+
+# GPT-2 models
+models=( "117M" "345M" "774M" "1558M" )
+
+# list available models
+function list_models {
+    printf "\n"
+    printf "  Available models:"
+    for model in "${models[@]}"; do
+        printf " $model"
+    done
+    printf "\n\n"
+}
+
+if [ "$#" -ne 1 ]; then
+    printf "Usage: $0 <model>\n"
+    list_models
+
+    exit 1
+fi
+
+model=$1
+
+if [[ ! " ${models[@]} " =~ " ${model} " ]]; then
+    printf "Invalid model: $model\n"
+    list_models
+
+    exit 1
+fi
+
+# download ggml model
+
+printf "Downloading ggml model $model ...\n"
+
+mkdir -p models/gpt-2-$model
+
+if [ -x "$(command -v wget)" ]; then
+    wget --quiet --show-progress -O models/gpt-2-$model/ggml-model.bin $src/$pfx-$model.bin
+elif [ -x "$(command -v curl)" ]; then
+    curl -L --output models/gpt-2-$model/ggml-model.bin $src/$pfx-$model.bin
+else
+    printf "Either wget or curl is required to download models.\n"
+    exit 1
+fi
+
+if [ $? -ne 0 ]; then
+    printf "Failed to download ggml model $model \n"
+    printf "Please try again later or download the original GPT-2 model files and convert them yourself.\n"
+    exit 1
+fi
+
+printf "Done! Model '$model' saved in 'models/gpt-2-$model/ggml-model.bin'\n"
+printf "You can now use it like this:\n\n"
+printf "  $ ./bin/gpt-2 -m models/gpt-2-$model/ggml-model.bin -p \"This is an example\"\n"
+printf "\n"

ggml/examples/gpt-2/download-model.sh

@@ -0,0 +1,48 @@
+#!/bin/bash
+
+ggml_path=$(dirname $(realpath $0))
+
+# GPT-2 models
+models=( "117M" "345M" "774M" "1558M" )
+
+# list available models
+function list_models {
+    printf "\n"
+    printf "  Available models:"
+    for model in "${models[@]}"; do
+        printf " $model"
+    done
+    printf "\n\n"
+}
+
+if [ "$#" -ne 1 ]; then
+    printf "Usage: $0 <model>\n"
+    list_models
+
+    exit 1
+fi
+
+model=$1
+
+if [[ ! " ${models[@]} " =~ " ${model} " ]]; then
+    printf "Invalid model: $model\n"
+    list_models
+
+    exit 1
+fi
+
+# download model
+
+printf "Downloading model $model ...\n"
+
+mkdir -p models/gpt-2-$model
+
+for file in checkpoint encoder.json hparams.json model.ckpt.data-00000-of-00001 model.ckpt.index model.ckpt.meta vocab.bpe; do
+    wget --quiet --show-progress -O models/gpt-2-$model/$file https://openaipublic.blob.core.windows.net/gpt-2/models/$model/$file
+done
+
+printf "Done! Model '$model' saved in 'models/gpt-2-$model/'\n\n"
+printf "Run the convert-ckpt-to-ggml.py script to convert the model to ggml format.\n"
+printf "\n"
+printf "  python $ggml_path/convert-ckpt-to-ggml.py models/gpt-2-$model/\n"
+printf "\n"

ggml/examples/gpt-2/main.cpp

@@ -0,0 +1,892 @@
+#include "ggml/ggml.h"
+#include "ggml/ggml-alloc.h"
+
+#include "common.h"
+#include "common-ggml.h"
+
+#include <cassert>
+#include <cmath>
+#include <cstdio>
+#include <cstring>
+#include <fstream>
+#include <map>
+#include <string>
+#include <vector>
+
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
+// default hparams (GPT-2 117M)
+struct gpt2_hparams {
+    int32_t n_vocab = 50257;
+    int32_t n_ctx   = 1024;
+    int32_t n_embd  = 768;
+    int32_t n_head  = 12;
+    int32_t n_layer = 12;
+    int32_t ftype   = 1;
+    float   eps     = 1e-5f;
+};
+
+struct gpt2_layer {
+    // normalization
+    struct ggml_tensor * ln_1_g;
+    struct ggml_tensor * ln_1_b;
+
+    struct ggml_tensor * ln_2_g;
+    struct ggml_tensor * ln_2_b;
+
+    // attention
+    struct ggml_tensor * c_attn_attn_w;
+    struct ggml_tensor * c_attn_attn_b;
+
+    struct ggml_tensor * c_attn_proj_w;
+    struct ggml_tensor * c_attn_proj_b;
+
+    // mlp
+    struct ggml_tensor * c_mlp_fc_w;
+    struct ggml_tensor * c_mlp_fc_b;
+
+    struct ggml_tensor * c_mlp_proj_w;
+    struct ggml_tensor * c_mlp_proj_b;
+};
+
+struct gpt2_model {
+    gpt2_hparams hparams;
+
+    // normalization
+    struct ggml_tensor * ln_f_g;
+    struct ggml_tensor * ln_f_b;
+
+    struct ggml_tensor * wte;     // position embedding
+    struct ggml_tensor * wpe;     //    token embedding
+    struct ggml_tensor * lm_head; // language model head
+
+    std::vector<gpt2_layer> layers;
+
+    // key + value memory
+    struct ggml_tensor * memory_k;
+    struct ggml_tensor * memory_v;
+
+    //
+    struct ggml_context * ctx;
+    std::map<std::string, struct ggml_tensor *> tensors;
+};
+
+// load the model's weights from a file
+bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab & vocab) {
+    printf("%s: loading model from '%s'\n", __func__, fname.c_str());
+
+    auto fin = std::ifstream(fname, std::ios::binary);
+    if (!fin) {
+        fprintf(stderr, "%s: failed to open '%s'\n", __func__, fname.c_str());
+        return false;
+    }
+
+    // verify magic
+    {
+        uint32_t magic;
+        fin.read((char *) &magic, sizeof(magic));
+        if (magic != GGML_FILE_MAGIC) {
+            fprintf(stderr, "%s: invalid model file '%s' (bad magic)\n", __func__, fname.c_str());
+            return false;
+        }
+    }
+
+    // load hparams
+    {
+        auto & hparams = model.hparams;
+
+        fin.read((char *) &hparams.n_vocab, sizeof(hparams.n_vocab));
+        fin.read((char *) &hparams.n_ctx,   sizeof(hparams.n_ctx));
+        fin.read((char *) &hparams.n_embd,  sizeof(hparams.n_embd));
+        fin.read((char *) &hparams.n_head,  sizeof(hparams.n_head));
+        fin.read((char *) &hparams.n_layer, sizeof(hparams.n_layer));
+        fin.read((char *) &hparams.ftype,   sizeof(hparams.ftype));
+
+        const int32_t qntvr = hparams.ftype / GGML_QNT_VERSION_FACTOR;
+
+        printf("%s: n_vocab = %d\n", __func__, hparams.n_vocab);
+        printf("%s: n_ctx   = %d\n", __func__, hparams.n_ctx);
+        printf("%s: n_embd  = %d\n", __func__, hparams.n_embd);
+        printf("%s: n_head  = %d\n", __func__, hparams.n_head);
+        printf("%s: n_layer = %d\n", __func__, hparams.n_layer);
+        printf("%s: ftype   = %d\n", __func__, hparams.ftype);
+        printf("%s: qntvr   = %d\n", __func__, qntvr);
+
+        hparams.ftype %= GGML_QNT_VERSION_FACTOR;
+    }
+
+    // load vocab
+    {
+        int32_t n_vocab = 0;
+        fin.read((char *) &n_vocab, sizeof(n_vocab));
+
+        if (n_vocab != model.hparams.n_vocab) {
+            fprintf(stderr, "%s: invalid model file '%s' (bad vocab size %d != %d)\n",
+                    __func__, fname.c_str(), n_vocab, model.hparams.n_vocab);
+            return false;
+        }
+
+        std::string word;
+        std::vector<char> buf(128);
+
+        for (int i = 0; i < n_vocab; i++) {
+            uint32_t len;
+            fin.read((char *) &len, sizeof(len));
+
+            buf.resize(len);
+            fin.read((char *) buf.data(), len);
+            word.assign(buf.data(), len);
+
+            vocab.token_to_id[word] = i;
+            vocab.id_to_token[i] = word;
+        }
+    }
+
+    // for the big tensors, we have the option to store the data in 16-bit floats or quantized
+    // in order to save memory and also to speed up the computation
+    ggml_type wtype = ggml_ftype_to_ggml_type((ggml_ftype) (model.hparams.ftype));
+    if (wtype == GGML_TYPE_COUNT) {
+        fprintf(stderr, "%s: invalid model file '%s' (bad ftype value %d)\n",
+                __func__, fname.c_str(), model.hparams.ftype);
+        return false;
+    }
+
+    auto & ctx = model.ctx;
+
+    size_t ctx_size = 0;
+
+    {
+        const auto & hparams = model.hparams;
+
+        const int n_embd  = hparams.n_embd;
+        const int n_layer = hparams.n_layer;
+        const int n_ctx   = hparams.n_ctx;
+        const int n_vocab = hparams.n_vocab;
+
+        ctx_size += n_embd*ggml_type_sizef(GGML_TYPE_F32); // ln_f_g
+        ctx_size += n_embd*ggml_type_sizef(GGML_TYPE_F32); // ln_f_b
+
+        ctx_size += n_vocab*n_embd*ggml_type_sizef(wtype);         // wte
+        ctx_size +=   n_ctx*n_embd*ggml_type_sizef(GGML_TYPE_F32); // wpe
+        ctx_size += n_vocab*n_embd*ggml_type_sizef(wtype);         // lm_head
+
+        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_1_g
+        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_1_b
+
+        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_2_g
+        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_2_b
+
+        ctx_size += n_layer*(3*n_embd*n_embd*ggml_type_sizef(wtype));         // c_attn_attn_w
+        ctx_size += n_layer*(       3*n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_attn_attn_b
+
+        ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype));           // c_attn_proj_w
+        ctx_size += n_layer*(       n_embd*ggml_type_sizef(GGML_TYPE_F32));   // c_attn_proj_b
+
+        ctx_size += n_layer*(4*n_embd*n_embd*ggml_type_sizef(wtype));         // c_mlp_fc_w
+        ctx_size += n_layer*(       4*n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_mlp_fc_b
+
+        ctx_size += n_layer*(4*n_embd*n_embd*ggml_type_sizef(wtype));         // c_mlp_proj_w
+        ctx_size += n_layer*(         n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_mlp_proj_b
+
+        ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(GGML_TYPE_F32); // memory_k
+        ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(GGML_TYPE_F32); // memory_v
+
+        ctx_size += (6 + 12*n_layer)*512; // object overhead
+
+        printf("%s: ggml tensor size = %d bytes\n", __func__, (int) sizeof(ggml_tensor));
+        printf("%s: ggml ctx size = %6.2f MB\n", __func__, ctx_size/(1024.0*1024.0));
+    }
+
+    // create the ggml context
+    {
+        struct ggml_init_params params = {
+            /*.mem_size   =*/ ctx_size,
+            /*.mem_buffer =*/ NULL,
+            /*.no_alloc   =*/ false,
+        };
+
+        model.ctx = ggml_init(params);
+        if (!model.ctx) {
+            fprintf(stderr, "%s: ggml_init() failed\n", __func__);
+            return false;
+        }
+    }
+
+    // prepare memory for the weights
+    {
+        const auto & hparams = model.hparams;
+
+        const int n_embd  = hparams.n_embd;
+        const int n_layer = hparams.n_layer;
+        const int n_ctx   = hparams.n_ctx;
+        const int n_vocab = hparams.n_vocab;
+
+        model.layers.resize(n_layer);
+
+        model.ln_f_g = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
+        model.ln_f_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
+
+        model.wte     = ggml_new_tensor_2d(ctx, wtype,         n_embd, n_vocab);
+        model.wpe     = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_ctx);
+        model.lm_head = ggml_new_tensor_2d(ctx, wtype,         n_embd, n_vocab);
+
+        // map by name
+        model.tensors["model/ln_f/g"] = model.ln_f_g;
+        model.tensors["model/ln_f/b"] = model.ln_f_b;
+
+        model.tensors["model/wte"]     = model.wte;
+        model.tensors["model/wpe"]     = model.wpe;
+        model.tensors["model/lm_head"] = model.lm_head;
+
+        for (int i = 0; i < n_layer; ++i) {
+            auto & layer = model.layers[i];
+
+            layer.ln_1_g        = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+            layer.ln_1_b        = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+
+            layer.ln_2_g        = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+            layer.ln_2_b        = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+
+            layer.c_attn_attn_w = ggml_new_tensor_2d(ctx, wtype,           n_embd, 3*n_embd);
+            layer.c_attn_attn_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 3*n_embd);
+
+            layer.c_attn_proj_w = ggml_new_tensor_2d(ctx, wtype,           n_embd, n_embd);
+            layer.c_attn_proj_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+
+            layer.c_mlp_fc_w    = ggml_new_tensor_2d(ctx, wtype,           n_embd, 4*n_embd);
+            layer.c_mlp_fc_b    = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 4*n_embd);
+
+            layer.c_mlp_proj_w  = ggml_new_tensor_2d(ctx, wtype,         4*n_embd, n_embd);
+            layer.c_mlp_proj_b  = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+
+            // map by name
+            model.tensors["model/h" + std::to_string(i) + "/ln_1/g"]        = layer.ln_1_g;
+            model.tensors["model/h" + std::to_string(i) + "/ln_1/b"]        = layer.ln_1_b;
+
+            model.tensors["model/h" + std::to_string(i) + "/ln_2/g"]        = layer.ln_2_g;
+            model.tensors["model/h" + std::to_string(i) + "/ln_2/b"]        = layer.ln_2_b;
+
+            model.tensors["model/h" + std::to_string(i) + "/attn/c_attn/w"] = layer.c_attn_attn_w;
+            model.tensors["model/h" + std::to_string(i) + "/attn/c_attn/b"] = layer.c_attn_attn_b;
+
+            model.tensors["model/h" + std::to_string(i) + "/attn/c_proj/w"] = layer.c_attn_proj_w;
+            model.tensors["model/h" + std::to_string(i) + "/attn/c_proj/b"] = layer.c_attn_proj_b;
+
+            model.tensors["model/h" + std::to_string(i) + "/mlp/c_fc/w"]    = layer.c_mlp_fc_w;
+            model.tensors["model/h" + std::to_string(i) + "/mlp/c_fc/b"]    = layer.c_mlp_fc_b;
+
+            model.tensors["model/h" + std::to_string(i) + "/mlp/c_proj/w"]  = layer.c_mlp_proj_w;
+            model.tensors["model/h" + std::to_string(i) + "/mlp/c_proj/b"]  = layer.c_mlp_proj_b;
+        }
+    }
+
+    // key + value memory
+    {
+        const auto & hparams = model.hparams;
+
+        const int n_embd  = hparams.n_embd;
+        const int n_layer = hparams.n_layer;
+        const int n_ctx   = hparams.n_ctx;
+
+        const int n_mem      = n_layer*n_ctx;
+        const int n_elements = n_embd*n_mem;
+
+        model.memory_k = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_elements);
+        model.memory_v = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_elements);
+
+        const size_t memory_size = ggml_nbytes(model.memory_k) + ggml_nbytes(model.memory_v);
+
+        printf("%s: memory size = %8.2f MB, n_mem = %d\n", __func__, memory_size/1024.0/1024.0, n_mem);
+    }
+
+    // load weights
+    {
+        size_t total_size = 0;
+
+        bool has_lm_head = false;
+
+        while (true) {
+            int32_t n_dims;
+            int32_t length;
+            int32_t ttype;
+
+            fin.read(reinterpret_cast<char *>(&n_dims), sizeof(n_dims));
+            fin.read(reinterpret_cast<char *>(&length), sizeof(length));
+            fin.read(reinterpret_cast<char *>(&ttype),  sizeof(ttype));
+
+            if (fin.eof()) {
+                break;
+            }
+
+            int32_t nelements = 1;
+            int32_t ne[2] = { 1, 1 };
+            for (int i = 0; i < n_dims; ++i) {
+                fin.read(reinterpret_cast<char *>(&ne[i]), sizeof(ne[i]));
+                nelements *= ne[i];
+            }
+
+            std::string name(length, 0);
+            fin.read(&name[0], length);
+
+            if (model.tensors.find(name) == model.tensors.end()) {
+                fprintf(stderr, "%s: unknown tensor '%s' in model file\n", __func__, name.c_str());
+                return false;
+            }
+
+            auto tensor = model.tensors[name];
+            if (ggml_nelements(tensor) != nelements) {
+                fprintf(stderr, "%s: tensor '%s' has wrong size in model file\n", __func__, name.c_str());
+                return false;
+            }
+
+            if (tensor->ne[0] != ne[0] || tensor->ne[1] != ne[1]) {
+                fprintf(stderr, "%s: tensor '%s' has wrong shape in model file: got [%d, %d], expected [%d, %d]\n",
+                        __func__, name.c_str(), (int) tensor->ne[0], (int) tensor->ne[1], ne[0], ne[1]);
+                return false;
+            }
+
+            // for debugging
+            if (0) {
+                printf("%24s - [%5d, %5d], type = %6s, %6.2f MB, %9zu bytes\n", name.c_str(), ne[0], ne[1], ggml_type_name(ggml_type(ttype)), ggml_nbytes(tensor)/1024.0/1024.0, ggml_nbytes(tensor));
+            }
+
+            const size_t bpe = ggml_type_size(ggml_type(ttype));
+
+            if ((nelements*bpe)/ggml_blck_size(tensor->type) != ggml_nbytes(tensor)) {
+                fprintf(stderr, "%s: tensor '%s' has wrong size in model file: got %zu, expected %zu\n",
+                        __func__, name.c_str(), ggml_nbytes(tensor), nelements*bpe);
+                return false;
+            }
+
+            fin.read(reinterpret_cast<char *>(tensor->data), ggml_nbytes(tensor));
+
+            // GPT-2 models share the WTE tensor as the LM head
+            if (name == "model/wte" && has_lm_head == false) {
+                memcpy(model.lm_head->data, tensor->data, ggml_nbytes(tensor));
+            }
+
+            if (name == "model/lm_head") {
+                has_lm_head = true;
+            }
+
+            total_size += ggml_nbytes(tensor);
+        }
+
+        printf("%s: model size  = %8.2f MB\n", __func__, total_size/1024.0/1024.0);
+    }
+
+    fin.close();
+
+    return true;
+}
+
+// build the computation graph
+struct ggml_cgraph * gpt2_graph(
+        const gpt2_model & model,
+        struct ggml_allocr * allocr,
+        const int n_past,
+        const std::vector<gpt_vocab::id> & embd_inp) {
+    const int N = embd_inp.size();
+
+    const auto & hparams = model.hparams;
+
+    const int n_embd  = hparams.n_embd;
+    const int n_layer = hparams.n_layer;
+    const int n_ctx   = hparams.n_ctx;
+    const int n_head  = hparams.n_head;
+
+    // since we are using ggml-alloc, this buffer only needs enough space to hold the ggml_tensor and ggml_cgraph structs, but not the tensor data
+    static size_t buf_size = ggml_tensor_overhead()*GGML_MAX_NODES + ggml_graph_overhead();
+    static std::vector<uint8_t> buf(buf_size);
+
+    struct ggml_init_params params = {
+        /*.mem_size   =*/ buf_size,
+        /*.mem_buffer =*/ buf.data(),
+        /*.no_alloc   =*/ true, // the tensors will be allocated later by ggml_allocr_alloc_graph()
+    };
+
+    struct ggml_context * ctx0 = ggml_init(params);
+
+    struct ggml_cgraph  * gf = ggml_new_graph(ctx0);
+
+    struct ggml_tensor * embd = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
+    ggml_allocr_alloc(allocr, embd);
+
+    // avoid writing to tensors if we are only measuring the memory usage
+    if (!ggml_allocr_is_measure(allocr)) {
+        memcpy(embd->data, embd_inp.data(), N*ggml_element_size(embd));
+    }
+
+    struct ggml_tensor * position = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
+    ggml_allocr_alloc(allocr, position);
+    if (!ggml_allocr_is_measure(allocr)) {
+        for (int i = 0; i < N; ++i) {
+            ((int32_t *) position->data)[i] = n_past + i;
+        }
+    }
+
+    struct ggml_tensor * KQ_scale = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, 1);
+    ggml_allocr_alloc(allocr, KQ_scale);
+    if (!ggml_allocr_is_measure(allocr)) {
+        ggml_set_f32(KQ_scale, 1.0f/sqrtf(float(n_embd)/n_head));
+    }
+
+    // wte + wpe
+    struct ggml_tensor * inpL =
+        ggml_add(ctx0,
+                ggml_get_rows(ctx0, model.wte, embd),
+                ggml_get_rows(ctx0, model.wpe, position));
+
+    for (int il = 0; il < n_layer; ++il) {
+        struct ggml_tensor * cur;
+
+        // norm
+        {
+            // [ 768, N]
+            cur = ggml_norm(ctx0, inpL, hparams.eps);
+
+            // cur = ln_1_g*cur + ln_1_b
+            // [ 768, N]
+            cur = ggml_add(ctx0,
+                    ggml_mul(ctx0,
+                        ggml_repeat(ctx0, model.layers[il].ln_1_g, cur),
+                        cur),
+                    ggml_repeat(ctx0, model.layers[il].ln_1_b, cur));
+        }
+
+        // attn
+        // [2304, 768] - model.layers[il].c_attn_attn_w
+        // [2304,   1] - model.layers[il].c_attn_attn_b
+        // [ 768,   N] - cur (in)
+        // [2304,   N] - cur (out)
+        //
+        // cur = attn_w*cur + attn_b
+        // [2304, N]
+        {
+            cur = ggml_mul_mat(ctx0,
+                    model.layers[il].c_attn_attn_w,
+                    cur);
+
+            cur = ggml_add(ctx0,
+                    ggml_repeat(ctx0, model.layers[il].c_attn_attn_b, cur),
+                    cur);
+        }
+
+        // self-attention
+        {
+            struct ggml_tensor * Qcur = ggml_view_2d(ctx0, cur, n_embd, N, cur->nb[1], 0*sizeof(float)*n_embd);
+            struct ggml_tensor * Kcur = ggml_view_2d(ctx0, cur, n_embd, N, cur->nb[1], 1*sizeof(float)*n_embd);
+            struct ggml_tensor * Vcur = ggml_view_2d(ctx0, cur, n_embd, N, cur->nb[1], 2*sizeof(float)*n_embd);
+
+            // store key and value to memory
+            if (N >= 1) {
+                struct ggml_tensor * k = ggml_view_1d(ctx0, model.memory_k, N*n_embd, (ggml_element_size(model.memory_k)*n_embd)*(il*n_ctx + n_past));
+                struct ggml_tensor * v = ggml_view_1d(ctx0, model.memory_v, N*n_embd, (ggml_element_size(model.memory_v)*n_embd)*(il*n_ctx + n_past));
+
+                ggml_build_forward_expand(gf, ggml_cpy(ctx0, Kcur, k));
+                ggml_build_forward_expand(gf, ggml_cpy(ctx0, Vcur, v));
+            }
+
+            // Q = Qcur.contiguous().view(n_embd/n_head, n_head, N).permute(0, 2, 1, 3)
+            // [64, N, 12]
+            struct ggml_tensor * Q =
+                ggml_permute(ctx0,
+                        ggml_cpy(ctx0,
+                            Qcur,
+                            ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_embd/n_head, n_head, N)),
+                        0, 2, 1, 3);
+
+            // K = Kmem.view(n_embd/n_head, n_head, n_past + N).permute(0, 2, 1, 3)
+            // [64, n_past + N, 12]
+            struct ggml_tensor * K =
+                ggml_permute(ctx0,
+                        ggml_reshape_3d(ctx0,
+                            ggml_view_1d(ctx0, model.memory_k, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(model.memory_k)*n_embd),
+                            n_embd/n_head, n_head, n_past + N),
+                        0, 2, 1, 3);
+
+            // GG: flash attention
+            //struct ggml_tensor * V =
+            //    ggml_cpy(ctx0,
+            //            ggml_permute(ctx0,
+            //                ggml_reshape_3d(ctx0,
+            //                    ggml_view_1d(ctx0, model.memory_v, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(model.memory_v)*n_embd),
+            //                    n_embd/n_head, n_head, n_past + N),
+            //                1, 2, 0, 3),
+            //            ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_past + N, n_embd/n_head, n_head));
+
+            //struct ggml_tensor * KQV = ggml_flash_attn(ctx0, Q, K, V, true);
+
+            // K * Q
+            // [n_past + N, N, 12]
+            struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);
+
+            // KQ_scaled = KQ / sqrt(n_embd/n_head)
+            // [n_past + N, N, 12]
+            struct ggml_tensor * KQ_scaled =
+                ggml_scale(ctx0,
+                        KQ,
+                        KQ_scale);
+
+            // KQ_masked = mask_past(KQ_scaled)
+            // [n_past + N, N, 12]
+            struct ggml_tensor * KQ_masked = ggml_diag_mask_inf(ctx0, KQ_scaled, n_past);
+
+            // KQ = soft_max(KQ_masked)
+            // [n_past + N, N, 12]
+            struct ggml_tensor * KQ_soft_max = ggml_soft_max(ctx0, KQ_masked);
+
+            // V_trans = Vmem.view(n_embd/n_head, n_head, n_past + N).permute(1, 2, 0, 3).contiguous()
+            // [n_past + N, 64, 12]
+            struct ggml_tensor * V_trans =
+                ggml_cpy(ctx0,
+                        ggml_permute(ctx0,
+                            ggml_reshape_3d(ctx0,
+                                ggml_view_1d(ctx0, model.memory_v, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(model.memory_v)*n_embd),
+                                n_embd/n_head, n_head, n_past + N),
+                            1, 2, 0, 3),
+                        ggml_new_tensor_3d(ctx0, model.memory_v->type, n_past + N, n_embd/n_head, n_head));
+
+            // KQV = transpose(V) * KQ_soft_max
+            // [64, N, 12]
+            struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V_trans, KQ_soft_max);
+
+            // KQV_merged = KQV.permute(0, 2, 1, 3)
+            // [64, 12, N]
+            struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
+
+            // cur = KQV_merged.contiguous().view(n_embd, N)
+            // [768, N]
+            cur = ggml_cpy(ctx0,
+                    KQV_merged,
+                    ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, N));
+        }
+
+        // projection
+        // [ 768, 768] - model.layers[il].c_attn_proj_w
+        // [ 768,   1] - model.layers[il].c_attn_proj_b
+        // [ 768,   N] - cur (in)
+        // [ 768,   N] - cur (out)
+        //
+        // cur = proj_w*cur + proj_b
+        // [768, N]
+        {
+            cur = ggml_mul_mat(ctx0,
+                    model.layers[il].c_attn_proj_w,
+                    cur);
+
+            cur = ggml_add(ctx0,
+                    ggml_repeat(ctx0, model.layers[il].c_attn_proj_b, cur),
+                    cur);
+        }
+
+        // add the input
+        cur = ggml_add(ctx0, cur, inpL);
+
+        struct ggml_tensor * inpFF = cur;
+
+        // feed-forward network
+        {
+            // norm
+            {
+                cur = ggml_norm(ctx0, inpFF, hparams.eps);
+
+                // cur = ln_2_g*cur + ln_2_b
+                // [ 768, N]
+                cur = ggml_add(ctx0,
+                        ggml_mul(ctx0,
+                            ggml_repeat(ctx0, model.layers[il].ln_2_g, cur),
+                            cur),
+                        ggml_repeat(ctx0, model.layers[il].ln_2_b, cur));
+            }
+
+            // fully connected
+            // [3072, 768] - model.layers[il].c_mlp_fc_w
+            // [3072,   1] - model.layers[il].c_mlp_fc_b
+            // [ 768,   N] - cur (in)
+            // [3072,   N] - cur (out)
+            //
+            // cur = fc_w*cur + fc_b
+            // [3072, N]
+            cur = ggml_mul_mat(ctx0,
+                    model.layers[il].c_mlp_fc_w,
+                    cur);
+
+            cur = ggml_add(ctx0,
+                    ggml_repeat(ctx0, model.layers[il].c_mlp_fc_b, cur),
+                    cur);
+
+            // GELU activation
+            // [3072, N]
+            cur = ggml_gelu(ctx0, cur);
+
+            // projection
+            // [ 768, 3072] - model.layers[il].c_mlp_proj_w
+            // [ 768,    1] - model.layers[il].c_mlp_proj_b
+            // [3072,    N] - cur (in)
+            // [ 768,    N] - cur (out)
+            //
+            // cur = proj_w*cur + proj_b
+            // [768, N]
+            cur = ggml_mul_mat(ctx0,
+                    model.layers[il].c_mlp_proj_w,
+                    cur);
+
+            cur = ggml_add(ctx0,
+                    ggml_repeat(ctx0, model.layers[il].c_mlp_proj_b, cur),
+                    cur);
+        }
+
+        // input for next layer
+        inpL = ggml_add(ctx0, cur, inpFF);
+    }
+
+    // norm
+    {
+        // [ 768, N]
+        inpL = ggml_norm(ctx0, inpL, hparams.eps);
+
+        // inpL = ln_f_g*inpL + ln_f_b
+        // [ 768, N]
+        inpL = ggml_add(ctx0,
+                ggml_mul(ctx0,
+                    ggml_repeat(ctx0, model.ln_f_g, inpL),
+                    inpL),
+                ggml_repeat(ctx0, model.ln_f_b, inpL));
+    }
+
+    // inpL = WTE * inpL
+    // [ 768, 50257] - model.lm_head
+    // [ 768, N]     - inpL
+    inpL = ggml_mul_mat(ctx0, model.lm_head, inpL);
+
+    // logits -> probs
+    //inpL = ggml_soft_max(ctx0, inpL);
+
+    ggml_build_forward_expand(gf, inpL);
+
+    ggml_free(ctx0);
+
+    return gf;
+}
+
+// evaluate the transformer
+//
+//   - model:     the model
+//   - allocr:    ggml_allocr to use to allocate the compute buffer
+//   - n_threads: number of threads to use
+//   - n_past:    the context size so far
+//   - embd_inp:  the embeddings of the tokens in the context
+//   - embd_w:    the predicted logits for the next token
+//
+bool gpt2_eval(
+        const gpt2_model & model,
+        struct ggml_allocr * allocr,
+        const int n_threads,
+        const int n_past,
+        const std::vector<gpt_vocab::id> & embd_inp,
+              std::vector<float>         & embd_w) {
+    const int N = embd_inp.size();
+
+    const auto & hparams = model.hparams;
+
+    const int n_vocab = hparams.n_vocab;
+
+    // reset the allocator to free all the memory allocated during the previous inference
+    ggml_allocr_reset(allocr);
+
+    struct ggml_cgraph * gf = gpt2_graph(model, allocr, n_past, embd_inp);
+
+    // allocate tensors
+    ggml_allocr_alloc_graph(allocr, gf);
+
+    // run the computation
+    struct ggml_cplan plan = ggml_graph_plan(gf, n_threads);
+    static std::vector<uint8_t> work_buffer;
+    work_buffer.resize(plan.work_size);
+    plan.work_data = work_buffer.data();
+    ggml_graph_compute(gf, &plan);
+
+    //if (n_past%100 == 0) {
+    //    ggml_graph_print   (&gf);
+    //    ggml_graph_dump_dot(&gf, NULL, "gpt-2.dot");
+    //}
+
+    // in this case, the output tensor is the last one in the graph
+    struct ggml_tensor * inpL = gf->nodes[gf->n_nodes - 1];
+
+    //embd_w.resize(n_vocab*N);
+    //memcpy(embd_w.data(), ggml_get_data(inpL), sizeof(float)*n_vocab*N);
+
+    // return result just for the last token
+    embd_w.resize(n_vocab);
+    memcpy(embd_w.data(), (float *) ggml_get_data(inpL) + (n_vocab*(N-1)), sizeof(float)*n_vocab);
+
+    return true;
+}
+
+int main(int argc, char ** argv) {
+    ggml_time_init();
+
+    const int64_t t_main_start_us = ggml_time_us();
+
+    gpt_params params;
+    params.model = "models/gpt-2-117M/ggml-model.bin";
+
+    if (gpt_params_parse(argc, argv, params) == false) {
+        return 1;
+    }
+
+    if (params.seed < 0) {
+        params.seed = time(NULL);
+    }
+
+    printf("%s: seed = %d\n", __func__, params.seed);
+
+    std::mt19937 rng(params.seed);
+    if (params.prompt.empty()) {
+        params.prompt = gpt_random_prompt(rng);
+    }
+
+    int64_t t_load_us = 0;
+
+    gpt_vocab vocab;
+    gpt2_model model;
+
+    // load the model
+    {
+        const int64_t t_start_us = ggml_time_us();
+
+        if (!gpt2_model_load(params.model, model, vocab)) {
+            fprintf(stderr, "%s: failed to load model from '%s'\n", __func__, params.model.c_str());
+            return 1;
+        }
+
+        t_load_us = ggml_time_us() - t_start_us;
+
+        test_gpt_tokenizer(vocab, params.token_test);
+    }
+
+    // keep this buffer alive while evaluating the model
+    std::vector<uint8_t> compute_buffer;
+
+    struct ggml_allocr * allocr = NULL;
+    // allocate the compute buffer
+    {
+        allocr = ggml_allocr_new_measure(GGML_MEM_ALIGN);
+
+        // create the worst case graph for memory usage estimation
+        int n_tokens = std::min(model.hparams.n_ctx, params.n_batch);
+        int n_past = model.hparams.n_ctx - n_tokens;
+        struct ggml_cgraph * gf = gpt2_graph(model, allocr, n_past, std::vector<gpt_vocab::id>(n_tokens, 0));
+
+        // compute the required memory
+        size_t mem_size = ggml_allocr_alloc_graph(allocr, gf) + GGML_MEM_ALIGN;
+
+        // recreate the allocator with the required memory
+        ggml_allocr_free(allocr);
+        compute_buffer.resize(mem_size);
+        allocr = ggml_allocr_new(compute_buffer.data(), mem_size, GGML_MEM_ALIGN);
+
+        fprintf(stderr, "%s: compute buffer size: %.2f MB\n", __func__, mem_size/1024.0/1024.0);
+    }
+
+    int n_past = 0;
+
+    int64_t t_sample_us  = 0;
+    int64_t t_predict_us = 0;
+
+    std::vector<float> logits;
+
+    // tokenize the prompt
+    std::vector<gpt_vocab::id> embd_inp = ::gpt_tokenize(vocab, params.prompt);
+
+    params.n_predict = std::min(params.n_predict, model.hparams.n_ctx - (int) embd_inp.size());
+
+    printf("%s: prompt: '%s'\n", __func__, params.prompt.c_str());
+    printf("%s: number of tokens in prompt = %zu, first 8 tokens: ", __func__, embd_inp.size());
+    for (int i = 0; i < std::min(8, (int) embd_inp.size()); i++) {
+        printf("%d ", embd_inp[i]);
+    }
+    printf("\n\n");
+
+    // submit the input prompt token-by-token
+    // this reduces the memory usage during inference, at the cost of a bit of speed at the beginning
+    std::vector<gpt_vocab::id> embd;
+
+    for (size_t i = embd.size(); i < embd_inp.size() + params.n_predict; i++) {
+        // predict
+        if (embd.size() > 0) {
+            const int64_t t_start_us = ggml_time_us();
+
+            if (!gpt2_eval(model, allocr, params.n_threads, n_past, embd, logits)) {
+                printf("Failed to predict\n");
+                return 1;
+            }
+
+            t_predict_us += ggml_time_us() - t_start_us;
+        }
+
+        n_past += embd.size();
+        embd.clear();
+
+        if (i >= embd_inp.size()) {
+            // sample next token
+            const int   top_k = params.top_k;
+            const float top_p = params.top_p;
+            const float temp  = params.temp;
+
+            const int n_vocab = model.hparams.n_vocab;
+
+            gpt_vocab::id id = 0;
+
+            {
+                const int64_t t_start_sample_us = ggml_time_us();
+
+                id = gpt_sample_top_k_top_p(vocab, logits.data() + (logits.size() - n_vocab), top_k, top_p, temp, rng);
+
+                t_sample_us += ggml_time_us() - t_start_sample_us;
+            }
+
+            // add it to the context
+            embd.push_back(id);
+        } else {
+            // if here, it means we are still processing the input prompt
+            for (size_t k = i; k < embd_inp.size(); k++) {
+                embd.push_back(embd_inp[k]);
+                if (int32_t(embd.size()) >= params.n_batch) {
+                    break;
+                }
+            }
+            i += embd.size() - 1;
+        }
+
+        // display text
+        for (auto id : embd) {
+            printf("%s", vocab.id_to_token[id].c_str());
+        }
+        fflush(stdout);
+
+        // end of text token
+        if (embd.back() == 50256) {
+            break;
+        }
+    }
+
+    // report timing
+    {
+        const int64_t t_main_end_us = ggml_time_us();
+
+        printf("\n\n");
+        printf("%s:     load time = %8.2f ms\n", __func__, t_load_us/1000.0f);
+        printf("%s:   sample time = %8.2f ms\n", __func__, t_sample_us/1000.0f);
+        printf("%s:  predict time = %8.2f ms / %.2f ms per token\n", __func__, t_predict_us/1000.0f, t_predict_us/1000.0f/n_past);
+        printf("%s:    total time = %8.2f ms\n", __func__, (t_main_end_us - t_main_start_us)/1000.0f);
+    }
+
+    ggml_free(model.ctx);
+
+    return 0;
+}

ggml/examples/gpt-2/quantize.cpp

@@ -0,0 +1,184 @@
+#include "ggml/ggml.h"
+
+#include "common.h"
+#include "common-ggml.h"
+
+#include <cassert>
+#include <cmath>
+#include <cstdio>
+#include <cstring>
+#include <fstream>
+#include <map>
+#include <string>
+#include <vector>
+#include <regex>
+
+// default hparams (GPT-2 117M)
+struct gpt2_hparams {
+    int32_t n_vocab = 50257;
+    int32_t n_ctx   = 1024;
+    int32_t n_embd  = 768;
+    int32_t n_head  = 12;
+    int32_t n_layer = 12;
+    int32_t ftype   = 1;
+};
+
+// quantize a model
+bool gpt2_model_quantize(const std::string & fname_inp, const std::string & fname_out, ggml_ftype ftype) {
+    gpt_vocab vocab;
+
+    printf("%s: loading model from '%s'\n", __func__, fname_inp.c_str());
+
+    auto finp = std::ifstream(fname_inp, std::ios::binary);
+    if (!finp) {
+        fprintf(stderr, "%s: failed to open '%s' for reading\n", __func__, fname_inp.c_str());
+        return false;
+    }
+
+    auto fout = std::ofstream(fname_out, std::ios::binary);
+    if (!fout) {
+        fprintf(stderr, "%s: failed to open '%s' for writing\n", __func__, fname_out.c_str());
+        return false;
+    }
+
+    // verify magic
+    {
+        uint32_t magic;
+        finp.read((char *) &magic, sizeof(magic));
+        if (magic != GGML_FILE_MAGIC) {
+            fprintf(stderr, "%s: invalid model file '%s' (bad magic)\n", __func__, fname_inp.c_str());
+            return false;
+        }
+
+        fout.write((char *) &magic, sizeof(magic));
+    }
+
+    gpt2_hparams hparams;
+
+    // load hparams
+    {
+        finp.read((char *) &hparams.n_vocab, sizeof(hparams.n_vocab));
+        finp.read((char *) &hparams.n_ctx,   sizeof(hparams.n_ctx));
+        finp.read((char *) &hparams.n_embd,  sizeof(hparams.n_embd));
+        finp.read((char *) &hparams.n_head,  sizeof(hparams.n_head));
+        finp.read((char *) &hparams.n_layer, sizeof(hparams.n_layer));
+        finp.read((char *) &hparams.ftype,   sizeof(hparams.ftype));
+
+        const int32_t qntvr_src =    hparams.ftype / GGML_QNT_VERSION_FACTOR;
+        const int32_t ftype_dst = GGML_QNT_VERSION * GGML_QNT_VERSION_FACTOR + ftype;
+
+        printf("%s: n_vocab     = %d\n", __func__, hparams.n_vocab);
+        printf("%s: n_ctx       = %d\n", __func__, hparams.n_ctx);
+        printf("%s: n_embd      = %d\n", __func__, hparams.n_embd);
+        printf("%s: n_head      = %d\n", __func__, hparams.n_head);
+        printf("%s: n_layer     = %d\n", __func__, hparams.n_layer);
+        printf("%s: ftype (src) = %d\n", __func__, hparams.ftype);
+        printf("%s: qntvr (src) = %d\n", __func__, qntvr_src);
+        printf("%s: ftype (dst) = %d\n", __func__, ftype_dst);
+        printf("%s: qntvr (dst) = %d\n", __func__, GGML_QNT_VERSION);
+
+        fout.write((char *) &hparams.n_vocab, sizeof(hparams.n_vocab));
+        fout.write((char *) &hparams.n_ctx,   sizeof(hparams.n_ctx));
+        fout.write((char *) &hparams.n_embd,  sizeof(hparams.n_embd));
+        fout.write((char *) &hparams.n_head,  sizeof(hparams.n_head));
+        fout.write((char *) &hparams.n_layer, sizeof(hparams.n_layer));
+        fout.write((char *) &ftype_dst,       sizeof(ftype_dst));
+    }
+
+    // load vocab
+    {
+        int32_t n_vocab = 0;
+        finp.read ((char *) &n_vocab, sizeof(n_vocab));
+        fout.write((char *) &n_vocab, sizeof(n_vocab));
+
+        if (n_vocab != hparams.n_vocab) {
+            fprintf(stderr, "%s: invalid model file '%s' (bad vocab size %d != %d)\n",
+                    __func__, fname_inp.c_str(), n_vocab, hparams.n_vocab);
+            return false;
+        }
+
+        std::string word;
+        for (int i = 0; i < n_vocab; i++) {
+            uint32_t len;
+            finp.read ((char *) &len, sizeof(len));
+            fout.write((char *) &len, sizeof(len));
+
+            word.resize(len);
+            finp.read ((char *) word.data(), len);
+            fout.write((char *) word.data(), len);
+
+            vocab.token_to_id[word] = i;
+            vocab.id_to_token[i] = word;
+        }
+    }
+
+    // regexes of tensor names to be quantized
+    const std::vector<std::string> to_quant = {
+        "model/wte",
+        "model/lm_head",
+        "model/h.*/attn/c_attn/w",
+        "model/h.*/attn/c_proj/w",
+        "model/h.*/mlp/c_fc/w",
+        "model/h.*/mlp/c_proj/w",
+    };
+
+    if (!ggml_common_quantize_0(finp, fout, ftype, to_quant, {})) {
+        fprintf(stderr, "%s: failed to quantize model '%s'\n", __func__, fname_inp.c_str());
+        return false;
+    }
+
+    finp.close();
+    fout.close();
+
+    return true;
+}
+
+// usage:
+//  ./gpt-2-quantize models/gpt-2-117M/ggml-model.bin models/gpt-2-117M/ggml-model-quant.bin type
+//
+int main(int argc, char ** argv) {
+    if (argc != 4) {
+        fprintf(stderr, "usage: %s model-f32.bin model-quant.bin type\n", argv[0]);
+        ggml_print_ftypes(stderr);
+        return 1;
+    }
+
+    // needed to initialize f16 tables
+    {
+        struct ggml_init_params params = { 0, NULL, false };
+        struct ggml_context * ctx = ggml_init(params);
+        ggml_free(ctx);
+    }
+
+    const std::string fname_inp = argv[1];
+    const std::string fname_out = argv[2];
+
+    const ggml_ftype ftype = ggml_parse_ftype(argv[3]);
+
+    const int64_t t_main_start_us = ggml_time_us();
+
+    int64_t t_quantize_us = 0;
+
+    // load the model
+    {
+        const int64_t t_start_us = ggml_time_us();
+
+        if (!gpt2_model_quantize(fname_inp, fname_out, ggml_ftype(ftype))) {
+            fprintf(stderr, "%s: failed to quantize model from '%s'\n", __func__, fname_inp.c_str());
+            return 1;
+        }
+
+        t_quantize_us = ggml_time_us() - t_start_us;
+    }
+
+    // report timing
+    {
+        const int64_t t_main_end_us = ggml_time_us();
+
+        printf("\n");
+        printf("%s: quantize time = %8.2f ms\n", __func__, t_quantize_us/1000.0f);
+        printf("%s:    total time = %8.2f ms\n", __func__, (t_main_end_us - t_main_start_us)/1000.0f);
+    }
+
+    return 0;
+}

ggml/examples/gpt-j/CMakeLists.txt

@@ -0,0 +1,13 @@
+#
+# gpt-j
+
+set(TEST_TARGET gpt-j)
+add_executable(${TEST_TARGET} main.cpp)
+target_link_libraries(${TEST_TARGET} PRIVATE ggml common common-ggml)
+
+#
+# gpt-j-quantize
+
+set(TEST_TARGET gpt-j-quantize)
+add_executable(${TEST_TARGET} quantize.cpp)
+target_link_libraries(${TEST_TARGET} PRIVATE ggml common common-ggml)

ggml/examples/gpt-j/README.md

@@ -0,0 +1,246 @@
+# gpt-j
+
+Local GPT-J inference on your computer using C/C++
+
+No video card required. You just need to have 16 GB of RAM.
+
+## Motivation
+
+The GPT-J 6B model is the open-source alternative to OpenAI's GPT-3. It's basically a neural network that allows you to
+generate coherent, human-like text given a certain context (prompt).
+
+The GPT-J model is quite big - the compact version of the model uses 16-bit floating point representation of the weights
+and is still 12 GB big. This means that in order to run inference on your computer, you would need to have a video card
+with at least 12 GB of video RAM. Alternatively, you can try to run the python implementations on the CPU, but that
+would probably not be very efficient as they are primarily optimized for running on a GPU (or at least this is my guess -
+I don't have much experience with python).
+
+I wanted to try and run the model on my MacBook, so I decided to implement the model inference from scratch using my own
+custom build tensor library. The tensor library (called [ggml](https://github.com/ggerganov/ggml), written in C) is in
+early development stage, but it already allows me to run the GPT-J model.
+
+On my 32GB MacBook M1 Pro, I achieve an inference speed of about `125 ms/token` or about ~6 words per second (1 word
+typically consists of 1 or 2 tokens).
+
+Here is a sample run with prompt `int main(int argc, char ** argv) {`:
+
+```
+$ time ./bin/gpt-j -p "int main(int argc, char ** argv) {"
+
+gptj_model_load: loading model from 'models/gpt-j-6B/ggml-model.bin' - please wait ...
+gptj_model_load: n_vocab = 50400
+gptj_model_load: n_ctx   = 2048
+gptj_model_load: n_embd  = 4096
+gptj_model_load: n_head  = 16
+gptj_model_load: n_layer = 28
+gptj_model_load: n_rot   = 64
+gptj_model_load: f16     = 1
+gptj_model_load: ggml ctx size = 13334.86 MB
+gptj_model_load: memory_size =  1792.00 MB, n_mem = 57344
+gptj_model_load: ................................... done
+gptj_model_load: model size = 11542.79 MB / num tensors = 285
+main: number of tokens in prompt = 13
+
+int main(int argc, char ** argv) {
+    (void)argc;
+    (void)argv;
+
+    {
+        struct sockaddr_in addr;
+        int addrlen;
+        char * ip = "192.168.1.4";
+        int i;
+
+        if ( (addrlen = sizeof(addr)) == -1 )
+            return -1;
+
+        for (i = 0; i < 10; ++i) {
+            addr.sin_family = AF_INET;
+            addr.sin_addr.s_addr = inet_addr(ip);
+
+main: mem per token = 16430420 bytes
+main:     load time =  6211.48 ms
+main:   sample time =    13.74 ms
+main:  predict time = 26420.34 ms / 124.62 ms per token
+main:    total time = 33035.37 ms
+
+real	0m33.171s
+user	3m32.269s
+sys      0m3.686s
+
+$
+```
+
+It took ~6.2 seconds to load the model to memory. After that, it took ~26.4 seconds to generate 200 tokens of what
+looks like to be the beginning of a networking program in C. Pretty cool!
+
+Here is another run, just for fun:
+
+```
+time ./bin/gpt-j -n 500 -t 8 -p "Ask HN: Inherited the worst code and tech team I have ever seen. How to fix it?
+"
+
+gptj_model_load: loading model from 'models/gpt-j-6B/ggml-model.bin' - please wait ...
+gptj_model_load: n_vocab = 50400
+gptj_model_load: n_ctx   = 2048
+gptj_model_load: n_embd  = 4096
+gptj_model_load: n_head  = 16
+gptj_model_load: n_layer = 28
+gptj_model_load: n_rot   = 64
+gptj_model_load: f16     = 1
+gptj_model_load: ggml ctx size = 13334.86 MB
+gptj_model_load: memory_size =  1792.00 MB, n_mem = 57344
+gptj_model_load: ................................... done
+gptj_model_load: model size = 11542.79 MB / num tensors = 285
+main: number of tokens in prompt = 24
+
+Ask HN: Inherited the worst code and tech team I have ever seen. How to fix it?
+
+I've inherited a team with some very strange and un-documented practices, one of them is that they use an old custom
+application with a very slow tech stack written in Python that the team doesn't want to touch but also doesn't want to
+throw away as it has some "legacy" code in it.
+
+The problem is, the tech stack is very very slow.
+
+They have a single web server on a VM that is slow.
+The server is a little bit busy (not very busy though) and they have a lot of processes (30+ that are constantly being
+spawned by the application)
+They have an application that is single threaded and was written in Python and the team don't want to touch this, and
+the application is very slow.
+
+My task as a new member of the team is to fix this.
+
+I'm a senior dev on the team (3 years on the project) and have been told that I will take the lead on this task. I know
+next to nothing about Python. So here is what I have so far.
+
+What I have done is I've been trying to debug the processes with the "ps" command. This way I can see what is running
+and where. From what I see, the application spawns 10 processes a minute and some of them are used for nothing.
+
+I have also started to look for the code. The application source is not in GitHub or any other repository, it is only on
+our internal GitLab.
+
+What I've found so far:
+
+The application uses a custom SQLAlchemy implementation to interact with the data. I've looked at the source, it looks
+like an object cache or something like that. But from what I've seen, the cache gets full every 20 minutes and then gets
+cleared with a special command.
+
+Another strange thing is that the application creates a file for every entry in the database (even if the entry already
+exists). I've looked at the file to see if it contains something, but it seems to be a JSON file with lots of records.
+
+The other strange thing is that I can only find the database tables in the GitLab repository and not the code. So I
+can't really understand how the application is supposed to interact with the database.
+
+I also found a "log" directory, but the code is encrypted with AES. From what I've found, it is in
+
+main: mem per token = 16430420 bytes
+main:     load time =  3900.10 ms
+main:   sample time =    32.58 ms
+main:  predict time = 68049.91 ms / 130.11 ms per token
+main:    total time = 73020.05 ms
+
+real	1m13.156s
+user	9m1.328s
+sys.    0m7.103s
+```
+
+## Implementation details
+
+The high level implementation of the model is contained in the [main.cpp](main.cpp) file. The core computations are
+performed by the [ggml](https://github.com/ggerganov/ggml/blob/master/include/ggml/ggml.h) library.
+
+
+#### Matrix multiplication
+
+The most performance critical part of the implementation is of course the matrix multiplication routine. 99% of the time
+is spent here, so it was important to optimize this as much as possible.
+
+On Arm64, I utilize the 128-bit NEON intrinsics for 16-bit floating point operations:
+
+https://github.com/ggerganov/ggml/blob/fb558f78d905f85c54813602649ddd628ffe0f3a/src/ggml.c#L187-L243
+
+These instructions allow each core to operate simultaneously on 64 16-bit floats. I'm no expert in SIMD, but after quite
+some trials this was the most efficient code for dot product of a row and column that I could come up with. Combined
+with the parallel computation on 8 CPU threads, I believe I'm close to the maximum performance that one could possibly
+get on the M1 CPU. Still, I'm curious to know if there is a more efficient way to implement this.
+
+
+#### Attempt to use the M1 GPU
+
+One interesting property of the GPT-J transformer architecture is that it allows you to perform part of the inference in
+parallel - i.e. the Feed-forward network can be computed in parallel to the Self-attention layer:
+
+https://github.com/ggerganov/ggml/blob/fb558f78d905f85c54813602649ddd628ffe0f3a/examples/gpt-j/main.cpp#L507-L531
+
+So I thought why not try and bring in the M1 GPU to compute half of the neural network in parallel to the CPU and
+potentially gain some extra performance. Thanks to the M1's shared memory model, it was relatively easy to offload part
+of the computation to the GPU using Apple's [Metal Performance
+Shaders](https://developer.apple.com/documentation/metalperformanceshaders). The GPU shares the host memory, so there is
+no need to copy the data back and forth as you would normally do with Cuda or OpenCL. The weight matrices are directly
+available to be used by the GPU.
+
+However, to my surprise, using MPS together with the CPU did not lead to any performance improvement at all. My
+conclusion was that the 8-thread NEON CPU computation is already saturating the memory bandwidth of the M1 and since
+the CPU and the GPU on the MacBook are sharing that bandwidth, it does not help to offload the computation to the GPU.
+Another observation was that the MPS GPU matrix multiplication using 16-bit floats had the same performance as the
+8-thread NEON CPU implementation. Again, I explain this with a saturated memory channel. But of course, my explanation
+could be totally wrong and somehow the implementation wasn't utilizing the resources correctly.
+
+In the end, I decided to not use MPS or the GPU all together.
+
+### Zero memory allocations
+
+Another property of my implementation is that it does not perform any memory allocations once the model is loaded into
+memory. All required memory is allocated at the start of the program with a single `malloc` (technically 2 calls, but
+that is not important).
+
+## Usage
+
+If you want to give this a try and you are on Linux or Mac OS, simply follow these instructions:
+
+```bash
+# Clone the ggml library and build the gpt-j example
+git clone https://github.com/ggerganov/ggml
+cd ggml
+mkdir build && cd build
+cmake ..
+make -j4 gpt-j
+
+# Download the ggml-compatible GPT-J 6B model (requires 12GB disk space)
+../examples/gpt-j/download-ggml-model.sh 6B
+
+# Run the inference (requires 16GB of CPU RAM)
+./bin/gpt-j -m models/gpt-j-6B/ggml-model.bin -p "This is an example"
+
+# Input prompt through pipe and run the inference.
+echo "This is an example" > prompt.txt
+cat prompt.txt | ./bin/gpt-j -m models/gpt-j-6B/ggml-model.bin
+```
+
+To run the `gpt-j` tool, you need the 12GB `ggml-model.bin` file which contains the GPT-J model in
+[ggml](https://github.com/ggerganov/ggml) compatible format. In the instructions above, the binary file
+is downloaded from my repository on Hugging Face using the [download-ggml-model.sh](download-ggml-model.sh) script.
+You can also, download the file manually from this link:
+
+https://huggingface.co/ggerganov/ggml/tree/main
+
+---
+
+Alternatively, if you don't want to download the 12GB ggml model file, you can perform the conversion yourself using
+python.
+
+First, you need to download the full GPT-J model from here: https://huggingface.co/EleutherAI/gpt-j-6B
+
+Note that the full model is quite big - about 72 GB. After you download it, you need to convert it to ggml format using
+the [convert-h5-to-ggml.py](convert-h5-to-ggml.py) script. This will generate the `ggml-model.bin` file, which you can
+then use with the `gpt-j` program.
+
+
+## GPT-2
+
+I also implemented a tool for CPU inference using the smaller GPT-2 models. They have worse quality compared to GPT-J,
+but are much faster to execute.
+
+For example, the Small GPT-2 model is only 240 MB big and the inference speed on my MacBook is about 200 tokens/sec.
+
+For more details, checkout the GPT-2 example here: [gpt-2](https://github.com/ggerganov/ggml/tree/master/examples/gpt-2)

ggml/examples/gpt-j/convert-h5-to-ggml.py

@@ -0,0 +1,173 @@
+# Convert GPT-J-6B h5 transformer model to ggml format
+#
+# Load the model using GPTJForCausalLM.
+# Iterate over all variables and write them to a binary file.
+#
+# For each variable, write the following:
+#   - Number of dimensions (int)
+#   - Name length (int)
+#   - Dimensions (int[n_dims])
+#   - Name (char[name_length])
+#   - Data (float[n_dims])
+#
+# By default, the bigger matrices are converted to 16-bit floats.
+# This can be disabled by adding the "use-f32" CLI argument.
+#
+# At the start of the ggml file we write the model parameters
+# and vocabulary.
+#
+
+import sys
+import struct
+import json
+import torch
+import numpy as np
+
+from transformers import GPTJForCausalLM
+
+# ref: https://github.com/openai/gpt-2/blob/master/src/encoder.py
+def bytes_to_unicode():
+    """
+    Returns list of utf-8 byte and a corresponding list of unicode strings.
+    The reversible bpe codes work on unicode strings.
+    This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.
+    When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.
+    This is a signficant percentage of your normal, say, 32K bpe vocab.
+    To avoid that, we want lookup tables between utf-8 bytes and unicode strings.
+    And avoids mapping to whitespace/control characters the bpe code barfs on.
+    """
+    bs = list(range(ord("!"), ord("~")+1))+list(range(ord("¡"), ord("¬")+1))+list(range(ord("®"), ord("ÿ")+1))
+    cs = bs[:]
+    n = 0
+    for b in range(2**8):
+        if b not in bs:
+            bs.append(b)
+            cs.append(2**8+n)
+            n += 1
+    cs = [chr(n) for n in cs]
+    return dict(zip(bs, cs))
+
+if len(sys.argv) < 3:
+    print("Usage: convert-h5-to-ggml.py dir-model [use-f32]\n")
+    print("  ftype == 0 -> float32")
+    print("  ftype == 1 -> float16")
+    sys.exit(1)
+
+# output in the same directory as the model
+dir_model = sys.argv[1]
+fname_out = sys.argv[1] + "/ggml-model.bin"
+
+with open(dir_model + "/vocab.json", "r", encoding="utf-8") as f:
+    encoder = json.load(f)
+
+with open(dir_model + "/added_tokens.json", "r", encoding="utf-8") as f:
+    encoder_added = json.load(f)
+
+with open(dir_model + "/config.json", "r", encoding="utf-8") as f:
+    hparams = json.load(f)
+
+# possible data types
+#   ftype == 0 -> float32
+#   ftype == 1 -> float16
+#
+# map from ftype to string
+ftype_str = ["f32", "f16"]
+
+ftype = 1
+if len(sys.argv) > 2:
+    ftype = int(sys.argv[2])
+    if ftype < 0 or ftype > 1:
+        print("Invalid ftype: " + str(ftype))
+        sys.exit(1)
+    fname_out = sys.argv[1] + "/ggml-model-" + ftype_str[ftype] + ".bin"
+
+
+model = GPTJForCausalLM.from_pretrained(dir_model, low_cpu_mem_usage=True)
+#print (model)
+
+list_vars = model.state_dict()
+#print (list_vars)
+
+fout = open(fname_out, "wb")
+
+fout.write(struct.pack("i", 0x67676d6c)) # magic: ggml in hex
+fout.write(struct.pack("i", hparams["vocab_size"]))
+fout.write(struct.pack("i", hparams["n_positions"]))
+fout.write(struct.pack("i", hparams["n_embd"]))
+fout.write(struct.pack("i", hparams["n_head"]))
+fout.write(struct.pack("i", hparams["n_layer"]))
+fout.write(struct.pack("i", hparams["rotary_dim"]))
+fout.write(struct.pack("i", ftype))
+
+byte_encoder = bytes_to_unicode()
+byte_decoder = {v:k for k, v in byte_encoder.items()}
+
+fout.write(struct.pack("i", len(encoder) + len(encoder_added)))
+
+for key in encoder:
+    text = bytearray([byte_decoder[c] for c in key])
+    fout.write(struct.pack("i", len(text)))
+    fout.write(text)
+
+for key in encoder_added:
+    text = bytearray([byte_decoder[c] for c in key])
+    fout.write(struct.pack("i", len(text)))
+    fout.write(text)
+
+for name in list_vars.keys():
+    data = list_vars[name].squeeze().numpy()
+    print("Processing variable: " + name + " with shape: ", data.shape)
+
+    # we don't need these
+    if name.endswith("attn.masked_bias") or name.endswith(".attn.bias"):
+        print("  Skipping variable: " + name)
+        continue
+
+    n_dims = len(data.shape);
+
+    # ftype == 0 -> float32, ftype == 1 -> float16
+    ftype_cur = 0;
+    if ftype != 0:
+        if name[-7:] == ".weight" and n_dims == 2:
+            print("  Converting to float16")
+            data = data.astype(np.float16)
+            ftype_cur = 1
+        else:
+            print("  Converting to float32")
+            data = data.astype(np.float32)
+            ftype_cur = 0
+    else:
+        if data.dtype != np.float32:
+            print("  Converting to float32")
+            data = data.astype(np.float32)
+            ftype_cur = 0
+
+    # for efficiency - transpose these matrices:
+    # (note - with latest ggml this is no longer more efficient, so disabling it)
+    #  "transformer.h.*.mlp.fc_in.weight"
+    #  "transformer.h.*.attn.out_proj.weight"
+    #  "transformer.h.*.attn.q_proj.weight"
+    #  "transformer.h.*.attn.k_proj.weight"
+    #  "transformer.h.*.attn.v_proj.weight"
+    #if name.endswith(".mlp.fc_in.weight")     or \
+    #   name.endswith(".attn.out_proj.weight") or \
+    #   name.endswith(".attn.q_proj.weight")   or \
+    #   name.endswith(".attn.k_proj.weight")   or \
+    #   name.endswith(".attn.v_proj.weight"):
+    #    print("  Transposing")
+    #    data = data.transpose()
+
+    # header
+    str = name.encode('utf-8')
+    fout.write(struct.pack("iii", n_dims, len(str), ftype_cur))
+    for i in range(n_dims):
+        fout.write(struct.pack("i", data.shape[n_dims - 1 - i]))
+    fout.write(str);
+
+    # data
+    data.tofile(fout)
+
+fout.close()
+
+print("Done. Output file: " + fname_out)
+print("")

ggml/examples/gpt-j/download-ggml-model.sh

@@ -0,0 +1,69 @@
+#!/bin/bash
+
+# This script downloads GPT-J model files that have already been converted to ggml format.
+# This way you don't have to convert them yourself.
+#
+# If you want to download the original GPT-J model files, use the "download-model.sh" script instead.
+
+#src="https://ggml.ggerganov.com"
+#pfx="ggml-model-gpt-j"
+
+src="https://huggingface.co/ggerganov/ggml"
+pfx="resolve/main/ggml-model-gpt-j"
+
+ggml_path=$(dirname $(realpath $0))
+
+# GPT-J models
+models=( "6B" )
+
+# list available models
+function list_models {
+    printf "\n"
+    printf "  Available models:"
+    for model in "${models[@]}"; do
+        printf " $model"
+    done
+    printf "\n\n"
+}
+
+if [ "$#" -ne 1 ]; then
+    printf "Usage: $0 <model>\n"
+    list_models
+
+    exit 1
+fi
+
+model=$1
+
+if [[ ! " ${models[@]} " =~ " ${model} " ]]; then
+    printf "Invalid model: $model\n"
+    list_models
+
+    exit 1
+fi
+
+# download ggml model
+
+printf "Downloading ggml model $model ...\n"
+
+mkdir -p models/gpt-j-$model
+
+if [ -x "$(command -v wget)" ]; then
+    wget --quiet --show-progress -O models/gpt-j-$model/ggml-model.bin $src/$pfx-$model.bin
+elif [ -x "$(command -v curl)" ]; then
+    curl -L --output models/gpt-j-$model/ggml-model.bin $src/$pfx-$model.bin
+else
+    printf "Either wget or curl is required to download models.\n"
+    exit 1
+fi
+
+if [ $? -ne 0 ]; then
+    printf "Failed to download ggml model $model \n"
+    printf "Please try again later or download the original GPT-J model files and convert them yourself.\n"
+    exit 1
+fi
+
+printf "Done! Model '$model' saved in 'models/gpt-j-$model/ggml-model.bin'\n"
+printf "You can now use it like this:\n\n"
+printf "  $ ./bin/gpt-j -m models/gpt-j-$model/ggml-model.bin -p \"This is an example\"\n"
+printf "\n"

ggml/examples/gpt-j/download-model.sh

@@ -0,0 +1,11 @@
+#!/bin/bash
+
+printf "To obtain the GPT-J 6B model files, please visit: https://huggingface.co/EleutherAI/gpt-j-6B\n\n"
+
+printf "The model is very big. For example, the reposirory above is 72GB in size.\n"
+printf "If you are sure that you want to clone it, simply run the following command:\n\n"
+
+printf " $ git clone https://huggingface.co/EleutherAI/gpt-j-6B models/gpt-j-6B\n\n"
+
+printf "Alternatively, use the 'download-ggml-model.sh' script to download a 12GB ggml version of the model.\n"
+printf "This version is enough to run inference using the ggml library.\n\n"

ggml/examples/gpt-j/main.cpp

@@ -0,0 +1,748 @@
+#include "ggml/ggml.h"
+
+#include "common.h"
+#include "common-ggml.h"
+
+#include <cassert>
+#include <cmath>
+#include <cstdio>
+#include <cstring>
+#include <fstream>
+#include <map>
+#include <string>
+#include <vector>
+
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
+
+// default hparams (GPT-J 6B)
+struct gptj_hparams {
+    int32_t n_vocab = 50400;
+    int32_t n_ctx   = 2048;
+    int32_t n_embd  = 4096;
+    int32_t n_head  = 16;
+    int32_t n_layer = 28;
+    int32_t n_rot   = 64;
+    int32_t ftype   = 1;
+    float   eps     = 1e-5f;
+};
+
+struct gptj_layer {
+    // normalization
+    struct ggml_tensor * ln_1_g;
+    struct ggml_tensor * ln_1_b;
+
+    // attention
+    struct ggml_tensor * c_attn_q_proj_w;
+    struct ggml_tensor * c_attn_k_proj_w;
+    struct ggml_tensor * c_attn_v_proj_w;
+
+    struct ggml_tensor * c_attn_proj_w;
+
+    // ff
+    struct ggml_tensor * c_mlp_fc_w;
+    struct ggml_tensor * c_mlp_fc_b;
+
+    struct ggml_tensor * c_mlp_proj_w;
+    struct ggml_tensor * c_mlp_proj_b;
+};
+
+struct gptj_model {
+    gptj_hparams hparams;
+
+    // normalization
+    struct ggml_tensor * ln_f_g;
+    struct ggml_tensor * ln_f_b;
+
+    struct ggml_tensor * wte; // position embedding
+
+    struct ggml_tensor * lmh_g; // language model head
+    struct ggml_tensor * lmh_b; // language model bias
+
+    std::vector<gptj_layer> layers;
+
+    // key + value memory
+    struct ggml_tensor * memory_k;
+    struct ggml_tensor * memory_v;
+
+    //
+    struct ggml_context * ctx;
+    std::map<std::string, struct ggml_tensor *> tensors;
+};
+
+// load the model's weights from a file
+bool gptj_model_load(const std::string & fname, gptj_model & model, gpt_vocab & vocab) {
+    printf("%s: loading model from '%s' - please wait ...\n", __func__, fname.c_str());
+
+    auto fin = std::ifstream(fname, std::ios::binary);
+    if (!fin) {
+        fprintf(stderr, "%s: failed to open '%s'\n", __func__, fname.c_str());
+        return false;
+    }
+
+    // verify magic
+    {
+        uint32_t magic;
+        fin.read((char *) &magic, sizeof(magic));
+        if (magic != GGML_FILE_MAGIC) {
+            fprintf(stderr, "%s: invalid model file '%s' (bad magic)\n", __func__, fname.c_str());
+            return false;
+        }
+    }
+
+    // load hparams
+    {
+        auto & hparams = model.hparams;
+
+        fin.read((char *) &hparams.n_vocab, sizeof(hparams.n_vocab));
+        fin.read((char *) &hparams.n_ctx,   sizeof(hparams.n_ctx));
+        fin.read((char *) &hparams.n_embd,  sizeof(hparams.n_embd));
+        fin.read((char *) &hparams.n_head,  sizeof(hparams.n_head));
+        fin.read((char *) &hparams.n_layer, sizeof(hparams.n_layer));
+        fin.read((char *) &hparams.n_rot,   sizeof(hparams.n_rot));
+        fin.read((char *) &hparams.ftype,   sizeof(hparams.ftype));
+
+        const int32_t qntvr = hparams.ftype / GGML_QNT_VERSION_FACTOR;
+
+        printf("%s: n_vocab = %d\n", __func__, hparams.n_vocab);
+        printf("%s: n_ctx   = %d\n", __func__, hparams.n_ctx);
+        printf("%s: n_embd  = %d\n", __func__, hparams.n_embd);
+        printf("%s: n_head  = %d\n", __func__, hparams.n_head);
+        printf("%s: n_layer = %d\n", __func__, hparams.n_layer);
+        printf("%s: n_rot   = %d\n", __func__, hparams.n_rot);
+        printf("%s: ftype   = %d\n", __func__, hparams.ftype);
+        printf("%s: qntvr   = %d\n", __func__, qntvr);
+
+        hparams.ftype %= GGML_QNT_VERSION_FACTOR;
+    }
+
+    // load vocab
+    {
+        int32_t n_vocab = 0;
+        fin.read((char *) &n_vocab, sizeof(n_vocab));
+
+        if (n_vocab != model.hparams.n_vocab) {
+            fprintf(stderr, "%s: invalid model file '%s' (bad vocab size %d != %d)\n",
+                    __func__, fname.c_str(), n_vocab, model.hparams.n_vocab);
+            return false;
+        }
+
+        std::string word;
+        std::vector<char> buf(128);
+
+        for (int i = 0; i < n_vocab; i++) {
+            uint32_t len;
+            fin.read((char *) &len, sizeof(len));
+
+            buf.resize(len);
+            fin.read((char *) buf.data(), len);
+            word.assign(buf.data(), len);
+
+            vocab.token_to_id[word] = i;
+            vocab.id_to_token[i] = word;
+        }
+    }
+
+    // for the big tensors, we have the option to store the data in 16-bit floats or quantized
+    // in order to save memory and also to speed up the computation
+    ggml_type wtype = ggml_ftype_to_ggml_type((ggml_ftype) (model.hparams.ftype));
+    if (wtype == GGML_TYPE_COUNT) {
+        fprintf(stderr, "%s: invalid model file '%s' (bad ftype value %d)\n",
+                __func__, fname.c_str(), model.hparams.ftype);
+        return false;
+    }
+
+    auto & ctx = model.ctx;
+
+    size_t ctx_size = 0;
+
+    {
+        const auto & hparams = model.hparams;
+
+        const int n_embd  = hparams.n_embd;
+        const int n_layer = hparams.n_layer;
+        const int n_ctx   = hparams.n_ctx;
+        const int n_vocab = hparams.n_vocab;
+
+        ctx_size += n_embd*ggml_type_sizef(GGML_TYPE_F32); // ln_f_g
+        ctx_size += n_embd*ggml_type_sizef(GGML_TYPE_F32); // ln_f_b
+
+        ctx_size += n_embd*n_vocab*ggml_type_sizef(wtype); // wte
+
+        ctx_size += n_embd*n_vocab*ggml_type_sizef(wtype);         // lmh_g
+        ctx_size +=        n_vocab*ggml_type_sizef(GGML_TYPE_F32); // lmh_b
+
+        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_1_g
+        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_1_b
+
+        ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // c_attn_q_proj_w
+        ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // c_attn_k_proj_w
+        ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // c_attn_v_proj_w
+
+        ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // c_attn_proj_w
+
+        ctx_size += n_layer*(4*n_embd*n_embd*ggml_type_sizef(wtype));         // c_mlp_fc_w
+        ctx_size += n_layer*(       4*n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_mlp_fc_b
+
+        ctx_size += n_layer*(4*n_embd*n_embd*ggml_type_sizef(wtype));         // c_mlp_proj_w
+        ctx_size += n_layer*(         n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_mlp_proj_b
+
+        ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(GGML_TYPE_F16); // memory_k
+        ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(GGML_TYPE_F16); // memory_v
+
+        ctx_size += (5 + 10*n_layer)*512; // object overhead
+
+        printf("%s: ggml ctx size = %6.2f MB\n", __func__, ctx_size/(1024.0*1024.0));
+    }
+
+    // create the ggml context
+    {
+        struct ggml_init_params params = {
+            /*.mem_size   =*/ ctx_size,
+            /*.mem_buffer =*/ NULL,
+            /*.no_alloc   =*/ false,
+        };
+
+        model.ctx = ggml_init(params);
+        if (!model.ctx) {
+            fprintf(stderr, "%s: ggml_init() failed\n", __func__);
+            return false;
+        }
+    }
+
+    // prepare memory for the weights
+    {
+        const auto & hparams = model.hparams;
+
+        const int n_embd  = hparams.n_embd;
+        const int n_layer = hparams.n_layer;
+        const int n_vocab = hparams.n_vocab;
+
+        model.layers.resize(n_layer);
+
+        model.wte    = ggml_new_tensor_2d(ctx, wtype,         n_embd, n_vocab);
+
+        model.ln_f_g = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
+        model.ln_f_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
+
+        model.lmh_g  = ggml_new_tensor_2d(ctx, wtype,         n_embd, n_vocab);
+        model.lmh_b  = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_vocab);
+
+        // map by name
+        model.tensors["transformer.wte.weight"] = model.wte;
+
+        model.tensors["transformer.ln_f.weight"] = model.ln_f_g;
+        model.tensors["transformer.ln_f.bias"]   = model.ln_f_b;
+
+        model.tensors["lm_head.weight"] = model.lmh_g;
+        model.tensors["lm_head.bias"]   = model.lmh_b;
+
+        for (int i = 0; i < n_layer; ++i) {
+            auto & layer = model.layers[i];
+
+            layer.ln_1_g          = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+            layer.ln_1_b          = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+
+            layer.c_attn_q_proj_w = ggml_new_tensor_2d(ctx, wtype,           n_embd,   n_embd);
+            layer.c_attn_k_proj_w = ggml_new_tensor_2d(ctx, wtype,           n_embd,   n_embd);
+            layer.c_attn_v_proj_w = ggml_new_tensor_2d(ctx, wtype,           n_embd,   n_embd);
+
+            layer.c_attn_proj_w   = ggml_new_tensor_2d(ctx, wtype,           n_embd,   n_embd);
+
+            layer.c_mlp_fc_w      = ggml_new_tensor_2d(ctx, wtype,           n_embd, 4*n_embd);
+            layer.c_mlp_fc_b      = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 4*n_embd);
+
+            layer.c_mlp_proj_w    = ggml_new_tensor_2d(ctx, wtype,         4*n_embd,   n_embd);
+            layer.c_mlp_proj_b    = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+
+            // map by name
+            model.tensors["transformer.h." + std::to_string(i) + ".ln_1.weight"]          = layer.ln_1_g;
+            model.tensors["transformer.h." + std::to_string(i) + ".ln_1.bias"]            = layer.ln_1_b;
+
+            model.tensors["transformer.h." + std::to_string(i) + ".attn.q_proj.weight"]   = layer.c_attn_q_proj_w;
+            model.tensors["transformer.h." + std::to_string(i) + ".attn.k_proj.weight"]   = layer.c_attn_k_proj_w;
+            model.tensors["transformer.h." + std::to_string(i) + ".attn.v_proj.weight"]   = layer.c_attn_v_proj_w;
+
+            model.tensors["transformer.h." + std::to_string(i) + ".attn.out_proj.weight"] = layer.c_attn_proj_w;
+
+            model.tensors["transformer.h." + std::to_string(i) + ".mlp.fc_in.weight"]     = layer.c_mlp_fc_w;
+            model.tensors["transformer.h." + std::to_string(i) + ".mlp.fc_in.bias"]       = layer.c_mlp_fc_b;
+
+            model.tensors["transformer.h." + std::to_string(i) + ".mlp.fc_out.weight"]    = layer.c_mlp_proj_w;
+            model.tensors["transformer.h." + std::to_string(i) + ".mlp.fc_out.bias"]      = layer.c_mlp_proj_b;
+        }
+    }
+
+    // key + value memory
+    {
+        const auto & hparams = model.hparams;
+
+        const int n_embd  = hparams.n_embd;
+        const int n_layer = hparams.n_layer;
+        const int n_ctx   = hparams.n_ctx;
+
+        const int n_mem      = n_layer*n_ctx;
+        const int n_elements = n_embd*n_mem;
+
+        model.memory_k = ggml_new_tensor_1d(ctx, GGML_TYPE_F16, n_elements);
+        model.memory_v = ggml_new_tensor_1d(ctx, GGML_TYPE_F16, n_elements);
+
+        const size_t memory_size = ggml_nbytes(model.memory_k) + ggml_nbytes(model.memory_v);
+
+        printf("%s: memory_size = %8.2f MB, n_mem = %d\n", __func__, memory_size/1024.0/1024.0, n_mem);
+    }
+
+    // load weights
+    {
+        int n_tensors = 0;
+        size_t total_size = 0;
+
+        printf("%s: ", __func__);
+
+        while (true) {
+            int32_t n_dims;
+            int32_t length;
+            int32_t ttype;
+
+            fin.read(reinterpret_cast<char *>(&n_dims), sizeof(n_dims));
+            fin.read(reinterpret_cast<char *>(&length), sizeof(length));
+            fin.read(reinterpret_cast<char *>(&ttype),  sizeof(ttype));
+
+            if (fin.eof()) {
+                break;
+            }
+
+            int32_t nelements = 1;
+            int32_t ne[2] = { 1, 1 };
+            for (int i = 0; i < n_dims; ++i) {
+                fin.read(reinterpret_cast<char *>(&ne[i]), sizeof(ne[i]));
+                nelements *= ne[i];
+            }
+
+            std::string name(length, 0);
+            fin.read(&name[0], length);
+
+            if (model.tensors.find(name) == model.tensors.end()) {
+                fprintf(stderr, "%s: unknown tensor '%s' in model file\n", __func__, name.c_str());
+                return false;
+            }
+
+            auto tensor = model.tensors[name];
+            if (ggml_nelements(tensor) != nelements) {
+                fprintf(stderr, "%s: tensor '%s' has wrong size in model file\n", __func__, name.c_str());
+                return false;
+            }
+
+            if (tensor->ne[0] != ne[0] || tensor->ne[1] != ne[1]) {
+                fprintf(stderr, "%s: tensor '%s' has wrong shape in model file: got [%d, %d], expected [%d, %d]\n",
+                        __func__, name.c_str(), (int) tensor->ne[0], (int) tensor->ne[1], ne[0], ne[1]);
+                return false;
+            }
+
+            // for debugging
+            if (0) {
+                printf("%24s - [%5d, %5d], type = %6s, %6.2f MB, %9zu bytes\n", name.c_str(), ne[0], ne[1], ggml_type_name(ggml_type(ttype)), ggml_nbytes(tensor)/1024.0/1024.0, ggml_nbytes(tensor));
+            }
+
+            const size_t bpe = ggml_type_size(ggml_type(ttype));
+
+            if ((nelements*bpe)/ggml_blck_size(tensor->type) != ggml_nbytes(tensor)) {
+                fprintf(stderr, "%s: tensor '%s' has wrong size in model file: got %zu, expected %zu\n",
+                        __func__, name.c_str(), ggml_nbytes(tensor), nelements*bpe);
+                return false;
+            }
+
+            fin.read(reinterpret_cast<char *>(tensor->data), ggml_nbytes(tensor));
+
+            //printf("%42s - [%5d, %5d], type = %6s, %6.2f MB\n", name.c_str(), ne[0], ne[1], ttype == 0 ? "float" : "f16", ggml_nbytes(tensor)/1024.0/1024.0);
+            total_size += ggml_nbytes(tensor);
+            if (++n_tensors % 8 == 0) {
+                printf(".");
+                fflush(stdout);
+            }
+        }
+
+        printf(" done\n");
+
+        printf("%s: model size = %8.2f MB / num tensors = %d\n", __func__, total_size/1024.0/1024.0, n_tensors);
+    }
+
+    fin.close();
+
+    return true;
+}
+
+// evaluate the transformer
+//
+//   - model:     the model
+//   - n_threads: number of threads to use
+//   - n_past:    the context size so far
+//   - embd_inp:  the embeddings of the tokens in the context
+//   - embd_w:    the predicted logits for the next token
+//
+// The GPT-J model requires about 16MB of memory per input token.
+//
+bool gptj_eval(
+        const gptj_model & model,
+        const int n_threads,
+        const int n_past,
+        const std::vector<gpt_vocab::id> & embd_inp,
+              std::vector<float>         & embd_w,
+              size_t                     & mem_per_token) {
+    const int N = embd_inp.size();
+
+    const auto & hparams = model.hparams;
+
+    const int n_embd  = hparams.n_embd;
+    const int n_layer = hparams.n_layer;
+    const int n_ctx   = hparams.n_ctx;
+    const int n_head  = hparams.n_head;
+    const int n_vocab = hparams.n_vocab;
+    const int n_rot   = hparams.n_rot;
+
+    static size_t buf_size = 256u*1024*1024;
+    static void * buf = malloc(buf_size);
+
+    if (mem_per_token > 0 && mem_per_token*N > buf_size) {
+        const size_t buf_size_new = 1.1*(mem_per_token*N); // add 10% to account for ggml object overhead
+        //printf("\n%s: reallocating buffer from %zu to %zu bytes\n", __func__, buf_size, buf_size_new);
+
+        // reallocate
+        buf_size = buf_size_new;
+        buf = realloc(buf, buf_size);
+        if (buf == nullptr) {
+            fprintf(stderr, "%s: failed to allocate %zu bytes\n", __func__, buf_size);
+            return false;
+        }
+    }
+
+    struct ggml_init_params params = {
+        /*.mem_size   =*/ buf_size,
+        /*.mem_buffer =*/ buf,
+        /*.no_alloc   =*/ false,
+    };
+
+    struct ggml_context * ctx0 = ggml_init(params);
+    struct ggml_cgraph gf = {};
+
+    struct ggml_tensor * embd = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
+    memcpy(embd->data, embd_inp.data(), N*ggml_element_size(embd));
+
+    // wte
+    struct ggml_tensor * inpL = ggml_get_rows(ctx0, model.wte, embd);
+
+    for (int il = 0; il < n_layer; ++il) {
+        struct ggml_tensor * cur;
+
+        // norm
+        {
+            cur = ggml_norm(ctx0, inpL, hparams.eps);
+
+            // cur = ln_1_g*cur + ln_1_b
+            cur = ggml_add(ctx0,
+                    ggml_mul(ctx0,
+                        ggml_repeat(ctx0, model.layers[il].ln_1_g, cur),
+                        cur),
+                    ggml_repeat(ctx0, model.layers[il].ln_1_b, cur));
+        }
+
+        struct ggml_tensor * inpSA = cur;
+
+        // self-attention
+        {
+            struct ggml_tensor * Qcur = ggml_rope_inplace(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model.layers[il].c_attn_q_proj_w, cur), n_embd/n_head, n_head, N), n_past, n_rot, 0, 0);
+            struct ggml_tensor * Kcur = ggml_rope_inplace(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model.layers[il].c_attn_k_proj_w, cur), n_embd/n_head, n_head, N), n_past, n_rot, 0, 0);
+
+            // store key and value to memory
+            {
+                struct ggml_tensor * Vcur = ggml_transpose(ctx0, ggml_mul_mat(ctx0, model.layers[il].c_attn_v_proj_w, cur));
+
+                struct ggml_tensor * k = ggml_view_1d(ctx0, model.memory_k, N*n_embd, (ggml_element_size(model.memory_k)*n_embd)*(il*n_ctx + n_past));
+                struct ggml_tensor * v = ggml_view_2d(ctx0, model.memory_v, N, n_embd,
+                        (   n_ctx)*ggml_element_size(model.memory_v),
+                        (il*n_ctx)*ggml_element_size(model.memory_v)*n_embd + n_past*ggml_element_size(model.memory_v));
+
+                ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Kcur, k));
+                ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Vcur, v));
+            }
+
+            // Q = Qcur.contiguous().view(n_embd/n_head, n_head, N).permute(0, 2, 1, 3)
+            struct ggml_tensor * Q =
+                ggml_permute(ctx0,
+                        Qcur,
+                        0, 2, 1, 3);
+
+            // K = Kmem.view(n_embd/n_head, n_head, n_past + N).permute(0, 2, 1, 3)
+            struct ggml_tensor * K =
+                ggml_permute(ctx0,
+                        ggml_reshape_3d(ctx0,
+                            ggml_view_1d(ctx0, model.memory_k, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(model.memory_k)*n_embd),
+                            n_embd/n_head, n_head, n_past + N),
+                        0, 2, 1, 3);
+
+            // K * Q
+            struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);
+
+            // KQ_scaled = KQ / sqrt(n_embd/n_head)
+            struct ggml_tensor * KQ_scaled =
+                ggml_scale_inplace(ctx0,
+                        KQ,
+                        ggml_new_f32(ctx0, 1.0f/sqrt(float(n_embd)/n_head))
+                        );
+
+            // KQ_masked = mask_past(KQ_scaled)
+            struct ggml_tensor * KQ_masked = ggml_diag_mask_inf_inplace(ctx0, KQ_scaled, n_past);
+
+            // KQ = soft_max(KQ_masked)
+            struct ggml_tensor * KQ_soft_max = ggml_soft_max_inplace(ctx0, KQ_masked);
+
+            // V_trans = Vmem.view(n_embd/n_head, n_head, n_past + N).permute(1, 2, 0, 3).contiguous()
+            struct ggml_tensor * V =
+                ggml_view_3d(ctx0, model.memory_v,
+                        n_past + N, n_embd/n_head, n_head,
+                        n_ctx*ggml_element_size(model.memory_v),
+                        n_ctx*ggml_element_size(model.memory_v)*n_embd/n_head,
+                        il*n_ctx*ggml_element_size(model.memory_v)*n_embd);
+
+            // KQV = transpose(V) * KQ_soft_max
+            struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max);
+
+            // KQV_merged = KQV.permute(0, 2, 1, 3)
+            struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
+
+            // cur = KQV_merged.contiguous().view(n_embd, N)
+            cur = ggml_cpy(ctx0,
+                    KQV_merged,
+                    ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, N));
+
+            // projection (no bias)
+            cur = ggml_mul_mat(ctx0,
+                    model.layers[il].c_attn_proj_w,
+                    cur);
+        }
+
+        struct ggml_tensor * inpFF = cur;
+
+        // feed-forward network
+        // this is independent of the self-attention result, so it could be done in parallel to the self-attention
+        {
+            // note here we pass inpSA instead of cur
+            cur = ggml_mul_mat(ctx0,
+                    model.layers[il].c_mlp_fc_w,
+                    inpSA);
+
+            cur = ggml_add(ctx0,
+                    ggml_repeat(ctx0, model.layers[il].c_mlp_fc_b, cur),
+                    cur);
+
+            // GELU activation
+            cur = ggml_gelu(ctx0, cur);
+
+            // projection
+            // cur = proj_w*cur + proj_b
+            cur = ggml_mul_mat(ctx0,
+                    model.layers[il].c_mlp_proj_w,
+                    cur);
+
+            cur = ggml_add(ctx0,
+                    ggml_repeat(ctx0, model.layers[il].c_mlp_proj_b, cur),
+                    cur);
+        }
+
+        // self-attention + FF
+        cur  = ggml_add(ctx0, cur, inpFF);
+
+        // input for next layer
+        inpL = ggml_add(ctx0, cur, inpL);
+    }
+
+    // norm
+    {
+        inpL = ggml_norm(ctx0, inpL, hparams.eps);
+
+        // inpL = ln_f_g*inpL + ln_f_b
+        inpL = ggml_add(ctx0,
+                ggml_mul(ctx0,
+                    ggml_repeat(ctx0, model.ln_f_g, inpL),
+                    inpL),
+                ggml_repeat(ctx0, model.ln_f_b, inpL));
+    }
+
+    // lm_head
+    {
+        inpL = ggml_mul_mat(ctx0, model.lmh_g, inpL);
+
+        inpL = ggml_add(ctx0,
+                ggml_repeat(ctx0, model.lmh_b, inpL),
+                inpL);
+    }
+
+    // logits -> probs
+    //inpL = ggml_soft_max_inplace(ctx0, inpL);
+
+    // run the computation
+    ggml_build_forward_expand(&gf, inpL);
+    ggml_graph_compute_with_ctx(ctx0, &gf, n_threads);
+
+    //if (n_past%100 == 0) {
+    //    ggml_graph_print   (&gf);
+    //    ggml_graph_dump_dot(&gf, NULL, "gpt-j.dot");
+    //}
+
+    //embd_w.resize(n_vocab*N);
+    //memcpy(embd_w.data(), ggml_get_data(inpL), sizeof(float)*n_vocab*N);
+
+    // return result for just the last token
+    embd_w.resize(n_vocab);
+    memcpy(embd_w.data(), (float *) ggml_get_data(inpL) + (n_vocab*(N-1)), sizeof(float)*n_vocab);
+
+    if (mem_per_token == 0) {
+        mem_per_token = ggml_used_mem(ctx0)/N;
+    }
+    //printf("used_mem = %zu\n", ggml_used_mem(ctx0));
+
+    ggml_free(ctx0);
+
+    return true;
+}
+
+int main(int argc, char ** argv) {
+    ggml_time_init();
+
+    const int64_t t_main_start_us = ggml_time_us();
+
+    gpt_params params;
+    params.model = "models/gpt-j-6B/ggml-model.bin";
+
+    if (gpt_params_parse(argc, argv, params) == false) {
+        return 1;
+    }
+
+    if (params.seed < 0) {
+        params.seed = time(NULL);
+    }
+
+    printf("%s: seed = %d\n", __func__, params.seed);
+
+    std::mt19937 rng(params.seed);
+    if (params.prompt.empty()) {
+        params.prompt = gpt_random_prompt(rng);
+    }
+
+    int64_t t_load_us = 0;
+
+    gpt_vocab vocab;
+    gptj_model model;
+
+    // load the model
+    {
+        const int64_t t_start_us = ggml_time_us();
+
+        if (!gptj_model_load(params.model, model, vocab)) {
+            fprintf(stderr, "%s: failed to load model from '%s'\n", __func__, params.model.c_str());
+            return 1;
+        }
+
+        t_load_us = ggml_time_us() - t_start_us;
+
+        test_gpt_tokenizer(vocab, params.token_test);
+    }
+
+    int n_past = 0;
+
+    int64_t t_sample_us  = 0;
+    int64_t t_predict_us = 0;
+
+    std::vector<float> logits;
+
+    // tokenize the prompt
+    std::vector<gpt_vocab::id> embd_inp = ::gpt_tokenize(vocab, params.prompt);
+
+    params.n_predict = std::min(params.n_predict, model.hparams.n_ctx - (int) embd_inp.size());
+
+    printf("%s: number of tokens in prompt = %zu\n", __func__, embd_inp.size());
+    printf("\n");
+
+    std::vector<gpt_vocab::id> embd;
+
+    // determine the required inference memory per token:
+    size_t mem_per_token = 0;
+    gptj_eval(model, params.n_threads, 0, { 0, 1, 2, 3 }, logits, mem_per_token);
+
+    for (size_t i = embd.size(); i < embd_inp.size() + params.n_predict; i++) {
+        // predict
+        if (embd.size() > 0) {
+            const int64_t t_start_us = ggml_time_us();
+
+            if (!gptj_eval(model, params.n_threads, n_past, embd, logits, mem_per_token)) {
+                printf("Failed to predict\n");
+                return 1;
+            }
+
+            t_predict_us += ggml_time_us() - t_start_us;
+        }
+
+        n_past += embd.size();
+        embd.clear();
+
+        if (i >= embd_inp.size()) {
+            // sample next token
+            const int   top_k = params.top_k;
+            const float top_p = params.top_p;
+            const float temp  = params.temp;
+
+            const int n_vocab = model.hparams.n_vocab;
+
+            gpt_vocab::id id = 0;
+
+            {
+                const int64_t t_start_sample_us = ggml_time_us();
+
+                id = gpt_sample_top_k_top_p(vocab, logits.data() + (logits.size() - n_vocab), top_k, top_p, temp, rng);
+
+                t_sample_us += ggml_time_us() - t_start_sample_us;
+            }
+
+            // add it to the context
+            embd.push_back(id);
+        } else {
+            // if here, it means we are still processing the input prompt
+            for (size_t k = i; k < embd_inp.size(); k++) {
+                embd.push_back(embd_inp[k]);
+                if (int32_t(embd.size()) > params.n_batch) {
+                    break;
+                }
+            }
+            i += embd.size() - 1;
+        }
+
+        // display text
+        for (auto id : embd) {
+            printf("%s", vocab.id_to_token[id].c_str());
+        }
+        fflush(stdout);
+
+        // end of text token
+        if (embd.back() == 50256) {
+            break;
+        }
+    }
+
+    // report timing
+    {
+        const int64_t t_main_end_us = ggml_time_us();
+
+        printf("\n\n");
+        printf("%s: mem per token = %8zu bytes\n", __func__, mem_per_token);
+        printf("%s:     load time = %8.2f ms\n", __func__, t_load_us/1000.0f);
+        printf("%s:   sample time = %8.2f ms\n", __func__, t_sample_us/1000.0f);
+        printf("%s:  predict time = %8.2f ms / %.2f ms per token\n", __func__, t_predict_us/1000.0f, t_predict_us/1000.0f/n_past);
+        printf("%s:    total time = %8.2f ms\n", __func__, (t_main_end_us - t_main_start_us)/1000.0f);
+    }
+
+    ggml_free(model.ctx);
+
+    return 0;
+}

ggml/examples/gpt-j/quantize.cpp

@@ -0,0 +1,182 @@
+#include "ggml/ggml.h"
+
+#include "common.h"
+#include "common-ggml.h"
+
+#include <cassert>
+#include <cmath>
+#include <cstdio>
+#include <cstring>
+#include <fstream>
+#include <map>
+#include <string>
+#include <vector>
+#include <regex>
+
+// default hparams (GPT-J 6B)
+struct gptj_hparams {
+    int32_t n_vocab = 50400;
+    int32_t n_ctx   = 2048;
+    int32_t n_embd  = 4096;
+    int32_t n_head  = 16;
+    int32_t n_layer = 28;
+    int32_t n_rot   = 64;
+    int32_t ftype   = 1;
+};
+
+// quantize a model
+bool gptj_model_quantize(const std::string & fname_inp, const std::string & fname_out, ggml_ftype ftype) {
+    gpt_vocab vocab;
+
+    printf("%s: loading model from '%s'\n", __func__, fname_inp.c_str());
+
+    auto finp = std::ifstream(fname_inp, std::ios::binary);
+    if (!finp) {
+        fprintf(stderr, "%s: failed to open '%s' for reading\n", __func__, fname_inp.c_str());
+        return false;
+    }
+
+    auto fout = std::ofstream(fname_out, std::ios::binary);
+    if (!fout) {
+        fprintf(stderr, "%s: failed to open '%s' for writing\n", __func__, fname_out.c_str());
+        return false;
+    }
+
+    // verify magic
+    {
+        uint32_t magic;
+        finp.read((char *) &magic, sizeof(magic));
+        if (magic != GGML_FILE_MAGIC) {
+            fprintf(stderr, "%s: invalid model file '%s' (bad magic)\n", __func__, fname_inp.c_str());
+            return false;
+        }
+
+        fout.write((char *) &magic, sizeof(magic));
+    }
+
+    gptj_hparams hparams;
+
+    // load hparams
+    {
+        finp.read((char *) &hparams.n_vocab, sizeof(hparams.n_vocab));
+        finp.read((char *) &hparams.n_ctx,   sizeof(hparams.n_ctx));
+        finp.read((char *) &hparams.n_embd,  sizeof(hparams.n_embd));
+        finp.read((char *) &hparams.n_head,  sizeof(hparams.n_head));
+        finp.read((char *) &hparams.n_layer, sizeof(hparams.n_layer));
+        finp.read((char *) &hparams.n_rot,   sizeof(hparams.n_rot));
+        finp.read((char *) &hparams.ftype,   sizeof(hparams.ftype));
+
+        const int32_t qntvr_src =    hparams.ftype / GGML_QNT_VERSION_FACTOR;
+        const int32_t ftype_dst = GGML_QNT_VERSION * GGML_QNT_VERSION_FACTOR + ftype;
+
+        printf("%s: n_vocab     = %d\n", __func__, hparams.n_vocab);
+        printf("%s: n_ctx       = %d\n", __func__, hparams.n_ctx);
+        printf("%s: n_embd      = %d\n", __func__, hparams.n_embd);
+        printf("%s: n_head      = %d\n", __func__, hparams.n_head);
+        printf("%s: n_layer     = %d\n", __func__, hparams.n_layer);
+        printf("%s: ftype (src) = %d\n", __func__, hparams.ftype);
+        printf("%s: qntvr (src) = %d\n", __func__, qntvr_src);
+        printf("%s: ftype (dst) = %d\n", __func__, ftype_dst);
+        printf("%s: qntvr (dst) = %d\n", __func__, GGML_QNT_VERSION);
+
+        fout.write((char *) &hparams.n_vocab, sizeof(hparams.n_vocab));
+        fout.write((char *) &hparams.n_ctx,   sizeof(hparams.n_ctx));
+        fout.write((char *) &hparams.n_embd,  sizeof(hparams.n_embd));
+        fout.write((char *) &hparams.n_head,  sizeof(hparams.n_head));
+        fout.write((char *) &hparams.n_layer, sizeof(hparams.n_layer));
+        fout.write((char *) &hparams.n_rot,   sizeof(hparams.n_rot));
+        fout.write((char *) &ftype_dst,       sizeof(ftype_dst));
+    }
+
+    // load vocab
+    {
+        int32_t n_vocab = 0;
+        finp.read ((char *) &n_vocab, sizeof(n_vocab));
+        fout.write((char *) &n_vocab, sizeof(n_vocab));
+
+        if (n_vocab != hparams.n_vocab) {
+            fprintf(stderr, "%s: invalid model file '%s' (bad vocab size %d != %d)\n",
+                    __func__, fname_inp.c_str(), n_vocab, hparams.n_vocab);
+            return false;
+        }
+
+        std::string word;
+        for (int i = 0; i < n_vocab; i++) {
+            uint32_t len;
+            finp.read ((char *) &len, sizeof(len));
+            fout.write((char *) &len, sizeof(len));
+
+            word.resize(len);
+            finp.read ((char *) word.data(), len);
+            fout.write((char *) word.data(), len);
+
+            vocab.token_to_id[word] = i;
+            vocab.id_to_token[i] = word;
+        }
+    }
+
+    // regexes of tensor names to be quantized
+    const std::vector<std::string> to_quant = {
+        ".*weight",
+    };
+
+    if (!ggml_common_quantize_0(finp, fout, ftype, to_quant, {})) {
+        fprintf(stderr, "%s: failed to quantize model '%s'\n", __func__, fname_inp.c_str());
+        return false;
+    }
+
+    finp.close();
+    fout.close();
+
+    return true;
+}
+
+// usage:
+//  ./gpt-2-quantize models/gpt-2-117M/ggml-model.bin models/gpt-2-117M/ggml-model-quant.bin type
+//
+int main(int argc, char ** argv) {
+    if (argc != 4) {
+        fprintf(stderr, "usage: %s model-f32.bin model-quant.bin type\n", argv[0]);
+        ggml_print_ftypes(stderr);
+        return 1;
+    }
+
+    // needed to initialize f16 tables
+    {
+        struct ggml_init_params params = { 0, NULL, false };
+        struct ggml_context * ctx = ggml_init(params);
+        ggml_free(ctx);
+    }
+
+    const std::string fname_inp = argv[1];
+    const std::string fname_out = argv[2];
+
+    const ggml_ftype ftype = ggml_parse_ftype(argv[3]);
+
+    const int64_t t_main_start_us = ggml_time_us();
+
+    int64_t t_quantize_us = 0;
+
+    // load the model
+    {
+        const int64_t t_start_us = ggml_time_us();
+
+        if (!gptj_model_quantize(fname_inp, fname_out, ggml_ftype(ftype))) {
+            fprintf(stderr, "%s: failed to quantize model from '%s'\n", __func__, fname_inp.c_str());
+            return 1;
+        }
+
+        t_quantize_us = ggml_time_us() - t_start_us;
+    }
+
+    // report timing
+    {
+        const int64_t t_main_end_us = ggml_time_us();
+
+        printf("\n");
+        printf("%s: quantize time = %8.2f ms\n", __func__, t_quantize_us/1000.0f);
+        printf("%s:    total time = %8.2f ms\n", __func__, (t_main_end_us - t_main_start_us)/1000.0f);
+    }
+
+    return 0;
+}

ggml/examples/gpt-neox/CMakeLists.txt

@@ -0,0 +1,13 @@
+#
+# gpt-neox
+
+set(TEST_TARGET gpt-neox)
+add_executable(${TEST_TARGET} main.cpp)
+target_link_libraries(${TEST_TARGET} PRIVATE ggml common common-ggml)
+
+#
+# gpt-neox-quantize
+
+set(TEST_TARGET gpt-neox-quantize)
+add_executable(${TEST_TARGET} quantize.cpp)
+target_link_libraries(${TEST_TARGET} PRIVATE ggml common common-ggml)

ggml/examples/gpt-neox/README.md

@@ -0,0 +1,110 @@
+# GPT-NeoX
+
+Transformer architecture: GPT-NeoX
+
+Ref: https://github.com/stability-AI/stableLM/#stablelm-alpha
+
+## Usage
+
+```bash
+# get the repo and build it
+git clone https://github.com/ggerganov/ggml
+cd ggml
+mkdir build && cd build
+cmake ..
+make -j
+
+# get the StableLM 3B Alpha model
+git clone https://huggingface.co/stabilityai/gpt_neox-base-alpha-3b
+
+# install Python dependencies
+python3 -m pip install -r ../requirements.txt
+
+# convert model to FP16
+python3 ../examples/gpt-neox/convert-h5-to-ggml.py ./stablelm-base-alpha-3b/ 1
+
+# run inference using FP16 precision
+make -j && ./bin/gpt-neox -m ./stablelm-base-alpha-3b/ggml-model-f16.bin -p "I believe the meaning of life is" -t 8 -n 64
+
+main: seed = 1681940611
+gpt_neox_model_load: loading model from 'models/stablelm-base-alpha-3b/ggml-model-f16.bin' - please wait ...
+gpt_neox_model_load: n_vocab = 50688
+gpt_neox_model_load: n_ctx   = 4096
+gpt_neox_model_load: n_embd  = 4096
+gpt_neox_model_load: n_head  = 32
+gpt_neox_model_load: n_layer = 16
+gpt_neox_model_load: n_rot   = 32
+gpt_neox_model_load: ftype   = 1
+gpt_neox_model_load: ggml ctx size = 10011.10 MB
+gpt_neox_model_load: memory_size =  2048.00 MB, n_mem = 65536
+gpt_neox_model_load: ................................ done
+gpt_neox_model_load: model size =  6939.28 MB / num tensors = 260
+main: number of tokens in prompt = 7
+main: token[0] =     42, I
+main: token[1] =   2868,  believe
+main: token[2] =    253,  the
+main: token[3] =   4495,  meaning
+main: token[4] =    273,  of
+main: token[5] =   1495,  life
+main: token[6] =    310,  is
+
+I believe the meaning of life is to grow, to find a way, to love, to find an appreciation for life, and to live it with all of its beauty.
+
+For I am the child of God. I am the offspring of God's love. I am the offspring of the light of the world. I am the offspring of the
+
+main: mem per token = 12186760 bytes
+main:     load time =  2118.55 ms
+main:   sample time =     9.59 ms
+main:  predict time =  4474.07 ms / 63.92 ms per token
+main:    total time =  6911.26 ms
+```
+
+## 5-bit integer quantization mode
+
+```bash
+# quantize the model to 5-bits using Q5_0 quantization
+./bin/gpt-neox-quantize ./stablelm-base-alpha-3b/ggml-model-f16.bin ./stablelm-base-alpha-3b/ggml-model-q5_0.bin q5_0
+
+# run the quantized model
+./bin/gpt-neox -m ./stablelm-base-alpha-3b/ggml-model-q5_0.bin -p "I believe the meaning of life is" -t 8 -n 64
+
+main: seed = 1682021489
+gpt_neox_model_load: loading model from 'models/stablelm-base-alpha-3b/ggml-model-q5_0.bin' - please wait ...
+gpt_neox_model_load: n_vocab = 50688
+gpt_neox_model_load: n_ctx   = 4096
+gpt_neox_model_load: n_embd  = 4096
+gpt_neox_model_load: n_head  = 32
+gpt_neox_model_load: n_layer = 16
+gpt_neox_model_load: n_rot   = 32
+gpt_neox_model_load: ftype   = 6
+gpt_neox_model_load: ggml ctx size = 5676.10 MB
+gpt_neox_model_load: memory_size =  1024.00 MB, n_mem = 65536
+gpt_neox_model_load: ........................ done
+gpt_neox_model_load: model size =  2604.28 MB / num tensors = 196
+main: number of tokens in prompt = 7
+main: token[0] =     42, I
+main: token[1] =   2868,  believe
+main: token[2] =    253,  the
+main: token[3] =   4495,  meaning
+main: token[4] =    273,  of
+main: token[5] =   1495,  life
+main: token[6] =    310,  is
+
+I believe the meaning of life is to love and be loved. The last three verses were enough to tie us all together. If you love someone you love them all. There are some things in this world that are just not equal in Heaven. - Be here in this moment.
+
+This world is not what is outside of us. It is what
+
+main: mem per token = 12958024 bytes
+main:     load time =   850.51 ms
+main:   sample time =     9.95 ms
+main:  predict time =  3103.81 ms / 44.34 ms per token
+main:    total time =  4177.68 ms
+
+```
+
+## Notes
+
+- No guarantees for correctness
+- The tokenizer is currently hacked - probably works only for English
+- Non-parallel residual is not supported
+- Contributions and improvements are welcome

ggml/examples/gpt-neox/convert-h5-to-ggml.py

@@ -0,0 +1,107 @@
+import sys
+import struct
+import json
+import numpy as np
+
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+if len(sys.argv) < 3:
+    print("Usage: convert-h5-to-ggml.py dir-model [use-f32]\n")
+    print("  ftype == 0 -> float32")
+    print("  ftype == 1 -> float16")
+    sys.exit(1)
+
+# output in the same directory as the model
+dir_model = sys.argv[1]
+fname_out = sys.argv[1] + "/ggml-model.bin"
+
+with open(dir_model + "/config.json", "r", encoding="utf-8") as f:
+    hparams = json.load(f)
+
+# possible data types
+#   ftype == 0 -> float32
+#   ftype == 1 -> float16
+#
+# map from ftype to string
+ftype_str = ["f32", "f16"]
+
+ftype = 1
+if len(sys.argv) > 2:
+    ftype = int(sys.argv[2])
+    if ftype < 0 or ftype > 1:
+        print("Invalid ftype: " + str(ftype))
+        sys.exit(1)
+    fname_out = sys.argv[1] + "/ggml-model-" + ftype_str[ftype] + ".bin"
+
+
+tokenizer = AutoTokenizer.from_pretrained(dir_model)
+model = AutoModelForCausalLM.from_pretrained(dir_model, low_cpu_mem_usage=True)
+
+list_vars = model.state_dict()
+for name in list_vars.keys():
+    print(name, list_vars[name].shape, list_vars[name].dtype)
+
+fout = open(fname_out, "wb")
+
+print(hparams)
+
+fout.write(struct.pack("i", 0x67676d6c)) # magic: ggml in hex
+fout.write(struct.pack("i", hparams["vocab_size"]))
+fout.write(struct.pack("i", hparams["max_position_embeddings"]))
+fout.write(struct.pack("i", hparams["hidden_size"]))
+fout.write(struct.pack("i", hparams["num_attention_heads"]))
+fout.write(struct.pack("i", hparams["num_hidden_layers"]))
+fout.write(struct.pack("i", int(hparams["rotary_pct"]*(hparams["hidden_size"]//hparams["num_attention_heads"]))))
+fout.write(struct.pack("i", hparams["use_parallel_residual"] if "use_parallel_residual" in hparams else True))
+fout.write(struct.pack("i", ftype))
+
+# TODO: temporary hack to not deal with implementing the tokenizer
+for i in range(hparams["vocab_size"]):
+    text = tokenizer.decode([i]).encode('utf-8')
+    fout.write(struct.pack("i", len(text)))
+    fout.write(text)
+
+for name in list_vars.keys():
+    data = list_vars[name].squeeze().numpy()
+    print("Processing variable: " + name + " with shape: ", data.shape)
+
+    # we don't need these
+    if name.endswith(".attention.masked_bias") or     \
+       name.endswith(".attention.bias") or \
+       name.endswith(".attention.rotary_emb.inv_freq"):
+        print("  Skipping variable: " + name)
+        continue
+
+    n_dims = len(data.shape)
+
+    # ftype == 0 -> float32, ftype == 1 -> float16
+    ftype_cur = 0
+    if ftype != 0:
+        if name[-7:] == ".weight" and n_dims == 2:
+            print("  Converting to float16")
+            data = data.astype(np.float16)
+            ftype_cur = 1
+        else:
+            print("  Converting to float32")
+            data = data.astype(np.float32)
+            ftype_cur = 0
+    else:
+        if data.dtype != np.float32:
+            print("  Converting to float32")
+            data = data.astype(np.float32)
+            ftype_cur = 0
+
+    # header
+    str = name.encode('utf-8')
+    fout.write(struct.pack("iii", n_dims, len(str), ftype_cur))
+    for i in range(n_dims):
+        fout.write(struct.pack("i", data.shape[n_dims - 1 - i]))
+    fout.write(str)
+
+    # data
+    data.tofile(fout)
+
+fout.close()
+
+print("Done. Output file: " + fname_out)
+print("")

ggml/examples/gpt-neox/main.cpp

@@ -0,0 +1,814 @@
+#include "ggml/ggml.h"
+
+#include "common.h"
+#include "common-ggml.h"
+
+#include <cassert>
+#include <cmath>
+#include <cstdio>
+#include <cstring>
+#include <cinttypes>
+#include <fstream>
+#include <map>
+#include <string>
+#include <vector>
+
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
+// default hparams (StableLM 3B)
+struct gpt_neox_hparams {
+    int32_t n_vocab = 50257;
+    int32_t n_ctx   = 4096;
+    int32_t n_embd  = 4096;
+    int32_t n_head  = 32;
+    int32_t n_layer = 16;
+    int32_t n_rot   = 32; // rotary_pct * (n_embd / n_head)
+    int32_t par_res = 1; // 1 = true, 0 = false
+    int32_t ftype   = 1;
+    float   eps     = 1e-5f;
+};
+
+struct gpt_neox_layer {
+    // pre normalization
+    struct ggml_tensor * ln_1_g;
+    struct ggml_tensor * ln_1_b;
+
+    // attention
+    struct ggml_tensor * c_attn_attn_w;
+    struct ggml_tensor * c_attn_attn_b;
+
+    struct ggml_tensor * c_attn_proj_w;
+    struct ggml_tensor * c_attn_proj_b;
+
+    // post normalization
+    struct ggml_tensor * ln_2_g;
+    struct ggml_tensor * ln_2_b;
+
+    // ff
+    struct ggml_tensor * c_mlp_fc_w;
+    struct ggml_tensor * c_mlp_fc_b;
+
+    struct ggml_tensor * c_mlp_proj_w;
+    struct ggml_tensor * c_mlp_proj_b;
+};
+
+struct gpt_neox_model {
+    gpt_neox_hparams hparams;
+
+    // normalization
+    struct ggml_tensor * ln_f_g;
+    struct ggml_tensor * ln_f_b;
+
+    struct ggml_tensor * wte; // position embedding
+
+    struct ggml_tensor * lmh_g; // language model head
+    //struct ggml_tensor * lmh_b; // language model bias
+
+    std::vector<gpt_neox_layer> layers;
+
+    // key + value memory
+    struct ggml_tensor * memory_k;
+    struct ggml_tensor * memory_v;
+
+    //
+    struct ggml_context * ctx;
+    std::map<std::string, struct ggml_tensor *> tensors;
+};
+
+// load the model's weights from a file
+bool gpt_neox_model_load(const std::string & fname, gpt_neox_model & model, gpt_vocab & vocab) {
+    printf("%s: loading model from '%s' - please wait ...\n", __func__, fname.c_str());
+
+    auto fin = std::ifstream(fname, std::ios::binary);
+    if (!fin) {
+        fprintf(stderr, "%s: failed to open '%s'\n", __func__, fname.c_str());
+        return false;
+    }
+
+    // verify magic
+    {
+        uint32_t magic;
+        fin.read((char *) &magic, sizeof(magic));
+        if (magic != GGML_FILE_MAGIC) {
+            fprintf(stderr, "%s: invalid model file '%s' (bad magic)\n", __func__, fname.c_str());
+            return false;
+        }
+    }
+
+    // load hparams
+    {
+        auto & hparams = model.hparams;
+
+        fin.read((char *) &hparams.n_vocab, sizeof(hparams.n_vocab));
+        fin.read((char *) &hparams.n_ctx,   sizeof(hparams.n_ctx));
+        fin.read((char *) &hparams.n_embd,  sizeof(hparams.n_embd));
+        fin.read((char *) &hparams.n_head,  sizeof(hparams.n_head));
+        fin.read((char *) &hparams.n_layer, sizeof(hparams.n_layer));
+        fin.read((char *) &hparams.n_rot,   sizeof(hparams.n_rot));
+        fin.read((char *) &hparams.par_res, sizeof(hparams.par_res));
+        fin.read((char *) &hparams.ftype,   sizeof(hparams.ftype));
+
+        const int32_t qntvr = hparams.ftype / GGML_QNT_VERSION_FACTOR;
+
+        printf("%s: n_vocab = %d\n", __func__, hparams.n_vocab);
+        printf("%s: n_ctx   = %d\n", __func__, hparams.n_ctx);
+        printf("%s: n_embd  = %d\n", __func__, hparams.n_embd);
+        printf("%s: n_head  = %d\n", __func__, hparams.n_head);
+        printf("%s: n_layer = %d\n", __func__, hparams.n_layer);
+        printf("%s: n_rot   = %d\n", __func__, hparams.n_rot);
+        printf("%s: par_res = %d\n", __func__, hparams.par_res);
+        printf("%s: ftype   = %d\n", __func__, hparams.ftype);
+        printf("%s: qntvr   = %d\n", __func__, qntvr);
+
+        hparams.ftype %= GGML_QNT_VERSION_FACTOR;
+    }
+
+    // load vocab
+    {
+        const int32_t n_vocab = model.hparams.n_vocab;
+
+        std::string word;
+        std::vector<char> buf(128);
+
+        for (int i = 0; i < n_vocab; i++) {
+            uint32_t len;
+            fin.read((char *) &len, sizeof(len));
+
+            buf.resize(len);
+            fin.read((char *) buf.data(), len);
+            word.assign(buf.data(), len);
+
+            vocab.token_to_id[word] = i;
+            vocab.id_to_token[i] = word;
+        }
+    }
+
+    // for the big tensors, we have the option to store the data in 16-bit floats or quantized
+    // in order to save memory and also to speed up the computation
+    ggml_type wtype = ggml_ftype_to_ggml_type((ggml_ftype) (model.hparams.ftype));
+    if (wtype == GGML_TYPE_COUNT) {
+        fprintf(stderr, "%s: invalid model file '%s' (bad ftype value %d)\n",
+                __func__, fname.c_str(), model.hparams.ftype);
+        return false;
+    }
+
+    auto & ctx = model.ctx;
+
+    size_t ctx_size = 0;
+
+    {
+        const auto & hparams = model.hparams;
+
+        const size_t n_embd  = hparams.n_embd;
+        const size_t n_layer = hparams.n_layer;
+        const size_t n_ctx   = hparams.n_ctx;
+        const size_t n_vocab = hparams.n_vocab;
+
+        ctx_size += n_embd*ggml_type_sizef(GGML_TYPE_F32); // ln_f_g
+        ctx_size += n_embd*ggml_type_sizef(GGML_TYPE_F32); // ln_f_b
+
+        ctx_size += n_embd*n_vocab*ggml_type_sizef(wtype); // wte
+
+        ctx_size += n_embd*n_vocab*ggml_type_sizef(wtype);           // lmh_g
+        //ctx_size +=        n_vocab*ggml_type_sizef(GGML_TYPE_F32); // lmh_b
+
+        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_1_g
+        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_1_b
+
+        ctx_size += n_layer*(3*n_embd*n_embd*ggml_type_sizef(wtype));         // c_attn_attn_w
+        ctx_size += n_layer*(       3*n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_attn_attn_b
+
+        ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype));         // c_attn_proj_w
+        ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_attn_proj_b
+
+        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_2_g
+        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_2_b
+
+        ctx_size += n_layer*(4*n_embd*n_embd*ggml_type_sizef(wtype));         // c_mlp_fc_w
+        ctx_size += n_layer*(       4*n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_mlp_fc_b
+
+        ctx_size += n_layer*(4*n_embd*n_embd*ggml_type_sizef(wtype));         // c_mlp_proj_w
+        ctx_size += n_layer*(         n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_mlp_proj_b
+
+        ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(GGML_TYPE_F32); // memory_k
+        ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(GGML_TYPE_F32); // memory_v
+
+        ctx_size += (6 + 16*n_layer)*1024; // object overhead
+
+        printf("%s: ggml ctx size = %6.2f MB\n", __func__, ctx_size/(1024.0*1024.0));
+    }
+
+    // create the ggml context
+    {
+        struct ggml_init_params params = {
+            /*.mem_size   =*/ ctx_size,
+            /*.mem_buffer =*/ NULL,
+            /*.no_alloc   =*/ false,
+        };
+
+        model.ctx = ggml_init(params);
+        if (!model.ctx) {
+            fprintf(stderr, "%s: ggml_init() failed\n", __func__);
+            return false;
+        }
+    }
+
+    // prepare memory for the weights
+    {
+        const auto & hparams = model.hparams;
+
+        const int n_embd  = hparams.n_embd;
+        const int n_layer = hparams.n_layer;
+        const int n_vocab = hparams.n_vocab;
+
+        model.layers.resize(n_layer);
+
+        model.wte    = ggml_new_tensor_2d(ctx, wtype,         n_embd, n_vocab);
+
+        model.ln_f_g = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
+        model.ln_f_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
+
+        model.lmh_g  = ggml_new_tensor_2d(ctx, wtype,         n_embd, n_vocab);
+        //model.lmh_b  = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_vocab);
+
+        // map by name
+        model.tensors["gpt_neox.embed_in.weight"] = model.wte;
+
+        model.tensors["gpt_neox.final_layer_norm.weight"] = model.ln_f_g;
+        model.tensors["gpt_neox.final_layer_norm.bias"]   = model.ln_f_b;
+
+        model.tensors["embed_out.weight"] = model.lmh_g;
+        //model.tensors["lm_head.bias"]   = model.lmh_b;
+
+        for (int i = 0; i < n_layer; ++i) {
+            auto & layer = model.layers[i];
+
+            layer.ln_1_g          = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+            layer.ln_1_b          = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+
+            layer.c_attn_attn_w   = ggml_new_tensor_2d(ctx, wtype,           n_embd, 3*n_embd);
+            layer.c_attn_attn_b   = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 3*n_embd);
+
+            layer.c_attn_proj_w   = ggml_new_tensor_2d(ctx, wtype,           n_embd,   n_embd);
+            layer.c_attn_proj_b   = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+
+            layer.ln_2_g          = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+            layer.ln_2_b          = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+
+            layer.c_mlp_fc_w      = ggml_new_tensor_2d(ctx, wtype,           n_embd, 4*n_embd);
+            layer.c_mlp_fc_b      = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 4*n_embd);
+
+            layer.c_mlp_proj_w    = ggml_new_tensor_2d(ctx, wtype,         4*n_embd,   n_embd);
+            layer.c_mlp_proj_b    = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+
+            // map by name
+            model.tensors["gpt_neox.layers." + std::to_string(i) + ".input_layernorm.weight"] = layer.ln_1_g;
+            model.tensors["gpt_neox.layers." + std::to_string(i) + ".input_layernorm.bias"]   = layer.ln_1_b;
+
+            model.tensors["gpt_neox.layers." + std::to_string(i) + ".attention.query_key_value.weight"] = layer.c_attn_attn_w;
+            model.tensors["gpt_neox.layers." + std::to_string(i) + ".attention.query_key_value.bias"]   = layer.c_attn_attn_b;
+
+            model.tensors["gpt_neox.layers." + std::to_string(i) + ".attention.dense.weight"] = layer.c_attn_proj_w;
+            model.tensors["gpt_neox.layers." + std::to_string(i) + ".attention.dense.bias"]   = layer.c_attn_proj_b;
+
+            model.tensors["gpt_neox.layers." + std::to_string(i) + ".post_attention_layernorm.weight"] = layer.ln_2_g;
+            model.tensors["gpt_neox.layers." + std::to_string(i) + ".post_attention_layernorm.bias"]   = layer.ln_2_b;
+
+            model.tensors["gpt_neox.layers." + std::to_string(i) + ".mlp.dense_h_to_4h.weight"] = layer.c_mlp_fc_w;
+            model.tensors["gpt_neox.layers." + std::to_string(i) + ".mlp.dense_h_to_4h.bias"]   = layer.c_mlp_fc_b;
+
+            model.tensors["gpt_neox.layers." + std::to_string(i) + ".mlp.dense_4h_to_h.weight"] = layer.c_mlp_proj_w;
+            model.tensors["gpt_neox.layers." + std::to_string(i) + ".mlp.dense_4h_to_h.bias"]   = layer.c_mlp_proj_b;
+        }
+    }
+
+    // key + value memory
+    {
+        const auto & hparams = model.hparams;
+
+        const int n_embd  = hparams.n_embd;
+        const int n_layer = hparams.n_layer;
+        const int n_ctx   = hparams.n_ctx;
+
+        const int64_t n_mem      = n_layer*n_ctx;
+        const int64_t n_elements = n_embd*n_mem;
+
+        model.memory_k = ggml_new_tensor_1d(ctx, GGML_TYPE_F16, n_elements);
+        model.memory_v = ggml_new_tensor_1d(ctx, GGML_TYPE_F16, n_elements);
+
+        const size_t memory_size = ggml_nbytes(model.memory_k) + ggml_nbytes(model.memory_v);
+
+        printf("%s: memory_size = %8.2f MB, n_mem = %" PRId64 "\n", __func__, memory_size/1024.0/1024.0, n_mem);
+    }
+
+    // load weights
+    {
+        int n_tensors = 0;
+        size_t total_size = 0;
+
+        printf("%s: ", __func__);
+
+        while (true) {
+            int32_t n_dims;
+            int32_t length;
+            int32_t ttype;
+
+            fin.read(reinterpret_cast<char *>(&n_dims), sizeof(n_dims));
+            fin.read(reinterpret_cast<char *>(&length), sizeof(length));
+            fin.read(reinterpret_cast<char *>(&ttype),  sizeof(ttype));
+
+            if (fin.eof()) {
+                break;
+            }
+
+            int32_t nelements = 1;
+            int32_t ne[2] = { 1, 1 };
+            for (int i = 0; i < n_dims; ++i) {
+                fin.read(reinterpret_cast<char *>(&ne[i]), sizeof(ne[i]));
+                nelements *= ne[i];
+            }
+
+            std::string name(length, 0);
+            fin.read(&name[0], length);
+
+            if (model.tensors.find(name) == model.tensors.end()) {
+                fprintf(stderr, "%s: unknown tensor '%s' in model file\n", __func__, name.c_str());
+                return false;
+            }
+
+            auto tensor = model.tensors[name];
+            if (ggml_nelements(tensor) != nelements) {
+                fprintf(stderr, "%s: tensor '%s' has wrong size in model file\n", __func__, name.c_str());
+                return false;
+            }
+
+            if (tensor->ne[0] != ne[0] || tensor->ne[1] != ne[1]) {
+                fprintf(stderr, "%s: tensor '%s' has wrong shape in model file: got [%5d, %5d], expected [%5d, %5d]\n",
+                        __func__, name.c_str(), (int) tensor->ne[0], (int) tensor->ne[1], ne[0], ne[1]);
+                return false;
+            }
+
+            // for debugging
+            if (0) {
+                printf("%24s - [%5d, %5d], type = %6s, %6.2f MB, %9zu bytes\n", name.c_str(), ne[0], ne[1], ggml_type_name(ggml_type(ttype)), ggml_nbytes(tensor)/1024.0/1024.0, ggml_nbytes(tensor));
+            }
+
+            const size_t bpe = ggml_type_size(ggml_type(ttype));
+
+            if ((nelements*bpe)/ggml_blck_size(tensor->type) != ggml_nbytes(tensor)) {
+                fprintf(stderr, "%s: tensor '%s' has wrong size in model file: got %zu, expected %zu\n",
+                        __func__, name.c_str(), ggml_nbytes(tensor), nelements*bpe);
+                return false;
+            }
+
+            fin.read(reinterpret_cast<char *>(tensor->data), ggml_nbytes(tensor));
+
+            total_size += ggml_nbytes(tensor);
+            if (++n_tensors % 8 == 0) {
+                printf(".");
+                fflush(stdout);
+            }
+        }
+
+        printf(" done\n");
+
+        printf("%s: model size = %8.2f MB / num tensors = %d\n", __func__, total_size/1024.0/1024.0, n_tensors);
+    }
+
+    fin.close();
+
+    return true;
+}
+
+
+// feed-forward network
+ggml_tensor * gpt_neox_ff(
+        const gpt_neox_layer & layer,
+        ggml_context         * ctx0,
+        ggml_tensor          * inp,
+        float                  eps) {
+    ggml_tensor * cur = ggml_norm(ctx0, inp, eps);
+
+    cur = ggml_add(ctx0,
+        ggml_mul(ctx0,
+            ggml_repeat(ctx0, layer.ln_2_g, cur),
+            cur),
+        ggml_repeat(ctx0, layer.ln_2_b, cur));
+
+    cur = ggml_mul_mat(ctx0,
+            layer.c_mlp_fc_w,
+            cur);
+
+    cur = ggml_add(ctx0,
+            ggml_repeat(ctx0, layer.c_mlp_fc_b, cur),
+            cur);
+
+    // GELU activation
+    cur = ggml_gelu(ctx0, cur);
+
+    // projection
+    // cur = proj_w*cur + proj_b
+    cur = ggml_mul_mat(ctx0,
+            layer.c_mlp_proj_w,
+            cur);
+
+    cur = ggml_add(ctx0,
+            ggml_repeat(ctx0, layer.c_mlp_proj_b, cur),
+            cur);
+    return cur;
+}
+
+// evaluate the transformer
+//
+//   - model:     the model
+//   - n_threads: number of threads to use
+//   - n_past:    the context size so far
+//   - embd_inp:  the embeddings of the tokens in the context
+//   - embd_w:    the predicted logits for the next token
+//
+bool gpt_neox_eval(
+        const gpt_neox_model & model,
+        const int n_threads,
+        const int n_past,
+        const std::vector<gpt_vocab::id> & embd_inp,
+              std::vector<float>         & embd_w,
+              size_t                     & mem_per_token) {
+    const int N = embd_inp.size();
+
+    const auto & hparams = model.hparams;
+
+    const int n_embd  = hparams.n_embd;
+    const int n_layer = hparams.n_layer;
+    const int n_ctx   = hparams.n_ctx;
+    const int n_head  = hparams.n_head;
+    const int n_vocab = hparams.n_vocab;
+    const int n_rot   = hparams.n_rot;
+
+    static size_t buf_size = 256u*1024*1024;
+    static void * buf = malloc(buf_size);
+
+    // use 2 scratch buffers
+    // TODO: very hacky solution - reimplement in a more elegant way
+    static size_t scr0_size = 256u*1024*1024;
+    static void * scr0 = malloc(scr0_size);
+
+    static size_t scr1_size = 256u*1024*1024;
+    static void * scr1 = malloc(scr1_size);
+
+    if (mem_per_token > 0 && mem_per_token*N > buf_size) {
+        const size_t buf_size_new = 1.1*(mem_per_token*N); // add 10% to account for ggml object overhead
+        //printf("\n%s: reallocating buffer from %zu to %zu bytes\n", __func__, buf_size, buf_size_new);
+
+        // reallocate
+        buf_size = buf_size_new;
+        buf = realloc(buf, buf_size);
+        if (buf == nullptr) {
+            fprintf(stderr, "%s: failed to allocate %zu bytes\n", __func__, buf_size);
+            return false;
+        }
+    }
+
+    struct ggml_init_params params = {
+        /*.mem_size   =*/ buf_size,
+        /*.mem_buffer =*/ buf,
+        /*.no_alloc   =*/ false,
+    };
+
+    struct ggml_context * ctx0 = ggml_init(params);
+    struct ggml_cgraph gf = {};
+
+    struct ggml_tensor * embd = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
+    memcpy(embd->data, embd_inp.data(), N*ggml_element_size(embd));
+
+    // wte
+    struct ggml_tensor * inpL = ggml_get_rows(ctx0, model.wte, embd);
+
+    for (int il = 0; il < n_layer; ++il) {
+        struct ggml_tensor * cur;
+
+        ggml_set_scratch(ctx0, { 0, scr0_size, scr0, });
+
+        // self-attention
+        {
+            {
+                cur = ggml_norm(ctx0, inpL, hparams.eps);
+
+                cur = ggml_add(ctx0,
+                        ggml_mul(ctx0,
+                            ggml_repeat(ctx0, model.layers[il].ln_1_g, cur),
+                            cur),
+                        ggml_repeat(ctx0, model.layers[il].ln_1_b, cur));
+            }
+
+            // compute QKV
+            {
+                cur = ggml_mul_mat(ctx0,
+                        model.layers[il].c_attn_attn_w,
+                        cur);
+
+                cur = ggml_add(ctx0,
+                        ggml_repeat(ctx0, model.layers[il].c_attn_attn_b, cur),
+                        cur);
+            }
+
+            struct ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_3d(ctx0, cur, n_embd/n_head, n_head, N, cur->nb[1]/n_head, cur->nb[1], 0*sizeof(float)*n_embd/n_head));
+            struct ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_3d(ctx0, cur, n_embd/n_head, n_head, N, cur->nb[1]/n_head, cur->nb[1], 1*sizeof(float)*n_embd/n_head));
+            struct ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_3d(ctx0, cur, n_embd/n_head, n_head, N, cur->nb[1]/n_head, cur->nb[1], 2*sizeof(float)*n_embd/n_head));
+
+            // using mode = 2 for GPT-NeoX mode
+            Qcur = ggml_rope_inplace(ctx0, Qcur, n_past, n_rot, 2, 0);
+            Kcur = ggml_rope_inplace(ctx0, Kcur, n_past, n_rot, 2, 0);
+
+            // store key and value to memory
+            {
+                Vcur = ggml_transpose(ctx0, ggml_reshape_2d(ctx0, Vcur, n_embd, N));
+
+                struct ggml_tensor * k = ggml_view_1d(ctx0, model.memory_k, N*n_embd, (ggml_element_size(model.memory_k)*n_embd)*(il*n_ctx + n_past));
+                struct ggml_tensor * v = ggml_view_2d(ctx0, model.memory_v, N, n_embd,
+                        (   n_ctx)*ggml_element_size(model.memory_v),
+                        (il*n_ctx)*ggml_element_size(model.memory_v)*n_embd + n_past*ggml_element_size(model.memory_v));
+
+                ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Kcur, k));
+                ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Vcur, v));
+            }
+
+            // Q = Qcur.contiguous().view(n_embd/n_head, n_head, N).permute(0, 2, 1, 3)
+            struct ggml_tensor * Q =
+                ggml_permute(ctx0,
+                        Qcur,
+                        0, 2, 1, 3);
+
+            // K = Kmem.view(n_embd/n_head, n_head, n_past + N).permute(0, 2, 1, 3)
+            struct ggml_tensor * K =
+                ggml_permute(ctx0,
+                        ggml_reshape_3d(ctx0,
+                            ggml_view_1d(ctx0, model.memory_k, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(model.memory_k)*n_embd),
+                            n_embd/n_head, n_head, n_past + N),
+                        0, 2, 1, 3);
+
+            // K * Q
+            struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);
+
+            // KQ_scaled = KQ / sqrt(n_embd/n_head)
+            struct ggml_tensor * KQ_scaled =
+                ggml_scale_inplace(ctx0,
+                        KQ,
+                        ggml_new_f32(ctx0, 1.0f/sqrt(float(n_embd)/n_head))
+                        );
+
+            // KQ_masked = mask_past(KQ_scaled)
+            struct ggml_tensor * KQ_masked = ggml_diag_mask_inf_inplace(ctx0, KQ_scaled, n_past);
+
+            // KQ = soft_max(KQ_masked)
+            struct ggml_tensor * KQ_soft_max = ggml_soft_max_inplace(ctx0, KQ_masked);
+
+            // V_trans = Vmem.view(n_embd/n_head, n_head, n_past + N).permute(1, 2, 0, 3).contiguous()
+            struct ggml_tensor * V =
+                ggml_view_3d(ctx0, model.memory_v,
+                        n_past + N, n_embd/n_head, n_head,
+                        n_ctx*ggml_element_size(model.memory_v),
+                        n_ctx*ggml_element_size(model.memory_v)*n_embd/n_head,
+                        il*n_ctx*ggml_element_size(model.memory_v)*n_embd);
+
+            // KQV = transpose(V) * KQ_soft_max
+            struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max);
+
+            // KQV_merged = KQV.permute(0, 2, 1, 3)
+            struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
+
+            // cur = KQV_merged.contiguous().view(n_embd, N)
+            cur = ggml_cpy(ctx0,
+                    KQV_merged,
+                    ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, N));
+
+            // projection
+            {
+                cur = ggml_mul_mat(ctx0,
+                        model.layers[il].c_attn_proj_w,
+                        cur);
+
+                cur = ggml_add(ctx0, ggml_repeat(ctx0, model.layers[il].c_attn_proj_b, cur), cur);
+            }
+        }
+
+        ggml_set_scratch(ctx0, { 0, scr1_size, scr1, });
+
+        if (hparams.par_res == 0) {
+            struct ggml_tensor * inpFF = ggml_add(ctx0, cur, inpL);
+
+            cur = gpt_neox_ff(model.layers[il], ctx0, inpFF, hparams.eps);
+
+            // input for next layer
+            inpL = ggml_add(ctx0, cur, inpFF);
+        } else {
+            struct ggml_tensor * inpFF = cur;
+
+            // this is independent of the self-attention result, so it could be done in parallel to the self-attention
+            // note here we pass inpL instead of cur
+            cur = gpt_neox_ff(model.layers[il], ctx0, inpL, hparams.eps);
+
+            // layer input + FF
+            cur  = ggml_add(ctx0, cur, inpFF);
+
+            // input for next layer
+            inpL = ggml_add(ctx0, cur, inpL);
+        }
+    }
+
+    ggml_set_scratch(ctx0, { 0, scr0_size, scr0, });
+
+    // norm
+    {
+        inpL = ggml_norm(ctx0, inpL, hparams.eps);
+
+        // inpL = ln_f_g*inpL + ln_f_b
+        inpL = ggml_add(ctx0,
+                ggml_mul(ctx0,
+                    ggml_repeat(ctx0, model.ln_f_g, inpL),
+                    inpL),
+                ggml_repeat(ctx0, model.ln_f_b, inpL));
+    }
+
+    ggml_set_scratch(ctx0, { 0, 0, nullptr, });
+
+    // lm_head
+    {
+        inpL = ggml_mul_mat(ctx0, model.lmh_g, inpL);
+
+        //inpL = ggml_add(ctx0,
+        //        ggml_repeat(ctx0, model.lmh_b, inpL),
+        //        inpL);
+    }
+
+    // logits -> probs
+    //inpL = ggml_soft_max_inplace(ctx0, inpL);
+
+    // run the computation
+    ggml_build_forward_expand(&gf, inpL);
+    ggml_graph_compute_with_ctx(ctx0, &gf, n_threads);
+
+    //if (n_past%100 == 0) {
+    //    ggml_graph_print   (&gf);
+    //    ggml_graph_dump_dot(&gf, NULL, "gpt-2.dot");
+    //}
+
+    //embd_w.resize(n_vocab*N);
+    //memcpy(embd_w.data(), ggml_get_data(inpL), sizeof(float)*n_vocab*N);
+
+    // return result for just the last token
+    embd_w.resize(n_vocab);
+    memcpy(embd_w.data(), (float *) ggml_get_data(inpL) + (n_vocab*(N-1)), sizeof(float)*n_vocab);
+
+    if (mem_per_token == 0) {
+        mem_per_token = ggml_used_mem(ctx0)/N;
+    }
+    //printf("used_mem = %zu\n", ggml_used_mem(ctx0));
+
+    ggml_free(ctx0);
+
+    return true;
+}
+
+int main(int argc, char ** argv) {
+    ggml_time_init();
+
+    const int64_t t_main_start_us = ggml_time_us();
+
+    gpt_params params;
+    params.model = "models/stablelm-base-alpha-3b/ggml-model-f16.bin";
+
+    if (gpt_params_parse(argc, argv, params) == false) {
+        return 1;
+    }
+
+    if (params.seed < 0) {
+        params.seed = time(NULL);
+    }
+
+    printf("%s: seed = %d\n", __func__, params.seed);
+
+    std::mt19937 rng(params.seed);
+    if (params.prompt.empty()) {
+        params.prompt = gpt_random_prompt(rng);
+    }
+
+    int64_t t_load_us = 0;
+
+    gpt_vocab vocab;
+    gpt_neox_model model;
+
+    // load the model
+    {
+        const int64_t t_start_us = ggml_time_us();
+
+        if (!gpt_neox_model_load(params.model, model, vocab)) {
+            fprintf(stderr, "%s: failed to load model from '%s'\n", __func__, params.model.c_str());
+            return 1;
+        }
+
+        t_load_us = ggml_time_us() - t_start_us;
+
+        test_gpt_tokenizer(vocab, params.token_test);
+    }
+
+    int n_past = 0;
+
+    int64_t t_sample_us  = 0;
+    int64_t t_predict_us = 0;
+
+    std::vector<float> logits;
+
+    // tokenize the prompt
+    std::vector<gpt_vocab::id> embd_inp = ::gpt_tokenize(vocab, params.prompt);
+
+    params.n_predict = std::min(params.n_predict, model.hparams.n_ctx - (int) embd_inp.size());
+
+    printf("%s: number of tokens in prompt = %zu\n", __func__, embd_inp.size());
+    for (size_t i = 0; i < embd_inp.size(); i++) {
+        printf("%s: token[%zu] = %6d, %s\n", __func__, i, embd_inp[i], vocab.id_to_token.at(embd_inp[i]).c_str());
+    }
+    printf("\n");
+
+    std::vector<gpt_vocab::id> embd;
+
+    // determine the required inference memory per token:
+    size_t mem_per_token = 0;
+    gpt_neox_eval(model, params.n_threads, 0, { 0, 1, 2, 3 }, logits, mem_per_token);
+
+    for (size_t i = embd.size(); i < embd_inp.size() + params.n_predict; i++) {
+        // predict
+        if (embd.size() > 0) {
+            const int64_t t_start_us = ggml_time_us();
+
+            if (!gpt_neox_eval(model, params.n_threads, n_past, embd, logits, mem_per_token)) {
+                printf("Failed to predict\n");
+                return 1;
+            }
+
+            t_predict_us += ggml_time_us() - t_start_us;
+        }
+
+        n_past += embd.size();
+        embd.clear();
+
+        if (i >= embd_inp.size()) {
+            // sample next token
+            const int   top_k = params.top_k;
+            const float top_p = params.top_p;
+            const float temp  = params.temp;
+
+            const int n_vocab = model.hparams.n_vocab;
+
+            gpt_vocab::id id = 0;
+
+            {
+                const int64_t t_start_sample_us = ggml_time_us();
+
+                id = gpt_sample_top_k_top_p(vocab, logits.data() + (logits.size() - n_vocab), top_k, top_p, temp, rng);
+
+                t_sample_us += ggml_time_us() - t_start_sample_us;
+            }
+
+            // add it to the context
+            embd.push_back(id);
+        } else {
+            // if here, it means we are still processing the input prompt
+            for (size_t k = i; k < embd_inp.size(); k++) {
+                embd.push_back(embd_inp[k]);
+                if (int32_t(embd.size()) > params.n_batch) {
+                    break;
+                }
+            }
+            i += embd.size() - 1;
+        }
+
+        // display text
+        for (auto id : embd) {
+            printf("%s", vocab.id_to_token[id].c_str());
+        }
+        fflush(stdout);
+
+        // end of text token
+        if (embd.back() == 0) {
+            break;
+        }
+    }
+
+    // report timing
+    {
+        const int64_t t_main_end_us = ggml_time_us();
+
+        printf("\n\n");
+        printf("%s: mem per token = %8zu bytes\n", __func__, mem_per_token);
+        printf("%s:     load time = %8.2f ms\n", __func__, t_load_us/1000.0f);
+        printf("%s:   sample time = %8.2f ms\n", __func__, t_sample_us/1000.0f);
+        printf("%s:  predict time = %8.2f ms / %.2f ms per token\n", __func__, t_predict_us/1000.0f, t_predict_us/1000.0f/n_past);
+        printf("%s:    total time = %8.2f ms\n", __func__, (t_main_end_us - t_main_start_us)/1000.0f);
+    }
+
+    ggml_free(model.ctx);
+
+    return 0;
+}

ggml/examples/gpt-neox/quantize.cpp

@@ -0,0 +1,178 @@
+#include "ggml/ggml.h"
+
+#include "common.h"
+#include "common-ggml.h"
+
+#include <cassert>
+#include <cmath>
+#include <cstdio>
+#include <cstring>
+#include <fstream>
+#include <map>
+#include <string>
+#include <vector>
+#include <regex>
+
+// default hparams (StableLM 3B)
+struct gpt_neox_hparams {
+    int32_t n_vocab = 50257;
+    int32_t n_ctx   = 4096;
+    int32_t n_embd  = 4096;
+    int32_t n_head  = 32;
+    int32_t n_layer = 16;
+    int32_t n_rot   = 32; // 0.25 * (n_embd / n_head)
+    int32_t par_res = 1; // 1 = true, 0 = false
+    int32_t ftype   = 1;
+};
+
+// quantize a model
+bool gpt_neox_model_quantize(const std::string & fname_inp, const std::string & fname_out, ggml_ftype ftype) {
+    gpt_vocab vocab;
+
+    printf("%s: loading model from '%s'\n", __func__, fname_inp.c_str());
+
+    auto finp = std::ifstream(fname_inp, std::ios::binary);
+    if (!finp) {
+        fprintf(stderr, "%s: failed to open '%s' for reading\n", __func__, fname_inp.c_str());
+        return false;
+    }
+
+    auto fout = std::ofstream(fname_out, std::ios::binary);
+    if (!fout) {
+        fprintf(stderr, "%s: failed to open '%s' for writing\n", __func__, fname_out.c_str());
+        return false;
+    }
+
+    // verify magic
+    {
+        uint32_t magic;
+        finp.read((char *) &magic, sizeof(magic));
+        if (magic != GGML_FILE_MAGIC) {
+            fprintf(stderr, "%s: invalid model file '%s' (bad magic)\n", __func__, fname_inp.c_str());
+            return false;
+        }
+
+        fout.write((char *) &magic, sizeof(magic));
+    }
+
+    gpt_neox_hparams hparams;
+
+    // load hparams
+    {
+        finp.read((char *) &hparams.n_vocab, sizeof(hparams.n_vocab));
+        finp.read((char *) &hparams.n_ctx,   sizeof(hparams.n_ctx));
+        finp.read((char *) &hparams.n_embd,  sizeof(hparams.n_embd));
+        finp.read((char *) &hparams.n_head,  sizeof(hparams.n_head));
+        finp.read((char *) &hparams.n_layer, sizeof(hparams.n_layer));
+        finp.read((char *) &hparams.n_rot,   sizeof(hparams.n_rot));
+        finp.read((char *) &hparams.par_res, sizeof(hparams.par_res));
+        finp.read((char *) &hparams.ftype,   sizeof(hparams.ftype));
+
+        const int32_t qntvr_src =    hparams.ftype / GGML_QNT_VERSION_FACTOR;
+        const int32_t ftype_dst = GGML_QNT_VERSION * GGML_QNT_VERSION_FACTOR + ftype;
+
+        printf("%s: n_vocab     = %d\n", __func__, hparams.n_vocab);
+        printf("%s: n_ctx       = %d\n", __func__, hparams.n_ctx);
+        printf("%s: n_embd      = %d\n", __func__, hparams.n_embd);
+        printf("%s: n_head      = %d\n", __func__, hparams.n_head);
+        printf("%s: n_layer     = %d\n", __func__, hparams.n_layer);
+        printf("%s: par_res     = %d\n", __func__, hparams.par_res);
+        printf("%s: ftype (src) = %d\n", __func__, hparams.ftype);
+        printf("%s: qntvr (src) = %d\n", __func__, qntvr_src);
+        printf("%s: ftype (dst) = %d\n", __func__, ftype_dst);
+        printf("%s: qntvr (dst) = %d\n", __func__, GGML_QNT_VERSION);
+
+        fout.write((char *) &hparams.n_vocab, sizeof(hparams.n_vocab));
+        fout.write((char *) &hparams.n_ctx,   sizeof(hparams.n_ctx));
+        fout.write((char *) &hparams.n_embd,  sizeof(hparams.n_embd));
+        fout.write((char *) &hparams.n_head,  sizeof(hparams.n_head));
+        fout.write((char *) &hparams.n_layer, sizeof(hparams.n_layer));
+        fout.write((char *) &hparams.n_rot,   sizeof(hparams.n_rot));
+        fout.write((char *) &hparams.par_res, sizeof(hparams.par_res));
+        fout.write((char *) &ftype_dst,       sizeof(ftype_dst));
+    }
+
+    // load vocab
+    {
+        const int32_t n_vocab = hparams.n_vocab;
+
+        std::string word;
+        for (int i = 0; i < n_vocab; i++) {
+            uint32_t len;
+            finp.read ((char *) &len, sizeof(len));
+            fout.write((char *) &len, sizeof(len));
+
+            word.resize(len);
+            finp.read ((char *) word.data(), len);
+            fout.write((char *) word.data(), len);
+
+            vocab.token_to_id[word] = i;
+            vocab.id_to_token[i] = word;
+        }
+    }
+
+    // regexes of tensor names to be quantized
+    const std::vector<std::string> to_quant = {
+        ".*weight",
+    };
+
+    if (!ggml_common_quantize_0(finp, fout, ftype, to_quant, {})) {
+        fprintf(stderr, "%s: failed to quantize model '%s'\n", __func__, fname_inp.c_str());
+        return false;
+    }
+
+    finp.close();
+    fout.close();
+
+    return true;
+}
+
+// usage:
+//  ./gpt-neox-quantize models/stalellm2-117M/ggml-model.bin models/stablelm2-117M/ggml-model-quant.bin type
+//
+int main(int argc, char ** argv) {
+    if (argc != 4) {
+        fprintf(stderr, "usage: %s model-f32.bin model-quant.bin type\n", argv[0]);
+        ggml_print_ftypes(stderr);
+        return 1;
+    }
+
+    // needed to initialize f16 tables
+    {
+        struct ggml_init_params params = { 0, NULL, false };
+        struct ggml_context * ctx = ggml_init(params);
+        ggml_free(ctx);
+    }
+
+    const std::string fname_inp = argv[1];
+    const std::string fname_out = argv[2];
+
+    const ggml_ftype ftype = ggml_parse_ftype(argv[3]);
+
+    const int64_t t_main_start_us = ggml_time_us();
+
+    int64_t t_quantize_us = 0;
+
+    // load the model
+    {
+        const int64_t t_start_us = ggml_time_us();
+
+        if (!gpt_neox_model_quantize(fname_inp, fname_out, ggml_ftype(ftype))) {
+            fprintf(stderr, "%s: failed to quantize model from '%s'\n", __func__, fname_inp.c_str());
+            return 1;
+        }
+
+        t_quantize_us = ggml_time_us() - t_start_us;
+    }
+
+    // report timing
+    {
+        const int64_t t_main_end_us = ggml_time_us();
+
+        printf("\n");
+        printf("%s: quantize time = %8.2f ms\n", __func__, t_quantize_us/1000.0f);
+        printf("%s:    total time = %8.2f ms\n", __func__, (t_main_end_us - t_main_start_us)/1000.0f);
+    }
+
+    return 0;
+}

ggml/examples/mnist/CMakeLists.txt

@@ -0,0 +1,40 @@
+#
+# mnist
+
+set(TEST_TARGET mnist)
+add_executable(${TEST_TARGET} main.cpp)
+target_link_libraries(${TEST_TARGET} PRIVATE ggml common)
+
+#
+# mnist-cnn
+
+set(TEST_TARGET mnist-cnn)
+add_executable(${TEST_TARGET} main-cnn.cpp)
+target_link_libraries(${TEST_TARGET} PRIVATE ggml common)
+
+#
+# mnist-cpu
+
+set(TEST_TARGET mnist-cpu)
+add_executable(${TEST_TARGET} main-cpu.cpp)
+target_link_libraries(${TEST_TARGET} PRIVATE ggml)
+
+if (APPLE)
+    #
+    # mnist-mtl
+
+    find_library(FOUNDATION_LIBRARY Foundation REQUIRED)
+    find_library(METAL_FRAMEWORK    Metal      REQUIRED)
+    find_library(METALKIT_FRAMEWORK MetalKit   REQUIRED)
+    find_library(METALPERFORMANCE_FRAMEWORK MetalPerformanceShaders REQUIRED)
+
+    set(TEST_TARGET mnist-mtl)
+    add_executable(${TEST_TARGET} main-mtl.cpp main-mtl.h main-mtl.m)
+    target_link_libraries(${TEST_TARGET} PRIVATE
+        ggml
+        ${FOUNDATION_LIBRARY}
+        ${METAL_FRAMEWORK}
+        ${METALKIT_FRAMEWORK}
+        ${METALPERFORMANCE_FRAMEWORK}
+    )
+endif()

ggml/examples/mnist/README.md

@@ -0,0 +1,128 @@
+# MNIST Examples for GGML
+
+These are simple examples of how to use GGML for inferencing.
+The first example uses convolutional neural network (CNN), the second one uses fully connected neural network.
+
+## Building the examples
+
+```bash
+git clone https://github.com/ggerganov/ggml
+cd ggml
+mkdir build && cd build
+cmake ..
+make -j4 mnist-cnn mnist
+```
+
+## MNIST with CNN
+
+This implementation achieves ~99% accuracy on the MNIST test set.
+
+### Training the model
+
+Use the `mnist-cnn.py` script to train the model and convert it to GGUF format:
+
+```
+$ python3 ../examples/mnist/mnist-cnn.py train mnist-cnn-model
+...
+Keras model saved to 'mnist-cnn-model'
+```
+
+Convert the model to GGUF format:
+
+```
+$ python3 ../examples/mnist/mnist-cnn.py convert mnist-cnn-model
+...
+Model converted and saved to 'mnist-cnn-model.gguf'
+```
+
+### Running the example
+
+```bash
+$ ./bin/mnist-cnn mnist-cnn-model.gguf ../examples/mnist/models/mnist/t10k-images.idx3-ubyte
+main: loaded model in     5.17 ms
+_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 
+_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 
+_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 
+_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 
+_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ * * * * * _ _ _ _ _ _ _ 
+_ _ _ _ _ _ _ _ _ _ _ _ _ _ * * * * * * * * _ _ _ _ _ _ 
+_ _ _ _ _ _ _ _ _ _ _ _ * * * * * _ _ _ * * _ _ _ _ _ _ 
+_ _ _ _ _ _ _ _ _ _ _ _ * * _ _ _ _ _ _ _ _ _ _ _ _ _ _ 
+_ _ _ _ _ _ _ _ _ _ _ * * * _ _ _ _ _ _ _ _ _ _ _ _ _ _ 
+_ _ _ _ _ _ _ _ _ _ _ * * * _ _ _ _ _ _ _ _ * _ _ _ _ _ 
+_ _ _ _ _ _ _ _ _ _ _ * * * _ _ _ _ _ _ _ _ * * _ _ _ _ 
+_ _ _ _ _ _ _ _ _ _ _ * * * _ _ _ _ _ _ _ _ * * _ _ _ _ 
+_ _ _ _ _ _ _ _ _ _ _ _ * * * _ _ _ _ * * * * * _ _ _ _ 
+_ _ _ _ _ _ _ _ _ _ _ _ * * * * * * * * * _ _ _ _ _ _ _ 
+_ _ _ _ _ _ _ _ * * * * * * * * * * _ _ _ _ _ _ _ _ _ _ 
+_ _ _ _ _ _ _ * * * * * * _ _ * * * _ _ _ _ _ _ _ _ _ _ 
+_ _ _ _ _ _ * * * _ _ _ _ _ _ _ * * * _ _ _ _ _ _ _ _ _ 
+_ _ _ _ _ _ * * _ _ _ _ _ _ _ _ _ * * _ _ _ _ _ _ _ _ _ 
+_ _ _ _ _ _ * * _ _ _ _ _ _ _ _ _ * * * _ _ _ _ _ _ _ _ 
+_ _ _ _ _ _ * * _ _ _ _ _ _ _ _ _ * * * _ _ _ _ _ _ _ _ 
+_ _ _ _ _ _ * * * _ _ _ _ _ _ _ _ * * * _ _ _ _ _ _ _ _ 
+_ _ _ _ _ _ _ _ * * * _ _ _ _ _ _ * * * _ _ _ _ _ _ _ _ 
+_ _ _ _ _ _ _ _ _ * * * * * * * * * * _ _ _ _ _ _ _ _ _ 
+_ _ _ _ _ _ _ _ _ _ _ * * * * * * _ _ _ _ _ _ _ _ _ _ _ 
+_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 
+_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 
+_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 
+_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 
+
+ggml_graph_dump_dot: dot -Tpng mnist-cnn.dot -o mnist-cnn.dot.png && open mnist-cnn.dot.png
+main: predicted digit is 8
+```
+
+Computation graph:
+
+![mnist dot](https://user-images.githubusercontent.com/1991296/263763842-3b679b45-7ca1-4ee9-b19a-82e34396624f.png)
+
+## MNIST with fully connected network
+
+A fully connected layer + relu, followed by a fully connected layer + softmax.
+
+### Training the Model
+
+A Google Colab notebook for training a simple two-layer network to recognize digits is located here. You can
+use this to save a pytorch model to be converted to ggml format.
+
+[Colab](https://colab.research.google.com/drive/12n_8VNJnolBnX5dVS0HNWubnOjyEaFSb?usp=sharing)
+
+GGML "format" is whatever you choose for efficient loading. In our case, we just save the hyperparameters used
+plus the model weights and biases. Run convert-h5-to-ggml.py to convert your pytorch model. The output format is:
+
+- magic constant (int32)
+- repeated list of tensors
+- number of dimensions of tensor (int32)
+- tensor dimension (int32 repeated)
+- values of tensor (int32)
+
+Run ```convert-h5-to-ggml.py mnist_model.state_dict``` where `mnist_model.state_dict` is the saved pytorch model from the Google Colab. For
+quickstart, it is included in the mnist/models directory.
+
+```bash
+mkdir -p models/mnist
+python3 ../examples/mnist/convert-h5-to-ggml.py ../examples/mnist/models/mnist/mnist_model.state_dict
+```
+
+### Running the example
+
+```bash
+./bin/mnist ./models/mnist/ggml-model-f32.bin ../examples/mnist/models/mnist/t10k-images.idx3-ubyte
+```
+
+Computation graph:
+
+![mnist dot](https://user-images.githubusercontent.com/1991296/231882071-84e29d53-b226-4d73-bdc2-5bd6dcb7efd1.png)
+
+
+## Web demo
+
+The example can be compiled with Emscripten like this:
+
+```bash
+cd examples/mnist
+emcc -I../../include -I../../include/ggml -I../../examples ../../src/ggml.c main.cpp -o web/mnist.js -s EXPORTED_FUNCTIONS='["_wasm_eval","_wasm_random_digit","_malloc","_free"]' -s EXPORTED_RUNTIME_METHODS='["ccall"]' -s ALLOW_MEMORY_GROWTH=1 --preload-file models/mnist
+```
+
+Online demo: https://mnist.ggerganov.com

ggml/examples/mnist/convert-h5-to-ggml.py

@@ -0,0 +1,63 @@
+# Convert MNIS h5 transformer model to ggml format
+#
+# Load the (state_dict) saved model using PyTorch
+# Iterate over all variables and write them to a binary file.
+#
+# For each variable, write the following:
+#   - Number of dimensions (int)
+#   - Name length (int)
+#   - Dimensions (int[n_dims])
+#   - Name (char[name_length])
+#   - Data (float[n_dims])
+#
+# At the start of the ggml file we write the model parameters
+
+import sys
+import struct
+import json
+import numpy as np
+import re
+
+
+import torch
+import torch.nn as nn
+import torchvision.datasets as dsets
+import torchvision.transforms as transforms
+from torch.autograd import Variable
+
+if len(sys.argv) != 2:
+    print("Usage: convert-h5-to-ggml.py model\n")
+    sys.exit(1)
+
+state_dict_file = sys.argv[1]
+fname_out = "models/mnist/ggml-model-f32.bin"
+
+state_dict = torch.load(state_dict_file, map_location=torch.device('cpu'))
+#print (model)
+
+list_vars = state_dict
+print (list_vars)
+
+fout = open(fname_out, "wb")
+
+fout.write(struct.pack("i", 0x67676d6c)) # magic: ggml in hex
+
+
+for name in list_vars.keys():
+    data = list_vars[name].squeeze().numpy()
+    print("Processing variable: " + name + " with shape: ", data.shape) 
+    n_dims = len(data.shape);
+   
+    fout.write(struct.pack("i", n_dims))
+    
+    data = data.astype(np.float32)
+    for i in range(n_dims):
+        fout.write(struct.pack("i", data.shape[n_dims - 1 - i]))
+
+    # data
+    data.tofile(fout)
+
+fout.close()
+
+print("Done. Output file: " + fname_out)
+print("")

ggml/examples/mnist/main-cnn.cpp

@@ -0,0 +1,169 @@
+#include "ggml/ggml.h"
+
+#include "common.h"
+
+#include <cmath>
+#include <cstdio>
+#include <cstring>
+#include <ctime>
+#include <fstream>
+#include <string>
+#include <vector>
+#include <algorithm>
+
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
+struct mnist_model {
+    struct ggml_tensor * conv2d_1_kernel;
+    struct ggml_tensor * conv2d_1_bias;
+    struct ggml_tensor * conv2d_2_kernel;
+    struct ggml_tensor * conv2d_2_bias;
+    struct ggml_tensor * dense_weight;
+    struct ggml_tensor * dense_bias;
+    struct ggml_context * ctx;
+};
+
+bool mnist_model_load(const std::string & fname, mnist_model & model) {
+    struct gguf_init_params params = {
+        /*.no_alloc   =*/ false,
+        /*.ctx        =*/ &model.ctx,
+    };
+    gguf_context * ctx = gguf_init_from_file(fname.c_str(), params);
+    if (!ctx) {
+        fprintf(stderr, "%s: gguf_init_from_file() failed\n", __func__);
+        return false;
+    }
+    model.conv2d_1_kernel = ggml_get_tensor(model.ctx, "kernel1");
+    model.conv2d_1_bias = ggml_get_tensor(model.ctx, "bias1");
+    model.conv2d_2_kernel = ggml_get_tensor(model.ctx, "kernel2");
+    model.conv2d_2_bias = ggml_get_tensor(model.ctx, "bias2");
+    model.dense_weight = ggml_get_tensor(model.ctx, "dense_w");
+    model.dense_bias = ggml_get_tensor(model.ctx, "dense_b");
+    return true;
+}
+
+int mnist_eval(
+        const mnist_model & model,
+        const int n_threads,
+        std::vector<float> digit,
+        const char * fname_cgraph
+        )
+{
+    static size_t buf_size = 100000 * sizeof(float) * 4;
+    static void * buf = malloc(buf_size);
+
+    struct ggml_init_params params = {
+        /*.mem_size   =*/ buf_size,
+        /*.mem_buffer =*/ buf,
+        /*.no_alloc   =*/ false,
+    };
+
+    struct ggml_context * ctx0 = ggml_init(params);
+    struct ggml_cgraph gf = {};
+
+    struct ggml_tensor * input = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, 28, 28, 1, 1);
+    memcpy(input->data, digit.data(), ggml_nbytes(input));
+    ggml_set_name(input, "input");
+    ggml_tensor * cur = ggml_conv_2d(ctx0, model.conv2d_1_kernel, input, 1, 1, 0, 0, 1, 1);
+    cur = ggml_add(ctx0, cur, model.conv2d_1_bias);
+    cur = ggml_relu(ctx0, cur);
+    // Output shape after Conv2D: (26 26 32 1)
+    cur = ggml_pool_2d(ctx0, cur, GGML_OP_POOL_MAX, 2, 2, 2, 2, 0, 0);
+    // Output shape after MaxPooling2D: (13 13 32 1)
+    cur = ggml_conv_2d(ctx0, model.conv2d_2_kernel, cur, 1, 1, 0, 0, 1, 1);
+    cur = ggml_add(ctx0, cur, model.conv2d_2_bias);
+    cur = ggml_relu(ctx0, cur);
+    // Output shape after Conv2D: (11 11 64 1)
+    cur = ggml_pool_2d(ctx0, cur, GGML_OP_POOL_MAX, 2, 2, 2, 2, 0, 0);
+    // Output shape after MaxPooling2D: (5 5 64 1)
+    cur = ggml_cont(ctx0, ggml_permute(ctx0, cur, 1, 2, 0, 3));
+    // Output shape after permute: (64 5 5 1)
+    cur = ggml_reshape_2d(ctx0, cur, 1600, 1);
+    // Final Dense layer
+    cur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.dense_weight, cur), model.dense_bias);
+    ggml_tensor * probs = ggml_soft_max(ctx0, cur);
+    ggml_set_name(probs, "probs");
+
+    ggml_build_forward_expand(&gf, probs);
+    ggml_graph_compute_with_ctx(ctx0, &gf, n_threads);
+
+    //ggml_graph_print(&gf);
+    ggml_graph_dump_dot(&gf, NULL, "mnist-cnn.dot");
+
+    if (fname_cgraph) {
+        // export the compute graph for later use
+        // see the "mnist-cpu" example
+        ggml_graph_export(&gf, fname_cgraph);
+
+        fprintf(stderr, "%s: exported compute graph to '%s'\n", __func__, fname_cgraph);
+    }
+
+    const float * probs_data = ggml_get_data_f32(probs);
+    const int prediction = std::max_element(probs_data, probs_data + 10) - probs_data;
+    ggml_free(ctx0);
+    return prediction;
+}
+
+int main(int argc, char ** argv) {
+    srand(time(NULL));
+    ggml_time_init();
+
+    if (argc != 3) {
+        fprintf(stderr, "Usage: %s models/mnist/mnist-cnn.gguf models/mnist/t10k-images.idx3-ubyte\n", argv[0]);
+        exit(0);
+    }
+
+    uint8_t buf[784];
+    mnist_model model;
+    std::vector<float> digit;
+
+    // load the model
+    {
+        const int64_t t_start_us = ggml_time_us();
+
+        if (!mnist_model_load(argv[1], model)) {
+            fprintf(stderr, "%s: failed to load model from '%s'\n", __func__, argv[1]);
+            return 1;
+        }
+
+        const int64_t t_load_us = ggml_time_us() - t_start_us;
+
+        fprintf(stdout, "%s: loaded model in %8.2f ms\n", __func__, t_load_us / 1000.0f);
+    }
+
+    // read a random digit from the test set
+    {
+        std::ifstream fin(argv[2], std::ios::binary);
+        if (!fin) {
+            fprintf(stderr, "%s: failed to open '%s'\n", __func__, argv[2]);
+            return 1;
+        }
+
+        // seek to a random digit: 16-byte header + 28*28 * (random 0 - 10000)
+        fin.seekg(16 + 784 * (rand() % 10000));
+        fin.read((char *) &buf, sizeof(buf));
+    }
+
+    // render the digit in ASCII
+    {
+        digit.resize(sizeof(buf));
+
+        for (int row = 0; row < 28; row++) {
+            for (int col = 0; col < 28; col++) {
+                fprintf(stderr, "%c ", (float)buf[row*28 + col] > 230 ? '*' : '_');
+                digit[row*28 + col] = ((float)buf[row*28 + col] / 255.0f);
+            }
+
+            fprintf(stderr, "\n");
+        }
+
+        fprintf(stderr, "\n");
+    }
+
+    const int prediction = mnist_eval(model, 1, digit, nullptr);
+    fprintf(stdout, "%s: predicted digit is %d\n", __func__, prediction);
+    ggml_free(model.ctx);
+    return 0;
+}

ggml/examples/mnist/main-cpu.cpp

@@ -0,0 +1,122 @@
+// Use a pre-generated MNIST compute graph for inference on the CPU
+//
+// You can generate a compute graph using the "mnist" tool:
+//
+// $ ./bin/mnist ./models/mnist/ggml-model-f32.bin ../examples/mnist/models/mnist/t10k-images.idx3-ubyte
+//
+// This command creates the "mnist.ggml" file, which contains the generated compute graph.
+// Now, you can re-use the compute graph with the "mnist-cpu" tool:
+//
+// $ ./bin/mnist-cpu ./models/mnist/mnist.ggml ../examples/mnist/models/mnist/t10k-images.idx3-ubyte
+//
+
+#include "ggml/ggml.h"
+
+#include <algorithm>
+#include <cmath>
+#include <cstdio>
+#include <cstring>
+#include <ctime>
+#include <fstream>
+#include <vector>
+
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
+// evaluate the MNIST compute graph
+//
+//   - fname_cgraph: path to the compute graph
+//   - n_threads:    number of threads to use
+//   - digit:        784 pixel values
+//
+// returns 0 - 9 prediction
+int mnist_eval(
+        const char * fname_cgraph,
+        const int n_threads,
+        std::vector<float> digit) {
+    // load the compute graph
+    struct ggml_context * ctx_data = NULL;
+    struct ggml_context * ctx_eval = NULL;
+
+    struct ggml_cgraph gfi = ggml_graph_import(fname_cgraph, &ctx_data, &ctx_eval);
+
+    // param export/import test
+    GGML_ASSERT(ggml_graph_get_tensor(&gfi, "fc1_bias")->op_params[0] == int(0xdeadbeef));
+
+    // allocate work context
+    // needed during ggml_graph_compute() to allocate a work tensor
+    static size_t buf_size = 128ull*1024*1024; // TODO
+    static void * buf = malloc(buf_size);
+
+    struct ggml_init_params params = {
+        /*.mem_size   =*/ buf_size,
+        /*.mem_buffer =*/ buf,
+        /*.no_alloc   =*/ false,
+    };
+
+    struct ggml_context * ctx_work = ggml_init(params);
+
+    struct ggml_tensor * input = ggml_graph_get_tensor(&gfi, "input");
+    memcpy(input->data, digit.data(), ggml_nbytes(input));
+
+    ggml_graph_compute_with_ctx(ctx_work, &gfi, n_threads);
+
+    const float * probs_data = ggml_get_data_f32(ggml_graph_get_tensor(&gfi, "probs"));
+
+    const int prediction = std::max_element(probs_data, probs_data + 10) - probs_data;
+
+    ggml_free(ctx_work);
+    ggml_free(ctx_data);
+    ggml_free(ctx_eval);
+
+    return prediction;
+}
+
+int main(int argc, char ** argv) {
+    srand(time(NULL));
+    ggml_time_init();
+
+    if (argc != 3) {
+        fprintf(stderr, "Usage: %s models/mnist/mnist.ggml models/mnist/t10k-images.idx3-ubyte\n", argv[0]);
+        exit(0);
+    }
+
+    uint8_t buf[784];
+    std::vector<float> digit;
+
+    // read a random digit from the test set
+    {
+        std::ifstream fin(argv[2], std::ios::binary);
+        if (!fin) {
+            fprintf(stderr, "%s: failed to open '%s'\n", __func__, argv[2]);
+            return 1;
+        }
+
+        // seek to a random digit: 16-byte header + 28*28 * (random 0 - 10000)
+        fin.seekg(16 + 784 * (rand() % 10000));
+        fin.read((char *) &buf, sizeof(buf));
+    }
+
+    // render the digit in ASCII
+    {
+        digit.resize(sizeof(buf));
+
+        for (int row = 0; row < 28; row++) {
+            for (int col = 0; col < 28; col++) {
+                fprintf(stderr, "%c ", (float)buf[row*28 + col] > 230 ? '*' : '_');
+                digit[row*28 + col] = ((float)buf[row*28 + col]);
+            }
+
+            fprintf(stderr, "\n");
+        }
+
+        fprintf(stderr, "\n");
+    }
+
+    const int prediction = mnist_eval(argv[1], 1, digit);
+
+    fprintf(stdout, "%s: predicted digit is %d\n", __func__, prediction);
+
+    return 0;
+}

ggml/examples/mnist/main-mtl.cpp

@@ -0,0 +1,125 @@
+// Use a pre-generated MNIST compute graph for inference on the M1 GPU via MPS
+//
+// You can generate a compute graph using the "mnist" tool:
+//
+// $ ./bin/mnist ./models/mnist/ggml-model-f32.bin ../examples/mnist/models/mnist/t10k-images.idx3-ubyte
+//
+// This command creates the "mnist.ggml" file, which contains the generated compute graph.
+// Now, you can re-use the compute graph on the GPU with the "mnist-mtl" tool:
+//
+// $ ./bin/mnist-mtl ./models/mnist/mnist.ggml ../examples/mnist/models/mnist/t10k-images.idx3-ubyte
+//
+
+#include "ggml/ggml.h"
+
+#include "main-mtl.h"
+
+#include <cmath>
+#include <cstdio>
+#include <cstring>
+#include <ctime>
+#include <fstream>
+#include <vector>
+
+// evaluate the MNIST compute graph
+//
+//   - fname_cgraph: path to the compute graph
+//   - digit:        784 pixel values
+//
+// returns 0 - 9 prediction
+int mnist_eval(
+        const char * fname_cgraph,
+        std::vector<float> digit
+        ) {
+    // load the compute graph
+    struct ggml_context * ctx_data = NULL;
+    struct ggml_context * ctx_eval = NULL;
+
+    struct ggml_cgraph gf = ggml_graph_import(fname_cgraph, &ctx_data, &ctx_eval);
+
+    // allocate work context
+    static size_t buf_size = 128ull*1024*1024; // TODO
+    static void * buf = malloc(buf_size);
+
+    struct ggml_init_params params = {
+        /*.mem_size   =*/ buf_size,
+        /*.mem_buffer =*/ buf,
+        /*.no_alloc   =*/ false,
+    };
+
+    struct ggml_context * ctx_work = ggml_init(params);
+
+    // this allocates all Metal resources and memory buffers
+    auto ctx_mtl = mnist_mtl_init(ctx_data, ctx_eval, ctx_work, &gf);
+
+    int prediction = -1;
+
+    for (int i = 0; i < 1; ++i) {
+        struct ggml_tensor * input = ggml_graph_get_tensor(&gf, "input");
+
+        if (i % 2 == 0) {
+            memcpy(input->data, digit.data(), ggml_nbytes(input));
+        } else {
+            memset(input->data, 0, ggml_nbytes(input));
+        }
+
+        // the actual inference happens here
+        prediction = mnist_mtl_eval(ctx_mtl, &gf);
+    }
+
+    mnist_mtl_free(ctx_mtl);
+
+    ggml_free(ctx_work);
+    ggml_free(ctx_data);
+    ggml_free(ctx_eval);
+
+    return prediction;
+}
+
+int main(int argc, char ** argv) {
+    srand(time(NULL));
+    ggml_time_init();
+
+    if (argc != 3) {
+        fprintf(stderr, "Usage: %s models/mnist/mnist.ggml models/mnist/t10k-images.idx3-ubyte\n", argv[0]);
+        exit(0);
+    }
+
+    uint8_t buf[784];
+    std::vector<float> digit;
+
+    // read a random digit from the test set
+    {
+        std::ifstream fin(argv[2], std::ios::binary);
+        if (!fin) {
+            fprintf(stderr, "%s: failed to open '%s'\n", __func__, argv[2]);
+            return 1;
+        }
+
+        // seek to a random digit: 16-byte header + 28*28 * (random 0 - 10000)
+        fin.seekg(16 + 784 * (rand() % 10000));
+        fin.read((char *) &buf, sizeof(buf));
+    }
+
+    // render the digit in ASCII
+    {
+        digit.resize(sizeof(buf));
+
+        for (int row = 0; row < 28; row++) {
+            for (int col = 0; col < 28; col++) {
+                fprintf(stderr, "%c ", (float)buf[row*28 + col] > 230 ? '*' : '_');
+                digit[row*28 + col] = ((float)buf[row*28 + col]);
+            }
+
+            fprintf(stderr, "\n");
+        }
+
+        fprintf(stderr, "\n");
+    }
+
+    const int prediction = mnist_eval(argv[1], digit);
+
+    fprintf(stdout, "%s: predicted digit is %d\n", __func__, prediction);
+
+    return 0;
+}

ggml/examples/mnist/main-mtl.h

@@ -0,0 +1,26 @@
+#pragma once
+
+struct ggml_context;
+struct ggml_cgraph;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct ggml_mtl_context;
+
+struct ggml_mtl_context * mnist_mtl_init(
+        struct ggml_context * ctx_data,
+        struct ggml_context * ctx_eval,
+        struct ggml_context * ctx_work,
+        struct ggml_cgraph  * gf);
+
+void mnist_mtl_free(struct ggml_mtl_context * ctx);
+
+int mnist_mtl_eval(
+        struct ggml_mtl_context * ctx,
+        struct ggml_cgraph      * gf);
+
+#ifdef __cplusplus
+}
+#endif

ggml/examples/mnist/main-mtl.m

@@ -0,0 +1,499 @@
+#import "main-mtl.h"
+
+#import "ggml/ggml.h"
+
+#import <Foundation/Foundation.h>
+#import <Metal/Metal.h>
+#import <MetalPerformanceShaders/MetalPerformanceShaders.h>
+
+// TODO: couldn't get this to work
+//#define GGML_MTL_HEAP
+
+struct ggml_mtl_context {
+    struct ggml_context * ctx_data;
+    struct ggml_context * ctx_eval;
+    struct ggml_context * ctx_work;
+
+    id<MTLDevice>       device;
+    id<MTLCommandQueue> queue;
+    id<MTLLibrary>      library;
+
+#ifdef GGML_MTL_HEAP
+    id<MTLHeap> heap_data;
+    id<MTLHeap> heap_eval;
+#else
+    id<MTLBuffer> buffer_data;
+    id<MTLBuffer> buffer_eval;
+#endif
+
+    id<MTLBuffer> out;
+
+    // custom kernels
+    id<MTLFunction>             function_add;
+    id<MTLComputePipelineState> pipeline_add;
+
+    id<MTLFunction>             function_relu;
+    id<MTLComputePipelineState> pipeline_relu;
+
+    id<MTLFunction>             function_soft_max;
+    id<MTLComputePipelineState> pipeline_soft_max;
+};
+
+// MSL code
+NSString * const msl_library_mnist = @"\
+#include <metal_stdlib>                                                                 \n\
+using namespace metal;                                                                  \n\
+                                                                                        \n\
+#define MAX(x, y) ((x) > (y) ? (x) : (y))                                               \n\
+                                                                                        \n\
+constant int k_digits [[function_constant(0)]];                                         \n\
+                                                                                        \n\
+kernel void kernel_add(                                                                 \n\
+        device const float * src0,                                                      \n\
+        device const float * src1,                                                      \n\
+        device float * dst,                                                             \n\
+        uint gid[[thread_position_in_grid]]) {                                          \n\
+    dst[gid] = src0[gid] + src1[gid];                                                   \n\
+}                                                                                       \n\
+                                                                                        \n\
+kernel void kernel_relu(                                                                \n\
+        device const float * src,                                                       \n\
+        device       float * dst,                                                       \n\
+        uint gid[[thread_position_in_grid]]) {                                          \n\
+    dst[gid] = max(0.0f, src[gid]);                                                     \n\
+}                                                                                       \n\
+                                                                                        \n\
+kernel void kernel_soft_max(                                                            \n\
+        device const float * src,                                                       \n\
+        device       float * dst,                                                       \n\
+        uint gid[[thread_position_in_grid]]) {                                          \n\
+    float max = 0.0f;                                                                   \n\
+    for (int i = 0; i < k_digits; i++) {                                                \n\
+        max = MAX(max, src[i]);                                                         \n\
+    }                                                                                   \n\
+    float sum = 0.0f;                                                                   \n\
+    for (int i = 0; i < k_digits; i++) {                                                \n\
+        dst[i] = exp(src[i] - max);                                                     \n\
+        sum += dst[i];                                                                  \n\
+    }                                                                                   \n\
+    for (int i = 0; i < k_digits; i++) {                                                \n\
+        dst[i] /= sum;                                                                  \n\
+    }                                                                                   \n\
+}                                                                                       \n\
+";
+
+struct ggml_mtl_context * mnist_mtl_init(
+    struct ggml_context * ctx_data,
+    struct ggml_context * ctx_eval,
+    struct ggml_context * ctx_work,
+    struct ggml_cgraph  * gf) {
+    fprintf(stderr, "%s: allocating\n", __func__);
+
+    struct ggml_mtl_context * ctx = malloc(sizeof(struct ggml_mtl_context));
+
+    ctx->ctx_data = ctx_data;
+    ctx->ctx_eval = ctx_eval;
+    ctx->ctx_work = ctx_work;
+
+    ctx->device = MTLCreateSystemDefaultDevice();
+    ctx->queue  = [ctx->device newCommandQueue];
+
+    // determine if we can use MPS
+    if (MPSSupportsMTLDevice(ctx->device)) {
+        fprintf(stderr, "%s: using MPS\n", __func__);
+    } else {
+        fprintf(stderr, "%s: not using MPS\n", __func__);
+        GGML_ASSERT(false && "MPS not supported");
+    }
+
+    // compile from source string and show compile log
+    {
+        NSError * error = nil;
+        ctx->library = [ctx->device newLibraryWithSource:msl_library_mnist options:nil error:&error];
+        if (error) {
+            fprintf(stderr, "%s: error: %s\n", __func__, [[error description] UTF8String]);
+            exit(1);
+        }
+    }
+
+    // load kernels
+    {
+        const int k_digits = ggml_graph_get_tensor(gf, "probs")->ne[0];
+
+        MTLFunctionConstantValues * constants = [MTLFunctionConstantValues new];
+        [constants setConstantValue:&k_digits type:MTLDataTypeInt withName:@"k_digits"];
+
+        ctx->function_add = [ctx->library newFunctionWithName:@"kernel_add"];
+        ctx->pipeline_add = [ctx->device newComputePipelineStateWithFunction:ctx->function_add error:nil];
+        fprintf(stderr, "%s: loaded kernel_add: %p\n", __func__, ctx->pipeline_add);
+
+        ctx->function_relu = [ctx->library newFunctionWithName:@"kernel_relu"];
+        ctx->pipeline_relu = [ctx->device newComputePipelineStateWithFunction:ctx->function_relu error:nil];
+        fprintf(stderr, "%s: loaded kernel_relu: %p\n", __func__, ctx->pipeline_relu);
+
+        ctx->function_soft_max = [ctx->library newFunctionWithName:@"kernel_soft_max" constantValues:constants error:nil];
+        ctx->pipeline_soft_max = [ctx->device newComputePipelineStateWithFunction:ctx->function_soft_max error:nil];
+        fprintf(stderr, "%s: loaded kernel_soft_max: %p\n", __func__, ctx->pipeline_soft_max);
+    }
+
+#ifdef GGML_MTL_HEAP
+    // MTLHeap approach
+
+    // pin ctx_data memory to GPU
+    // use MTLStorageModeShared to allow us to initialize the weights from the CPU
+    // TODO: how to use MTLStorageModeManaged?
+    // TODO: see if we can avoid this copy somehow
+    {
+        const void * mem_buffer = ggml_get_mem_buffer(ctx_data);
+        const size_t mem_size   = ggml_get_mem_size(ctx_data);
+
+        MTLHeapDescriptor * heap_desc = [MTLHeapDescriptor new];
+        heap_desc.storageMode = MTLStorageModeShared;
+        heap_desc.size        = mem_size;
+
+        printf("heap_desc.size = %zu\n", mem_size);
+
+        ctx->heap_data = [ctx->device newHeapWithDescriptor:heap_desc];
+        [ctx->heap_data setPurgeableState:MTLPurgeableStateNonVolatile]; // TODO: is this needed?
+        ctx->heap_data.label = @"heap_data";
+
+        printf("ctx->heap_data.size = %zu\n", [ctx->heap_data size]);
+
+        id<MTLBuffer> buffer = [ctx->heap_data newBufferWithLength:mem_size options:MTLResourceStorageModeShared];
+        if (!buffer) {
+            fprintf(stderr, "%s: error: failed to allocate buffer\n", __func__);
+            exit(1);
+        }
+
+        // copy data from CPU to GPU
+        memcpy([buffer contents], mem_buffer, mem_size);
+
+        fprintf(stderr, "%s: allocated data heap, size = %zu\n", __func__, mem_size);
+    }
+
+    // pin ctx_eval memory to GPU
+    // this heap will be used for the intermediate results of the evaluation
+    {
+        const size_t mem_size = ggml_get_mem_size(ctx_eval);
+
+        MTLHeapDescriptor * heap_desc = [MTLHeapDescriptor new];
+        heap_desc.storageMode = MTLStorageModePrivate; // GPU only
+        heap_desc.size        = mem_size;
+
+        ctx->heap_eval = [ctx->device newHeapWithDescriptor:heap_desc];
+        [ctx->heap_eval setPurgeableState:MTLPurgeableStateNonVolatile]; // TODO: is this needed?
+
+        fprintf(stderr, "%s: allocated eval heap, size = %zu\n", __func__, mem_size);
+    }
+#else
+    // MTLBuffer approach
+
+    // pin ctx_data memory to GPU
+    // use MTLStorageModeShared to allow us to initialize the weights from the CPU
+    // TODO: how to use MTLStorageModeManaged?
+    // TODO: see if we can avoid this copy somehow
+    {
+        const void * mem_buffer = ggml_get_mem_buffer(ctx_data);
+        const size_t mem_size   = ggml_get_mem_size(ctx_data);
+
+        ctx->buffer_data = [ctx->device newBufferWithBytes:mem_buffer length:mem_size options:MTLResourceStorageModeShared];
+
+        fprintf(stderr, "%s: allocated data buffer, size = %zu\n", __func__, mem_size);
+    }
+
+    // pin ctx_eval memory to GPU
+    // this buffer will be used for the intermediate results of the evaluation
+    {
+        const size_t mem_size = ggml_get_mem_size(ctx_eval);
+
+        ctx->buffer_eval = [ctx->device newBufferWithLength:mem_size options:MTLResourceStorageModePrivate];
+
+        fprintf(stderr, "%s: allocated eval buffer, size = %zu\n", __func__, mem_size);
+    }
+#endif
+
+    // allocate buffer for result extraction
+    {
+        const size_t mem_size = ggml_nbytes(gf->nodes[gf->n_nodes - 1]);
+
+        ctx->out = [ctx->device newBufferWithLength:mem_size options:MTLResourceStorageModeShared];
+
+        fprintf(stderr, "%s: allocated out buffer, size = %zu\n", __func__, mem_size);
+    }
+
+    return ctx;
+}
+
+void mnist_mtl_free(struct ggml_mtl_context * ctx) {
+    fprintf(stderr, "%s: deallocating\n", __func__);
+
+    free(ctx);
+}
+
+#ifdef GGML_MTL_HEAP
+
+// make a view of the respective MTL heap
+id<MTLBuffer> mnist_mtl_get_buffer_on_heap(struct ggml_mtl_context * ctx, struct ggml_tensor * t) {
+    const int64_t offs_data = (int64_t) t->data - (int64_t) ggml_get_mem_buffer(ctx->ctx_data);
+    const int64_t offs_eval = (int64_t) t->data - (int64_t) ggml_get_mem_buffer(ctx->ctx_eval);
+
+    const bool is_data = (offs_eval < 0) || (offs_data >= 0 && offs_data < offs_eval);
+
+    const size_t t_size = ggml_nbytes(t);
+    const size_t t_offs = is_data ? offs_data : offs_eval;
+
+    id<MTLBuffer> result;
+
+    if (is_data) {
+        fprintf(stderr, "%s: data tensor '%16s', offs = %8ld, size = %8ld\n", __func__, t->name, t_offs, t_size);
+        result = [ctx->heap_data newBufferWithLength:t_size options:MTLResourceStorageModeShared offset:t_offs];
+    } else {
+        fprintf(stderr, "%s: eval tensor '%16s', offs = %8ld, size = %8ld\n", __func__, t->name, t_offs, t_size);
+        result = [ctx->heap_eval newBufferWithLength:t_size options:MTLResourceStorageModePrivate offset:t_offs];
+    }
+
+    if (result == nil) {
+        fprintf(stderr, "%s: error: buffer is nil\n", __func__);
+        GGML_ASSERT(false);
+    }
+
+    return result;
+}
+
+#else
+
+// get data / eval buffer + offset
+id<MTLBuffer> mnist_mtl_get_buffer(struct ggml_mtl_context * ctx, struct ggml_tensor * t, size_t * offs) {
+    const int64_t offs_data = (int64_t) t->data - (int64_t) ggml_get_mem_buffer(ctx->ctx_data);
+    const int64_t offs_eval = (int64_t) t->data - (int64_t) ggml_get_mem_buffer(ctx->ctx_eval);
+
+    const bool is_data = (offs_eval < 0) || (offs_data >= 0 && offs_data < offs_eval);
+
+    const size_t t_size = ggml_nbytes(t);
+    const size_t t_offs = is_data ? offs_data : offs_eval;
+
+    id<MTLBuffer> result;
+
+    if (is_data) {
+        fprintf(stderr, "%s: data tensor '%16s', offs = %8ld, size = %8ld\n", __func__, t->name, t_offs, t_size);
+        result = ctx->buffer_data;
+    } else {
+        fprintf(stderr, "%s: eval tensor '%16s', offs = %8ld, size = %8ld\n", __func__, t->name, t_offs, t_size);
+        result = ctx->buffer_eval;
+    }
+
+    if (result == nil) {
+        fprintf(stderr, "%s: error: buffer is nil\n", __func__);
+        GGML_ASSERT(false);
+    }
+
+    if (offs != nil) {
+        *offs = t_offs;
+    }
+
+    return result;
+}
+
+#endif
+
+int mnist_mtl_eval(
+        struct ggml_mtl_context * ctx,
+        struct ggml_cgraph      * gf) {
+    fprintf(stderr, "%s: evaluating\n", __func__);
+
+    id<MTLCommandBuffer> command_buffer  = [ctx->queue commandBuffer];
+    id<MTLComputeCommandEncoder> encoder = nil;
+
+    size_t offs_src0;
+    size_t offs_src1;
+    size_t offs_dst;
+
+    // copy the input data to the GPU
+    {
+        struct ggml_tensor * inp = ggml_graph_get_tensor(gf, "input");
+
+        id<MTLBuffer> id_dst = mnist_mtl_get_buffer(ctx, inp, &offs_src0);
+
+        memcpy(id_dst.contents + offs_src0, inp->data, ggml_nbytes(inp));
+    }
+
+    for (int i = 0; i < gf->n_nodes; ++i) {
+        fprintf(stderr, "%s: encoding node %3d, op = %8s\n", __func__, i, ggml_op_name(gf->nodes[i]->op));
+
+        switch (gf->nodes[i]->op) {
+            case GGML_OP_ADD:
+                {
+                    if (encoder == nil) {
+                        encoder = [command_buffer computeCommandEncoder];
+                    }
+
+                    id<MTLBuffer> id_src0 = mnist_mtl_get_buffer(ctx, gf->nodes[i]->src[0], &offs_src0);
+                    id<MTLBuffer> id_src1 = mnist_mtl_get_buffer(ctx, gf->nodes[i]->src[1], &offs_src1);
+                    id<MTLBuffer> id_dst  = mnist_mtl_get_buffer(ctx, gf->nodes[i],         &offs_dst);
+
+                    [encoder setComputePipelineState:ctx->pipeline_add];
+                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                    [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
+                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
+
+                    const int64_t n = ggml_nelements(gf->nodes[i]);
+
+                    [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                } break;
+            case GGML_OP_UNARY:
+                switch (ggml_get_unary_op(gf->nodes[i])) {
+                    case GGML_UNARY_OP_RELU:
+                        {
+                            if (encoder == nil) {
+                                encoder = [command_buffer computeCommandEncoder];
+                            }
+
+                            id<MTLBuffer> id_src = mnist_mtl_get_buffer(ctx, gf->nodes[i]->src[0], &offs_src0);
+                            id<MTLBuffer> id_dst = mnist_mtl_get_buffer(ctx, gf->nodes[i],       &offs_dst);
+
+                            [encoder setComputePipelineState:ctx->pipeline_relu];
+                            [encoder setBuffer:id_src offset:offs_src0 atIndex:0];
+                            [encoder setBuffer:id_dst offset:offs_dst  atIndex:1];
+
+                            const int64_t n = ggml_nelements(gf->nodes[i]);
+
+                            [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                        } break;
+                    default:
+                        {
+                            fprintf(stderr, "%s: node %3d, op = %8s, unary op %d not implemented\n", __func__, i, ggml_op_name(gf->nodes[i]->op), (int) ggml_get_unary_op(gf->nodes[i]));
+                            GGML_ASSERT(false);
+                            return -1;
+                        }
+                        break;
+                } break;
+            case GGML_OP_SOFT_MAX:
+                {
+#if 0
+                    // NOTE: MPSMatrixSoftMax is not working properly, probably there is a bug
+
+                    if (encoder != nil) {
+                        [encoder endEncoding];
+                        encoder = nil;
+                    }
+
+                    // use MPSMatrixSoftMax
+                    id<MTLBuffer> id_src = mnist_mtl_get_buffer(ctx, gf->nodes[i]->src0, &offs_src0);
+                    id<MTLBuffer> id_dst = mnist_mtl_get_buffer(ctx, gf->nodes[i],       &offs_dst);
+
+                    MPSMatrixDescriptor * desc = [MPSMatrixDescriptor
+                        matrixDescriptorWithRows:1 columns:gf->nodes[i]->ne[0] rowBytes:gf->nodes[i]->nb[1] dataType:MPSDataTypeFloat32];
+
+                    MPSMatrix * mat_src = [[MPSMatrix alloc] initWithBuffer:id_src offset:offs_src0 descriptor:desc];
+                    MPSMatrix * mat_dst = [[MPSMatrix alloc] initWithBuffer:id_dst offset:offs_dst  descriptor:desc];
+
+                    MPSMatrixSoftMax * softmax = [[MPSMatrixSoftMax alloc] initWithDevice:ctx->device];
+
+                    [softmax encodeToCommandBuffer:command_buffer inputMatrix:mat_src resultMatrix:mat_dst];
+#else
+                    if (encoder == nil) {
+                        encoder = [command_buffer computeCommandEncoder];
+                    }
+
+                    id<MTLBuffer> id_src = mnist_mtl_get_buffer(ctx, gf->nodes[i]->src[0], &offs_src0);
+                    id<MTLBuffer> id_dst = mnist_mtl_get_buffer(ctx, gf->nodes[i],       &offs_dst);
+
+                    [encoder setComputePipelineState:ctx->pipeline_soft_max];
+                    [encoder setBuffer:id_src offset:offs_src0 atIndex:0];
+                    [encoder setBuffer:id_dst offset:offs_dst  atIndex:1];
+
+                    [encoder dispatchThreadgroups:MTLSizeMake(1, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+#endif
+                } break;
+            case GGML_OP_MUL_MAT:
+                {
+                    if (encoder != nil) {
+                        [encoder endEncoding];
+                        encoder = nil;
+                    }
+
+                    // use MPSMatrixMultiplication
+                    id<MTLBuffer> id_src0 = mnist_mtl_get_buffer(ctx, gf->nodes[i]->src[0], &offs_src0);
+                    id<MTLBuffer> id_src1 = mnist_mtl_get_buffer(ctx, gf->nodes[i]->src[1], &offs_src1);
+                    id<MTLBuffer> id_dst  = mnist_mtl_get_buffer(ctx, gf->nodes[i],         &offs_dst);
+
+                    const int64_t ncols0 = gf->nodes[i]->src[0]->ne[0];
+                    const int64_t nrows0 = gf->nodes[i]->src[0]->ne[1];
+
+                    const int64_t ncols1 = gf->nodes[i]->src[1]->ne[0];
+                    const int64_t nrows1 = gf->nodes[i]->src[1]->ne[1];
+
+                    const int64_t ncols2 = gf->nodes[i]->ne[0];
+                    const int64_t nrows2 = gf->nodes[i]->ne[1];
+
+                    GGML_ASSERT(ncols0 == ncols1);
+
+                    MPSMatrixDescriptor * desc0 = [MPSMatrixDescriptor
+                        matrixDescriptorWithRows:nrows0 columns:ncols0 rowBytes:gf->nodes[i]->src[0]->nb[1] dataType:MPSDataTypeFloat32];
+                    MPSMatrixDescriptor * desc1 = [MPSMatrixDescriptor
+                        matrixDescriptorWithRows:nrows1 columns:ncols1 rowBytes:gf->nodes[i]->src[1]->nb[1] dataType:MPSDataTypeFloat32];
+                    MPSMatrixDescriptor * desc2 = [MPSMatrixDescriptor
+                        matrixDescriptorWithRows:nrows2 columns:ncols2 rowBytes:gf->nodes[i]->nb[1] dataType:MPSDataTypeFloat32];
+
+                    MPSMatrix * mat_src0 = [[MPSMatrix alloc] initWithBuffer:id_src0 offset:offs_src0 descriptor:desc0];
+                    MPSMatrix * mat_src1 = [[MPSMatrix alloc] initWithBuffer:id_src1 offset:offs_src1 descriptor:desc1];
+                    MPSMatrix * mat_dst  = [[MPSMatrix alloc] initWithBuffer:id_dst  offset:offs_dst  descriptor:desc2];
+
+                    MPSMatrixMultiplication * mul = [[MPSMatrixMultiplication alloc] initWithDevice:ctx->device
+                        transposeLeft:false transposeRight:true resultRows:nrows1 resultColumns:nrows0 interiorColumns:ncols0 alpha:1.0 beta:0.0];
+
+                    [mul encodeToCommandBuffer:command_buffer leftMatrix:mat_src1 rightMatrix:mat_src0 resultMatrix:mat_dst];
+                } break;
+            default:
+                {
+                    fprintf(stderr, "%s: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(gf->nodes[i]->op));
+                    GGML_ASSERT(false);
+                    return -1;
+                }
+        }
+    }
+
+    // extract results from the GPU
+    {
+        if (encoder != nil) {
+            [encoder endEncoding];
+            encoder = nil;
+        }
+
+        struct ggml_tensor * out = gf->nodes[gf->n_nodes - 1];
+
+        id<MTLBuffer> id_src = mnist_mtl_get_buffer(ctx, out, &offs_src0);
+        id<MTLBuffer> id_dst = ctx->out;
+
+        id<MTLBlitCommandEncoder> encoder_blit = [command_buffer blitCommandEncoder];
+        [encoder_blit copyFromBuffer:id_src sourceOffset:offs_src0 toBuffer:id_dst destinationOffset:0 size:ggml_nbytes(out)];
+        [encoder_blit endEncoding];
+    }
+
+    [command_buffer commit];
+    [command_buffer waitUntilCompleted];
+
+    {
+        const double time_elapsed = [command_buffer GPUEndTime] - [command_buffer GPUStartTime];
+        fprintf(stderr, "%s: time elapsed = %f\n", __func__, time_elapsed);
+    }
+
+    // select the most probable digit
+    int result = -1;
+    {
+        const float * probs = ctx->out.contents;
+
+        float prob = probs[0];
+
+        for (int i = 0; i < 10; ++i) {
+            fprintf(stderr, "%s: probs[%2d] = %f\n", __func__, i, probs[i]);
+
+            if (probs[i] > prob) {
+                result = i;
+                prob = probs[i];
+            }
+        }
+    }
+
+    return result;
+}

ggml/examples/mnist/main.cpp

@@ -0,0 +1,328 @@
+#include "ggml/ggml.h"
+
+#include "common.h"
+
+#include <cmath>
+#include <cstdio>
+#include <cstring>
+#include <ctime>
+#include <fstream>
+#include <string>
+#include <vector>
+#include <algorithm>
+
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
+// default hparams
+struct mnist_hparams {
+    int32_t n_input   = 784;
+    int32_t n_hidden  = 500;
+    int32_t n_classes = 10;
+};
+
+struct mnist_model {
+    mnist_hparams hparams;
+
+    struct ggml_tensor * fc1_weight;
+    struct ggml_tensor * fc1_bias;
+
+    struct ggml_tensor * fc2_weight;
+    struct ggml_tensor * fc2_bias;
+
+    struct ggml_context * ctx;
+};
+
+// load the model's weights from a file
+bool mnist_model_load(const std::string & fname, mnist_model & model) {
+    printf("%s: loading model from '%s'\n", __func__, fname.c_str());
+
+    auto fin = std::ifstream(fname, std::ios::binary);
+    if (!fin) {
+        fprintf(stderr, "%s: failed to open '%s'\n", __func__, fname.c_str());
+        return false;
+    }
+
+    // verify magic
+    {
+        uint32_t magic;
+        fin.read((char *) &magic, sizeof(magic));
+        if (magic != GGML_FILE_MAGIC) {
+            fprintf(stderr, "%s: invalid model file '%s' (bad magic)\n", __func__, fname.c_str());
+            return false;
+        }
+    }
+
+    auto & ctx = model.ctx;
+
+    size_t ctx_size = 0;
+
+    {
+        const auto & hparams = model.hparams;
+
+        const int n_input   = hparams.n_input;
+        const int n_hidden  = hparams.n_hidden;
+        const int n_classes = hparams.n_classes;
+
+        ctx_size += n_input * n_hidden * ggml_type_sizef(GGML_TYPE_F32); // fc1 weight
+        ctx_size +=           n_hidden * ggml_type_sizef(GGML_TYPE_F32); // fc1 bias
+
+        ctx_size += n_hidden * n_classes * ggml_type_sizef(GGML_TYPE_F32); // fc2 weight
+        ctx_size +=            n_classes * ggml_type_sizef(GGML_TYPE_F32); // fc2 bias
+
+        printf("%s: ggml ctx size = %6.2f MB\n", __func__, ctx_size/(1024.0*1024.0));
+    }
+
+    // create the ggml context
+    {
+        struct ggml_init_params params = {
+            /*.mem_size   =*/ ctx_size + 1024*1024,
+            /*.mem_buffer =*/ NULL,
+            /*.no_alloc   =*/ false,
+        };
+
+        model.ctx = ggml_init(params);
+        if (!model.ctx) {
+            fprintf(stderr, "%s: ggml_init() failed\n", __func__);
+            return false;
+        }
+    }
+
+    // Read FC1 layer 1
+    {
+        // Read dimensions
+        int32_t n_dims;
+        fin.read(reinterpret_cast<char *>(&n_dims), sizeof(n_dims));
+
+        {
+            int32_t ne_weight[2] = { 1, 1 };
+            for (int i = 0; i < n_dims; ++i) {
+                fin.read(reinterpret_cast<char *>(&ne_weight[i]), sizeof(ne_weight[i]));
+            }
+
+            // FC1 dimensions taken from file, eg. 768x500
+            model.hparams.n_input  = ne_weight[0];
+            model.hparams.n_hidden = ne_weight[1];
+
+            model.fc1_weight = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, model.hparams.n_input, model.hparams.n_hidden);
+            fin.read(reinterpret_cast<char *>(model.fc1_weight->data), ggml_nbytes(model.fc1_weight));
+            ggml_set_name(model.fc1_weight, "fc1_weight");
+        }
+
+        {
+            int32_t ne_bias[2] = { 1, 1 };
+            for (int i = 0; i < n_dims; ++i) {
+                fin.read(reinterpret_cast<char *>(&ne_bias[i]), sizeof(ne_bias[i]));
+            }
+
+            model.fc1_bias = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, model.hparams.n_hidden);
+            fin.read(reinterpret_cast<char *>(model.fc1_bias->data), ggml_nbytes(model.fc1_bias));
+            ggml_set_name(model.fc1_bias, "fc1_bias");
+
+            // just for testing purposes, set some parameters to non-zero
+            model.fc1_bias->op_params[0] = 0xdeadbeef;
+        }
+    }
+
+    // Read FC2 layer 2
+    {
+        // Read dimensions
+        int32_t n_dims;
+        fin.read(reinterpret_cast<char *>(&n_dims), sizeof(n_dims));
+
+        {
+            int32_t ne_weight[2] = { 1, 1 };
+            for (int i = 0; i < n_dims; ++i) {
+                fin.read(reinterpret_cast<char *>(&ne_weight[i]), sizeof(ne_weight[i]));
+            }
+
+            // FC1 dimensions taken from file, eg. 10x500
+            model.hparams.n_classes = ne_weight[1];
+
+            model.fc2_weight = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, model.hparams.n_hidden, model.hparams.n_classes);
+            fin.read(reinterpret_cast<char *>(model.fc2_weight->data), ggml_nbytes(model.fc2_weight));
+            ggml_set_name(model.fc2_weight, "fc2_weight");
+        }
+
+        {
+            int32_t ne_bias[2] = { 1, 1 };
+            for (int i = 0; i < n_dims; ++i) {
+                fin.read(reinterpret_cast<char *>(&ne_bias[i]), sizeof(ne_bias[i]));
+            }
+
+            model.fc2_bias = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, model.hparams.n_classes);
+            fin.read(reinterpret_cast<char *>(model.fc2_bias->data), ggml_nbytes(model.fc2_bias));
+            ggml_set_name(model.fc2_bias, "fc2_bias");
+        }
+    }
+
+    fin.close();
+
+    return true;
+}
+
+// evaluate the model
+//
+//   - model:     the model
+//   - n_threads: number of threads to use
+//   - digit:     784 pixel values
+//
+// returns 0 - 9 prediction
+int mnist_eval(
+        const mnist_model & model,
+        const int n_threads,
+        std::vector<float> digit,
+        const char * fname_cgraph
+        ) {
+
+    const auto & hparams = model.hparams;
+
+    static size_t buf_size = hparams.n_input * sizeof(float) * 4;
+    static void * buf = malloc(buf_size);
+
+    struct ggml_init_params params = {
+        /*.mem_size   =*/ buf_size,
+        /*.mem_buffer =*/ buf,
+        /*.no_alloc   =*/ false,
+    };
+
+    struct ggml_context * ctx0 = ggml_init(params);
+    struct ggml_cgraph gf = {};
+
+    struct ggml_tensor * input = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, hparams.n_input);
+    memcpy(input->data, digit.data(), ggml_nbytes(input));
+    ggml_set_name(input, "input");
+
+    // fc1 MLP = Ax + b
+    ggml_tensor * fc1 = ggml_add(ctx0, ggml_mul_mat(ctx0, model.fc1_weight, input),                model.fc1_bias);
+    ggml_tensor * fc2 = ggml_add(ctx0, ggml_mul_mat(ctx0, model.fc2_weight, ggml_relu(ctx0, fc1)), model.fc2_bias);
+
+    // soft max
+    ggml_tensor * probs = ggml_soft_max(ctx0, fc2);
+    ggml_set_name(probs, "probs");
+
+    // build / export / run the computation graph
+    ggml_build_forward_expand(&gf, probs);
+    ggml_graph_compute_with_ctx(ctx0, &gf, n_threads);
+
+    //ggml_graph_print   (&gf);
+    ggml_graph_dump_dot(&gf, NULL, "mnist.dot");
+
+    if (fname_cgraph) {
+        // export the compute graph for later use
+        // see the "mnist-cpu" example
+        ggml_graph_export(&gf, "mnist.ggml");
+
+        fprintf(stderr, "%s: exported compute graph to '%s'\n", __func__, fname_cgraph);
+    }
+
+    const float * probs_data = ggml_get_data_f32(probs);
+
+    const int prediction = std::max_element(probs_data, probs_data + 10) - probs_data;
+
+    ggml_free(ctx0);
+
+    return prediction;
+}
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+int wasm_eval(uint8_t * digitPtr) {
+    mnist_model model;
+    if (!mnist_model_load("models/mnist/ggml-model-f32.bin", model)) {
+        fprintf(stderr, "error loading model\n");
+        return -1;
+    }
+    std::vector<float> digit(digitPtr, digitPtr + 784);
+    int result = mnist_eval(model, 1, digit, nullptr);
+    ggml_free(model.ctx);
+
+    return result;
+}
+
+int wasm_random_digit(char * digitPtr) {
+    auto fin = std::ifstream("models/mnist/t10k-images.idx3-ubyte", std::ios::binary);
+    if (!fin) {
+        fprintf(stderr, "failed to open digits file\n");
+        return 0;
+    }
+    srand(time(NULL));
+
+    // Seek to a random digit: 16-byte header + 28*28 * (random 0 - 10000)
+    fin.seekg(16 + 784 * (rand() % 10000));
+    fin.read(digitPtr, 784);
+
+    return 1;
+}
+
+#ifdef __cplusplus
+}
+#endif
+
+int main(int argc, char ** argv) {
+    srand(time(NULL));
+    ggml_time_init();
+
+    if (argc != 3) {
+        fprintf(stderr, "Usage: %s models/mnist/ggml-model-f32.bin models/mnist/t10k-images.idx3-ubyte\n", argv[0]);
+        exit(0);
+    }
+
+    uint8_t buf[784];
+    mnist_model model;
+    std::vector<float> digit;
+
+    // load the model
+    {
+        const int64_t t_start_us = ggml_time_us();
+
+        if (!mnist_model_load(argv[1], model)) {
+            fprintf(stderr, "%s: failed to load model from '%s'\n", __func__, "models/ggml-model-f32.bin");
+            return 1;
+        }
+
+        const int64_t t_load_us = ggml_time_us() - t_start_us;
+
+        fprintf(stdout, "%s: loaded model in %8.2f ms\n", __func__, t_load_us / 1000.0f);
+    }
+
+    // read a random digit from the test set
+    {
+        std::ifstream fin(argv[2], std::ios::binary);
+        if (!fin) {
+            fprintf(stderr, "%s: failed to open '%s'\n", __func__, argv[2]);
+            return 1;
+        }
+
+        // seek to a random digit: 16-byte header + 28*28 * (random 0 - 10000)
+        fin.seekg(16 + 784 * (rand() % 10000));
+        fin.read((char *) &buf, sizeof(buf));
+    }
+
+    // render the digit in ASCII
+    {
+        digit.resize(sizeof(buf));
+
+        for (int row = 0; row < 28; row++) {
+            for (int col = 0; col < 28; col++) {
+                fprintf(stderr, "%c ", (float)buf[row*28 + col] > 230 ? '*' : '_');
+                digit[row*28 + col] = ((float)buf[row*28 + col]);
+            }
+
+            fprintf(stderr, "\n");
+        }
+
+        fprintf(stderr, "\n");
+    }
+
+    const int prediction = mnist_eval(model, 1, digit, "mnist.ggml");
+
+    fprintf(stdout, "%s: predicted digit is %d\n", __func__, prediction);
+
+    ggml_free(model.ctx);
+
+    return 0;
+}

ggml/examples/mnist/mnist-cnn.py

@@ -0,0 +1,101 @@
+#!/usr/bin/env python3
+import sys
+import gguf
+import numpy as np
+from tensorflow import keras
+from tensorflow.keras import layers
+
+def train(model_name):
+    # Model / data parameters
+    num_classes = 10
+    input_shape = (28, 28, 1)
+
+    # Load the data and split it between train and test sets
+    (x_train, y_train), (x_test, y_test) = keras.datasets.mnist.load_data()
+
+    # Scale images to the [0, 1] range
+    x_train = x_train.astype("float32") / 255
+    x_test = x_test.astype("float32") / 255
+    # Make sure images have shape (28, 28, 1)
+    x_train = np.expand_dims(x_train, -1)
+    x_test = np.expand_dims(x_test, -1)
+    print("x_train shape:", x_train.shape)
+    print(x_train.shape[0], "train samples")
+    print(x_test.shape[0], "test samples")
+
+    # convert class vectors to binary class matrices
+    y_train = keras.utils.to_categorical(y_train, num_classes)
+    y_test = keras.utils.to_categorical(y_test, num_classes)
+
+    model = keras.Sequential(
+        [
+            keras.Input(shape=input_shape),
+            layers.Conv2D(32, kernel_size=(3, 3), activation="relu"),
+            layers.MaxPooling2D(pool_size=(2, 2)),
+            layers.Conv2D(64, kernel_size=(3, 3), activation="relu"),
+            layers.MaxPooling2D(pool_size=(2, 2)),
+            layers.Flatten(),
+            layers.Dropout(0.5),
+            layers.Dense(num_classes, activation="softmax"),
+        ]
+    )
+
+    model.summary()
+    batch_size = 128
+    epochs = 15
+    model.compile(loss="categorical_crossentropy", optimizer="adam", metrics=["accuracy"])
+    model.fit(x_train, y_train, batch_size=batch_size, epochs=epochs, validation_split=0.1)
+
+    score = model.evaluate(x_test, y_test, verbose=0)
+    print("Test loss:", score[0])
+    print("Test accuracy:", score[1])
+    model.save(model_name)
+    print("Keras model saved to '" + model_name + "'")
+
+def convert(model_name):
+    model = keras.models.load_model(model_name)
+    gguf_model_name = model_name + ".gguf"
+    gguf_writer = gguf.GGUFWriter(gguf_model_name, "mnist-cnn")
+
+    kernel1 = model.layers[0].weights[0].numpy()
+    kernel1 = np.moveaxis(kernel1, [2,3], [0,1])
+    kernel1 = kernel1.astype(np.float16)
+    gguf_writer.add_tensor("kernel1", kernel1, raw_shape=(32, 1, 3, 3))
+
+    bias1 = model.layers[0].weights[1].numpy()
+    bias1 = np.repeat(bias1, 26*26)
+    gguf_writer.add_tensor("bias1", bias1, raw_shape=(1, 32, 26, 26))
+
+    kernel2 = model.layers[2].weights[0].numpy()
+    kernel2 = np.moveaxis(kernel2, [0,1,2,3], [2,3,1,0])
+    kernel2 = kernel2.astype(np.float16)
+    gguf_writer.add_tensor("kernel2", kernel2, raw_shape=(64, 32, 3, 3))
+
+    bias2 = model.layers[2].weights[1].numpy()
+    bias2 = np.repeat(bias2, 11*11)
+    gguf_writer.add_tensor("bias2", bias2, raw_shape=(1, 64, 11, 11))
+
+    dense_w = model.layers[-1].weights[0].numpy()
+    dense_w = dense_w.transpose()
+    gguf_writer.add_tensor("dense_w", dense_w, raw_shape=(10, 1600))
+
+    dense_b = model.layers[-1].weights[1].numpy()
+    gguf_writer.add_tensor("dense_b", dense_b)
+
+    gguf_writer.write_header_to_file()
+    gguf_writer.write_kv_data_to_file()
+    gguf_writer.write_tensors_to_file()
+    gguf_writer.close()
+    print("Model converted and saved to '{}'".format(gguf_model_name))
+
+if __name__ == '__main__':
+    if len(sys.argv) < 3:
+        print("Usage: %s <train|convert> <model_name>".format(sys.argv[0]))
+        sys.exit(1)
+    if sys.argv[1] == 'train':
+        train(sys.argv[2])
+    elif sys.argv[1] == 'convert':
+        convert(sys.argv[2])
+    else:
+        print("Usage: %s <train|convert> <model_name>".format(sys.argv[0]))
+        sys.exit(1)

ggml/examples/mnist/models/mnist/.gitignore

@@ -0,0 +1 @@
+ggml-model-f32.bin

ggml/examples/mpt/CMakeLists.txt

@@ -0,0 +1,13 @@
+#
+# mpt
+
+set(TEST_TARGET mpt)
+add_executable(${TEST_TARGET} main.cpp)
+target_link_libraries(${TEST_TARGET} PRIVATE ggml common common-ggml)
+
+#
+# mpt-quantize
+
+set(TEST_TARGET mpt-quantize)
+add_executable(${TEST_TARGET} quantize.cpp)
+target_link_libraries(${TEST_TARGET} PRIVATE ggml common common-ggml)

ggml/examples/mpt/README.md

@@ -0,0 +1,27 @@
+# MPT
+
+Ref: https://github.com/mosaicml/llm-foundry#mpt
+
+## Usage
+
+```bash
+# get the repo and build it
+git clone https://github.com/ggerganov/ggml
+cd ggml
+mkdir build && cd build
+cmake ..
+make -j
+
+# get the model from HuggingFace
+# be sure to have git-lfs installed
+git clone https://huggingface.co/mosaicml/mpt-30b
+
+# convert model to FP16
+python3 ../examples/mpt/convert-h5-to-ggml.py ./mpt-30b 1
+
+# run inference using FP16 precision
+./bin/mpt -m ./mpt-30b/ggml-model-f16.bin -p "I believe the meaning of life is" -t 8 -n 64
+
+# quantize the model to 5-bits using Q5_0 quantization
+./bin/mpt-quantize ./mpt-30b/ggml-model-f16.bin ./mpt-30b/ggml-model-q5_0.bin q5_0
+```

ggml/examples/mpt/convert-h5-to-ggml.py

@@ -0,0 +1,169 @@
+import os
+import struct
+import sys
+
+import torch
+from transformers import AutoConfig, AutoTokenizer
+
+
+# ref: https://github.com/openai/gpt-2/blob/master/src/encoder.py
+def bytes_to_unicode():
+    """
+    Returns list of utf-8 byte and a corresponding list of unicode strings.
+    The reversible bpe codes work on unicode strings.
+    This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.
+    When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.
+    This is a signficant percentage of your normal, say, 32K bpe vocab.
+    To avoid that, we want lookup tables between utf-8 bytes and unicode strings.
+    And avoids mapping to whitespace/control characters the bpe code barfs on.
+    """
+    bs = (
+        list(range(ord("!"), ord("~") + 1))
+        + list(range(ord("¡"), ord("¬") + 1))
+        + list(range(ord("®"), ord("ÿ") + 1))
+    )
+    cs = bs[:]
+    n = 0
+    for b in range(2**8):
+        if b not in bs:
+            bs.append(b)
+            cs.append(2**8 + n)
+            n += 1
+
+    cs = [chr(n) for n in cs]
+
+    return dict(zip(bs, cs))
+
+
+def count_model_parts(dir_model: str) -> int:
+    """Returns the number of model parts in the model directory."""
+    num_parts = 0
+    for filename in os.listdir(dir_model):
+        if filename.startswith("pytorch_model-"):
+            num_parts += 1
+
+    if num_parts > 0:
+        print(f"Found {num_parts} model parts in {dir_model}")
+    return num_parts
+
+
+if len(sys.argv) < 3:
+    print("Usage: convert-h5-to-ggml.py dir-model [use-f32]\n")
+    print("  ftype == 0 -> float32")
+    print("  ftype == 1 -> float16")
+    sys.exit(1)
+
+
+# output in the same directory as the model
+dir_model = sys.argv[1]
+# get number of model parts
+num_parts = count_model_parts(dir_model)
+
+# possible data types
+#   ftype == 0 -> float32
+#   ftype == 1 -> float16
+#
+# map from ftype to string
+ftype_str = ["f32", "f16"]
+
+ftype = 1
+if len(sys.argv) > 2:
+    ftype = int(sys.argv[2])
+    if ftype < 0 or ftype > 1:
+        print("Invalid ftype: " + str(ftype))
+        sys.exit(1)
+    fname_out = dir_model + "/ggml-model-" + ftype_str[ftype] + ".bin"
+
+
+tokenizer = AutoTokenizer.from_pretrained(dir_model, trust_remote_code=True)
+config = AutoConfig.from_pretrained(dir_model, trust_remote_code=True)
+hparams = config.to_dict()
+
+fout = open(fname_out, "wb")
+
+fout.write(struct.pack("i", 0x67676D6C))  # magic: ggml in hex
+fout.write(struct.pack("i", hparams["d_model"]))
+fout.write(struct.pack("i", hparams["max_seq_len"]))
+fout.write(struct.pack("i", hparams["n_heads"]))
+fout.write(struct.pack("i", hparams["n_layers"]))
+fout.write(struct.pack("i", hparams["vocab_size"]))
+fout.write(struct.pack("f", hparams["attn_config"]["alibi_bias_max"]))
+fout.write(struct.pack("f", hparams["attn_config"]["clip_qkv"] or 0.0))
+fout.write(struct.pack("i", ftype))
+
+vocab_size = hparams["vocab_size"]
+
+encoder = tokenizer.vocab
+# Add added_tokens (special tokens) to the encoder
+encoder.update(tokenizer.get_added_vocab())
+
+byte_encoder = bytes_to_unicode()
+byte_decoder = {v: k for k, v in byte_encoder.items()}
+
+counter = 0
+# sort by value
+for key in sorted(encoder, key=encoder.get):
+    # workaround for key error when c not found
+    text = ""
+    for c in key:
+        if c not in byte_decoder:
+            text += c
+        else:
+            text += chr(byte_decoder[c])
+    text = bytearray(text, encoding="utf-8")
+    fout.write(struct.pack("i", len(text)))
+    fout.write(text)
+    counter += 1
+
+# Repeat last token until vocab_size
+while counter < vocab_size:
+    fout.write(struct.pack("i", len(text)))
+    fout.write(text)
+    counter += 1
+
+if num_parts == 0:
+    part_names = ("pytorch_model.bin",)
+else:
+    part_names = (
+        f"pytorch_model-{n:05}-of-{num_parts:05}.bin" for n in range(1, num_parts + 1)
+    )
+
+for part_name in part_names:
+    print(f"\n* Loading part: {part_name}")
+    model_part = torch.load(f"{dir_model}/{part_name}", map_location="cpu")
+
+    for name in model_part.keys():
+        data = model_part[name].squeeze()
+        n_dims = len(data.shape)
+
+        # ftype == 0 -> float32, ftype == 1 -> float16
+        # default type is fp32
+        ftype_cur = 0
+        if ftype == 1 and name[-7:] == ".weight" and n_dims > 1:
+            ftype_cur = 1
+        data = data.to(dtype=torch.float16 if ftype_cur == 1 else torch.float32).numpy()
+
+        print(
+            "Processing variable: " + name + " with shape: ",
+            data.shape,
+            "->",
+            data.dtype,
+        )
+
+        # header
+        str = name.encode("utf-8")
+        fout.write(struct.pack("iii", n_dims, len(str), ftype_cur))
+        for i in range(n_dims):
+            fout.write(struct.pack("i", data.shape[n_dims - 1 - i]))
+        fout.write(str)
+
+        # data
+        data.tofile(fout)
+
+    # release memory
+    del model_part
+
+fout.close()
+
+print("Done. Output file: " + fname_out)
+print("")

ggml/examples/mpt/main.cpp

@@ -0,0 +1,1039 @@
+#include "ggml/ggml.h"
+
+#include "common-ggml.h"
+#include "common.h"
+
+#include <cmath>
+#include <cstddef>
+#include <cstdio>
+#include <cstring>
+#include <fstream>
+#include <cinttypes>
+#include <map>
+#include <string>
+#include <utility>
+#include <vector>
+
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
+// no defaults for now
+struct mpt_hparams {
+    int32_t d_model      = 0;
+    int32_t max_seq_len  = 0;
+    int32_t n_heads      = 0;
+    int32_t n_layers     = 0;
+    int32_t n_vocab      = 0;
+    float alibi_bias_max = 0;
+    float clip_qkv       = 0;
+    int32_t ftype        = 0;
+    int32_t n_ctx        = 0;
+
+};
+
+struct mpt_layer {
+    // pre normalization
+    struct ggml_tensor * norm_1_weight;
+
+    // attention
+    struct ggml_tensor * c_attn_wqkv_weight;
+    struct ggml_tensor * c_attn_out_proj_weight;
+
+    // post normalization
+    struct ggml_tensor * norm_2_weight;
+
+    // ff
+    struct ggml_tensor * ffn_up_proj;
+    struct ggml_tensor * ffn_down_proj;
+};
+
+struct mpt_model {
+    mpt_hparams hparams;
+
+    struct ggml_tensor * wte_weight;    // position embedding
+    struct ggml_tensor * norm_f_weight; // language model head
+
+    std::vector<mpt_layer> layers;
+
+    // key + value memory
+    struct ggml_tensor * memory_k;
+    struct ggml_tensor * memory_v;
+
+    struct ggml_context * ctx;
+    std::map<std::string, struct ggml_tensor *> tensors;
+};
+
+struct mpt_params {
+    int32_t n_threads = std::min(4, (int32_t) std::thread::hardware_concurrency());
+
+    int32_t seed           = -1; // RNG seed
+    int32_t n_predict      = 200; // new tokens to predict
+    int32_t n_batch        = 8; // batch size for prompt processing
+    int32_t n_ctx          = 512;
+
+    std::string model      = ""; // model path
+    std::string prompt     = "";
+    std::string token_test = "";
+
+    bool    perplexity     = false;
+
+    // sampling parameters
+    int32_t top_k          = 0;
+    float   top_p          = 1.0f;
+    float   temp           = 0.8f;
+    int32_t repeat_last_n  = 64;
+    float   repeat_penalty = 1.02f;
+
+};
+
+void mpt_print_usage(int /*argc*/, char ** argv, const mpt_params & params) {
+    fprintf(stderr, "usage: %s [options]\n", argv[0]);
+    fprintf(stderr, "\n");
+    fprintf(stderr, "options:\n");
+    fprintf(stderr, "  -h, --help            show this help message and exit\n");
+    fprintf(stderr, "  -s SEED, --seed SEED  RNG seed (default: -1)\n");
+    fprintf(stderr, "  -t N, --threads N     number of threads to use during computation (default: %d)\n", params.n_threads);
+    fprintf(stderr, "  -p PROMPT, --prompt PROMPT\n");
+    fprintf(stderr, "                        prompt to start generation with (default: random)\n");
+    fprintf(stderr, "  -f FNAME, --file FNAME\n");
+    fprintf(stderr, "                        load prompt from a file\n");
+    fprintf(stderr, "  -tt TOKEN_TEST, --token_test TOKEN_TEST\n");
+    fprintf(stderr, "                        test tokenization\n");
+    fprintf(stderr, "  -n N, --n_predict N   number of tokens to predict (default: %d)\n", params.n_predict);
+    fprintf(stderr, "  --top_k N             top-k sampling (default: %d, 0 = n_vocab)\n", params.top_k);
+    fprintf(stderr, "  --top_p N             top-p sampling (default: %.2f)\n", params.top_p);
+    fprintf(stderr, "  --temp N              temperature (default: %.2f)\n", params.temp);
+    fprintf(stderr, "  --repeat-last-n N     last n tokens to consider for penalize (default: %d, 0 = disabled, -1 = ctx_size)\n", params.repeat_last_n);
+    fprintf(stderr, "  --repeat-penalty N    penalize repeat sequence of tokens (default: %.2f, 1.0 = disabled)\n", (double)params.repeat_penalty);
+    fprintf(stderr, "  --perplexity          compute perplexity over the prompt\n");
+    fprintf(stderr, "  -c N, --ctx-size N    size of the prompt context (default: %d)\n", params.n_ctx);
+    fprintf(stderr, "  -b N, --batch_size N  batch size for prompt processing (default: %d)\n", params.n_batch);
+    fprintf(stderr, "  -m FNAME, --model FNAME\n");
+    fprintf(stderr, "                        model path (default: %s)\n", params.model.c_str());
+    fprintf(stderr, "\n");
+}
+
+bool mpt_params_parse(int argc, char ** argv, mpt_params & params) {
+    for (int i = 1; i < argc; i++) {
+        std::string arg = argv[i];
+
+        if (arg == "-s" || arg == "--seed") {
+            params.seed = std::stoi(argv[++i]);
+        } else if (arg == "-t" || arg == "--threads") {
+            params.n_threads = std::stoi(argv[++i]);
+        } else if (arg == "-p" || arg == "--prompt") {
+            params.prompt = argv[++i];
+        } else if (arg == "-n" || arg == "--n_predict") {
+            params.n_predict = std::stoi(argv[++i]);
+        } else if (arg == "--top_k") {
+            params.top_k = std::max(1, std::stoi(argv[++i]));
+        } else if (arg == "--top_p") {
+            params.top_p = std::stof(argv[++i]);
+        } else if (arg == "--temp") {
+            params.temp = std::stof(argv[++i]);
+        } else if (arg == "--repeat-last-n") {
+            params.repeat_last_n = std::stof(argv[++i]);
+        } else if (arg == "--repeat-penalty") {
+            params.repeat_penalty = std::stof(argv[++i]);
+        } else if (arg == "--perplexity") {
+            params.perplexity = true;
+        } else if (arg == "-c" || arg == "--ctx-size") {
+            params.n_ctx = std::stoi(argv[++i]);
+        } else if (arg == "-b" || arg == "--batch_size") {
+            params.n_batch = std::stoi(argv[++i]);
+        } else if (arg == "-m" || arg == "--model") {
+            params.model = argv[++i];
+        } else if (arg == "-h" || arg == "--help") {
+            mpt_print_usage(argc, argv, params);
+            exit(0);
+        } else if (arg == "-f" || arg == "--file") {
+            if (++i > argc) {
+                fprintf(stderr, "Invalid file param");
+                break;
+            }
+            std::ifstream file(argv[i]);
+            if (!file) {
+                fprintf(stderr, "error: failed to open file '%s'\n", argv[i]);
+                break;
+            }
+            params.prompt.clear();
+            std::copy(std::istreambuf_iterator<char>(file), std::istreambuf_iterator<char>(), back_inserter(params.prompt));
+            if (params.prompt.back() == '\n') {
+                params.prompt.pop_back();
+            }
+        } else if (arg == "-tt" || arg == "--token_test") {
+            params.token_test = argv[++i];
+        } else {
+            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
+            mpt_print_usage(argc, argv, params);
+            exit(0);
+        }
+    }
+
+    return true;
+}
+
+// load the model's weights from a file
+bool mpt_model_load(const std::string & fname, mpt_model & model, gpt_vocab & vocab) {
+    printf("%s: loading model from '%s' - please wait ...\n", __func__, fname.c_str());
+
+    auto fin = std::ifstream(fname, std::ios::binary);
+    if (!fin) {
+        fprintf(stderr, "%s: failed to open '%s'\n", __func__, fname.c_str());
+        return false;
+    }
+
+    // verify magic
+    {
+        uint32_t magic;
+        fin.read((char *)&magic, sizeof(magic));
+        if (magic != GGML_FILE_MAGIC) {
+            fprintf(stderr, "%s: invalid model file '%s' (bad magic)\n", __func__, fname.c_str());
+            return false;
+        }
+    }
+
+    // load hparams
+    {
+        auto & hparams = model.hparams;
+
+        fin.read((char *) &hparams.d_model,        sizeof(hparams.d_model));
+        fin.read((char *) &hparams.max_seq_len,    sizeof(hparams.max_seq_len));
+        fin.read((char *) &hparams.n_heads,        sizeof(hparams.n_heads));
+        fin.read((char *) &hparams.n_layers,       sizeof(hparams.n_layers));
+        fin.read((char *) &hparams.n_vocab,        sizeof(hparams.n_vocab));
+        fin.read((char *) &hparams.alibi_bias_max, sizeof(hparams.alibi_bias_max));
+        fin.read((char *) &hparams.clip_qkv,       sizeof(hparams.clip_qkv));
+        fin.read((char *) &hparams.ftype,          sizeof(hparams.ftype));
+
+        hparams.n_ctx = std::min(hparams.max_seq_len, hparams.n_ctx);
+
+        const int32_t qntvr = hparams.ftype / GGML_QNT_VERSION_FACTOR;
+
+        printf("%s: d_model        = %d\n", __func__, hparams.d_model);
+        printf("%s: max_seq_len    = %d\n", __func__, hparams.max_seq_len);
+        printf("%s: n_ctx          = %d\n", __func__, hparams.n_ctx);
+        printf("%s: n_heads        = %d\n", __func__, hparams.n_heads);
+        printf("%s: n_layers       = %d\n", __func__, hparams.n_layers);
+        printf("%s: n_vocab        = %d\n", __func__, hparams.n_vocab);
+        printf("%s: alibi_bias_max = %f\n", __func__, hparams.alibi_bias_max);
+        printf("%s: clip_qkv       = %f\n", __func__, hparams.clip_qkv);
+        printf("%s: ftype          = %d\n", __func__, hparams.ftype);
+        printf("%s: qntvr          = %d\n", __func__, qntvr);
+
+        hparams.ftype %= GGML_QNT_VERSION_FACTOR;
+    }
+
+    // load vocab
+    {
+        const int32_t n_vocab = model.hparams.n_vocab;
+
+        std::string word;
+        std::vector<char> buf(128);
+
+        for (int i = 0; i < n_vocab; i++) {
+            uint32_t len;
+            fin.read((char *) &len, sizeof(len));
+
+            buf.resize(len);
+            fin.read((char *) buf.data(), len);
+            word.assign(buf.data(), len);
+
+            // Convert token from utf-8
+            std::wstring word_multibytes = convert_to_wstring(word);
+            word.resize(word_multibytes.size());
+            for (size_t w = 0; w < word_multibytes.size(); w++) {
+                word[w] = uint8_t(word_multibytes[w]);
+            }
+
+            vocab.token_to_id[word] = i;
+            vocab.id_to_token[i] = word;
+        }
+    }
+
+    // for the big tensors, we have the option to store the data in 16-bit
+    // floats or quantized in order to save memory and also to speed up the
+    // computation
+    ggml_type wtype = ggml_ftype_to_ggml_type((ggml_ftype)(model.hparams.ftype));
+    if (wtype == GGML_TYPE_COUNT) {
+        fprintf(stderr, "%s: invalid model file '%s' (bad ftype value %d)\n", __func__, fname.c_str(),
+                model.hparams.ftype);
+        return false;
+    }
+
+    auto & ctx = model.ctx;
+
+    size_t ctx_size = 0;
+
+    const auto & hparams = model.hparams;
+    const size_t n_ctx = hparams.n_ctx;
+
+    {
+        const size_t n_embd = hparams.d_model;
+        const size_t n_layer = hparams.n_layers;
+        const size_t n_vocab = hparams.n_vocab;
+
+        ctx_size += n_embd * n_vocab * ggml_type_sizef(wtype); // wte_weight
+        ctx_size += n_embd * ggml_type_sizef(GGML_TYPE_F32);   // norm_f_weight
+
+        ctx_size += n_layer * (n_embd * ggml_type_sizef(GGML_TYPE_F32));      // ln_1_weight
+        ctx_size += n_layer * (3 * n_embd * n_embd * ggml_type_sizef(wtype)); // attn_Wqkv_weight
+        ctx_size += n_layer * (n_embd * n_embd * ggml_type_sizef(wtype));     // attn_out_proj_weight
+        ctx_size += n_layer * (n_embd * ggml_type_sizef(GGML_TYPE_F32));      // ln_2_weight
+        ctx_size += n_layer * (4 * n_embd * n_embd * ggml_type_sizef(wtype)); // mlp_mlp_up_weight
+        ctx_size += n_layer * (n_embd * n_embd * 4 * ggml_type_sizef(wtype)); // mlp_mlp_down_weight
+
+        ctx_size += n_ctx * n_layer * n_embd * ggml_type_sizef(GGML_TYPE_F16); // memory_k
+        ctx_size += n_ctx * n_layer * n_embd * ggml_type_sizef(GGML_TYPE_F16); // memory_v
+
+        ctx_size += (1 + 6 * n_layer) * 512; // object overhead
+
+        printf("%s: ggml ctx size = %6.2f MB\n", __func__, ctx_size / (1024.0 * 1024.0));
+    }
+
+    // create the ggml context
+    {
+        struct ggml_init_params params = {
+            /*.mem_size   =*/ ctx_size,
+            /*.mem_buffer =*/ NULL,
+            /*.no_alloc   =*/ false,
+        };
+
+        model.ctx = ggml_init(params);
+        if (!model.ctx) {
+            fprintf(stderr, "%s: ggml_init() failed\n", __func__);
+            return false;
+        }
+    }
+
+    // prepare memory for the weights
+    {
+        const auto & hparams = model.hparams;
+
+        const size_t n_embd = hparams.d_model;
+        const size_t n_layer = hparams.n_layers;
+        const size_t n_vocab = hparams.n_vocab;
+
+        model.layers.resize(n_layer);
+
+        model.wte_weight    = ggml_new_tensor_2d(ctx, wtype, n_embd, n_vocab);
+        model.norm_f_weight = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
+
+        // map by name
+        model.tensors["transformer.wte.weight"]    = model.wte_weight;
+        model.tensors["transformer.norm_f.weight"] = model.norm_f_weight;
+
+        for (int i = 0; i < (int) n_layer; ++i) {
+            auto & layer = model.layers[i];
+
+            layer.norm_1_weight          = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,     n_embd);
+            layer.c_attn_wqkv_weight     = ggml_new_tensor_2d(ctx, wtype,             n_embd, 3 * n_embd);
+            layer.c_attn_out_proj_weight = ggml_new_tensor_2d(ctx, wtype,             n_embd,     n_embd);
+            layer.norm_2_weight          = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,     n_embd);
+            layer.ffn_up_proj            = ggml_new_tensor_2d(ctx, wtype,             n_embd, 4 * n_embd);
+            layer.ffn_down_proj          = ggml_new_tensor_2d(ctx, wtype,         4 * n_embd,     n_embd);
+
+            // map by name
+            model.tensors["transformer.blocks." + std::to_string(i) + ".norm_1.weight"]        = layer.norm_1_weight;
+            model.tensors["transformer.blocks." + std::to_string(i) + ".attn.Wqkv.weight"]     = layer.c_attn_wqkv_weight;
+            model.tensors["transformer.blocks." + std::to_string(i) + ".attn.out_proj.weight"] = layer.c_attn_out_proj_weight;
+            model.tensors["transformer.blocks." + std::to_string(i) + ".norm_2.weight"]        = layer.norm_2_weight;
+            model.tensors["transformer.blocks." + std::to_string(i) + ".ffn.up_proj.weight"]   = layer.ffn_up_proj;
+            model.tensors["transformer.blocks." + std::to_string(i) + ".ffn.down_proj.weight"] = layer.ffn_down_proj;
+        }
+    }
+
+    // key + value memory
+    {
+        const auto & hparams = model.hparams;
+
+        const size_t n_embd  = hparams.d_model;
+        const size_t n_layer = hparams.n_layers;
+
+        const int64_t n_mem      = n_layer * n_ctx;
+        const int64_t n_elements = n_embd  * n_mem;
+
+        model.memory_k = ggml_new_tensor_1d(ctx, GGML_TYPE_F16, n_elements);
+        model.memory_v = ggml_new_tensor_1d(ctx, GGML_TYPE_F16, n_elements);
+
+        const size_t memory_size = ggml_nbytes(model.memory_k) + ggml_nbytes(model.memory_v);
+
+        printf("%s: memory_size = %8.2f MB, n_mem = %" PRId64 "\n", __func__, memory_size / 1024.0 / 1024.0, n_mem);
+    }
+
+    // load weights
+    {
+        int n_tensors = 0;
+        size_t total_size = 0;
+
+        printf("%s: ", __func__);
+
+        while (true) {
+            int32_t n_dims;
+            int32_t length;
+            int32_t ttype;
+
+            fin.read(reinterpret_cast<char *>(&n_dims), sizeof(n_dims));
+            fin.read(reinterpret_cast<char *>(&length), sizeof(length));
+            fin.read(reinterpret_cast<char *>(&ttype), sizeof(ttype));
+
+            if (fin.eof()) {
+                break;
+            }
+
+            int32_t nelements = 1;
+            int32_t ne[2] = {1, 1};
+            for (int i = 0; i < n_dims; ++i) {
+                fin.read(reinterpret_cast<char *>(&ne[i]), sizeof(ne[i]));
+                nelements *= ne[i];
+            }
+
+            std::string name(length, 0);
+            fin.read(&name[0], length);
+
+            if (model.tensors.find(name) == model.tensors.end()) {
+                fprintf(stderr, "%s: unknown tensor '%s' in model file\n", __func__, name.c_str());
+                return false;
+            }
+
+            auto tensor = model.tensors[name];
+            if (ggml_nelements(tensor) != nelements) {
+                fprintf(stderr, "%s: tensor '%s' has wrong size in model file\n", __func__, name.c_str());
+                return false;
+            }
+
+            if (tensor->ne[0] != ne[0] || tensor->ne[1] != ne[1]) {
+                fprintf(stderr,
+                        "%s: tensor '%s' has wrong shape in model file: got [%5d, "
+                        "%5d], expected [%5d, %5d]\n",
+                        __func__, name.c_str(), (int)tensor->ne[0], (int)tensor->ne[1], ne[0], ne[1]);
+                return false;
+            }
+
+            // for debugging
+            if (0) {
+                printf("%24s - [%5d, %5d], type = %6s, %6.2f MB, %9zu bytes\n", name.c_str(), ne[0], ne[1],
+                       ggml_type_name(ggml_type(ttype)), ggml_nbytes(tensor) / 1024.0 / 1024.0, ggml_nbytes(tensor));
+            }
+
+            const size_t bpe = ggml_type_size(ggml_type(ttype));
+
+            if ((nelements * bpe) / ggml_blck_size(tensor->type) != ggml_nbytes(tensor)) {
+                fprintf(stderr,
+                        "%s: tensor '%s' has wrong size in model file: got %zu, "
+                        "expected %zu\n",
+                        __func__, name.c_str(), ggml_nbytes(tensor), nelements * bpe);
+                return false;
+            }
+
+            fin.read(reinterpret_cast<char *>(tensor->data), ggml_nbytes(tensor));
+
+            total_size += ggml_nbytes(tensor);
+            if (++n_tensors % 8 == 0) {
+                printf(".");
+                fflush(stdout);
+            }
+        }
+
+        printf(" done\n");
+
+        printf("%s: model size = %8.2f MB / num tensors = %d\n", __func__, total_size / 1024.0 / 1024.0, n_tensors);
+    }
+
+    fin.close();
+
+    return true;
+}
+
+// evaluate the transformer
+//
+//   - model:     the model
+//   - n_threads: number of threads to use
+//   - n_past:    the context size so far
+//   - embd_inp:  the embeddings of the tokens in the context
+//   - embd_w:    the predicted logits for the next token
+//
+bool mpt_eval(const mpt_model & model, const int n_threads, const int n_past,
+              const std::vector<gpt_vocab::id> & embd_inp, std::vector<float> & embd_w, bool logits_all, size_t & mem_per_token) {
+    const int N = embd_inp.size();
+
+    const auto & hparams = model.hparams;
+
+    const int n_embd  = hparams.d_model;
+    const int n_layer = hparams.n_layers;
+    const int n_head  = hparams.n_heads;
+    const int n_vocab = hparams.n_vocab;
+    const int n_ctx   = hparams.n_ctx;
+    const float eps   = 1e-5f;
+
+    static size_t buf_size = 256u * 1024 * 1024;
+    static void * buf = malloc(buf_size);
+
+    // use 2 scratch buffers
+    // TODO: very hacky solution - reimplement in a more elegant way
+    static size_t scr0_size = 256u*1024*1024;
+    static void * scr0 = malloc(scr0_size);
+
+    static size_t scr1_size = 256u*1024*1024;
+    static void * scr1 = malloc(scr1_size);
+
+    if (mem_per_token > 0 && mem_per_token * N > buf_size) {
+        const size_t buf_size_new = 1.1 * (mem_per_token * N); // add 10% to account for ggml object overhead
+        // printf("\n%s: reallocating buffer from %zu to %zu bytes\n", __func__,
+        // buf_size, buf_size_new);
+
+        // reallocate
+        buf_size = buf_size_new;
+        buf = realloc(buf, buf_size);
+        if (buf == nullptr) {
+            fprintf(stderr, "%s: failed to allocate %zu bytes\n", __func__, buf_size);
+            return false;
+        }
+    }
+
+    struct ggml_init_params params = {
+        /*.mem_size   =*/ buf_size,
+        /*.mem_buffer =*/ buf,
+        /*.no_alloc   =*/ false,
+    };
+
+    struct ggml_context * ctx0 = ggml_init(params);
+    struct ggml_cgraph gf = {};
+
+    struct ggml_tensor * embd = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
+    memcpy(embd->data, embd_inp.data(), N * ggml_element_size(embd));
+
+    struct ggml_tensor * inpL = ggml_get_rows(ctx0, model.wte_weight, embd);
+
+    for (int il = 0; il < n_layer; ++il) {
+
+        struct ggml_tensor * cur;
+
+        ggml_set_scratch(ctx0, { 0, scr0_size, scr0, });
+
+        // a = self.ln_1(x)
+        {
+            cur = ggml_norm(ctx0, inpL, eps);
+
+            cur = ggml_mul(ctx0, ggml_repeat(ctx0, model.layers[il].norm_1_weight, cur), cur);
+        }
+
+        // self-attention
+        //  b, _, past_key_value = self.attn(a, past_key_value=past_key_value,
+        //  attn_bias=attn_bias, attention_mask=attention_mask,
+        //  is_causal=is_causal)
+        {
+            // compute QKV
+            cur = ggml_mul_mat(ctx0, model.layers[il].c_attn_wqkv_weight, cur);
+
+            if (model.hparams.clip_qkv > 0.0f) {
+                cur = ggml_clamp(ctx0, cur, -model.hparams.clip_qkv, model.hparams.clip_qkv);
+            }
+
+            struct ggml_tensor * Qcur = ggml_view_2d(ctx0, cur, n_embd, N, cur->nb[1], 0 * sizeof(float) * n_embd);
+            struct ggml_tensor * Kcur = ggml_view_2d(ctx0, cur, n_embd, N, cur->nb[1], 1 * sizeof(float) * n_embd);
+            struct ggml_tensor * Vcur = ggml_view_2d(ctx0, cur, n_embd, N, cur->nb[1], 2 * sizeof(float) * n_embd);
+
+            // store key and value to memory
+            {
+                struct ggml_tensor * k =
+                    ggml_view_1d(ctx0, model.memory_k, N * n_embd,
+                                 (ggml_element_size(model.memory_k) * n_embd) * (il * n_ctx + n_past));
+                struct ggml_tensor * v =
+                    ggml_view_1d(ctx0, model.memory_v, N * n_embd,
+                                 (ggml_element_size(model.memory_v) * n_embd) * (il * n_ctx + n_past));
+
+                ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Kcur, k));
+                ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Vcur, v));
+            }
+
+            // Q = Qcur.contiguous().view(n_embd/n_head, n_head, N).permute(0,
+            // 2, 1, 3) [64, N, 12]
+            struct ggml_tensor * Q = ggml_permute(
+                ctx0, ggml_cpy(ctx0, Qcur, ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_embd / n_head, n_head, N)), 0, 2,
+                1, 3);
+
+            // K = Kmem.view(n_embd/n_head, n_head, n_past + N).permute(0, 2, 1,
+            // 3) [64, n_past + N, 12]
+            struct ggml_tensor * K =
+                ggml_permute(ctx0,
+                             ggml_reshape_3d(ctx0,
+                                             ggml_view_1d(ctx0, model.memory_k, (n_past + N) * n_embd,
+                                                          il * n_ctx * ggml_element_size(model.memory_k) * n_embd),
+                                             n_embd / n_head, n_head, n_past + N),
+                             0, 2, 1, 3);
+            // K * Q
+            struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);
+
+            // KQ_scaled = KQ / sqrt(n_embd/n_head)
+            struct ggml_tensor * KQ_scaled =
+                ggml_scale(ctx0, KQ, ggml_new_f32(ctx0, 1.0f / sqrt(float(n_embd) / n_head)));
+
+            struct ggml_tensor * KQ_scaled_alibi =
+                ggml_alibi(ctx0, KQ_scaled, n_past, n_head, model.hparams.alibi_bias_max);
+
+            // KQ_masked = mask_past(KQ_scaled)
+            struct ggml_tensor * KQ_masked = ggml_diag_mask_inf(ctx0, KQ_scaled_alibi, n_past);
+
+            // KQ = soft_max(KQ_masked)
+            struct ggml_tensor * KQ_soft_max = ggml_soft_max(ctx0, KQ_masked);
+
+            // V_trans = Vmem.view(n_embd/n_head, n_head, n_past + N).permute(1,
+            // 2, 0, 3).contiguous() [n_past + N, 64, 12]
+            struct ggml_tensor * V_trans = ggml_cpy(
+                ctx0,
+                ggml_permute(ctx0,
+                             ggml_reshape_3d(ctx0,
+                                             ggml_view_1d(ctx0, model.memory_v, (n_past + N) * n_embd,
+                                                          il * n_ctx * ggml_element_size(model.memory_v) * n_embd),
+                                             n_embd / n_head, n_head, n_past + N),
+                             1, 2, 0, 3),
+                ggml_new_tensor_3d(ctx0, model.memory_v->type, n_past + N, n_embd / n_head, n_head));
+
+            // KQV = transpose(V) * KQ_soft_max
+            struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V_trans, KQ_soft_max);
+
+            // KQV_merged = KQV.permute(0, 2, 1, 3)
+            struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
+
+            // cur = KQV_merged.contiguous().view(n_embd, N)
+            cur = ggml_cpy(ctx0, KQV_merged, ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, N));
+
+            // projection
+            { cur = ggml_mul_mat(ctx0, model.layers[il].c_attn_out_proj_weight, cur); }
+        }
+
+        inpL = ggml_add(ctx0, inpL, cur);
+
+        ggml_set_scratch(ctx0, { 0, scr1_size, scr1, });
+
+        // m = self.ln_2(x)
+        {
+            cur = ggml_norm(ctx0, inpL, eps);
+
+            cur = ggml_mul(ctx0, ggml_repeat(ctx0, model.layers[il].norm_2_weight, cur), cur);
+        }
+
+        // n = self.mlp(m)
+        {
+
+            cur = ggml_mul_mat(ctx0, model.layers[il].ffn_up_proj, cur);
+
+            // GELU activation
+            cur = ggml_gelu(ctx0, cur);
+
+            // projection
+            // cur = proj_w*cur + proj_b
+            cur = ggml_mul_mat(ctx0, model.layers[il].ffn_down_proj, cur);
+        }
+
+        // x = x + n
+        inpL = ggml_add(ctx0, inpL, cur);
+    }
+
+    ggml_set_scratch(ctx0, { 0, scr0_size, scr0, });
+
+    // norm
+    {
+        inpL = ggml_norm(ctx0, inpL, eps);
+        // inpL = ln_f_g*inpL
+        inpL = ggml_mul(ctx0, ggml_repeat(ctx0, model.norm_f_weight, inpL), inpL);
+    }
+
+    ggml_set_scratch(ctx0, { 0, 0, nullptr, });
+
+    // output embedding weight tied to input embedding
+    inpL = ggml_mul_mat(ctx0, model.wte_weight, inpL);
+
+    // logits -> probs
+    // inpL = ggml_soft_max(ctx0, inpL);
+
+    // run the computation
+    ggml_build_forward_expand(&gf, inpL);
+    ggml_graph_compute_with_ctx(ctx0, &gf, n_threads);
+
+    // std::cout << "Qcur" << std::endl;
+    // print_tensor(Qcur);
+
+    // if (n_past%100 == 0) {
+    // ggml_graph_print(&gf);
+    // ggml_graph_dump_dot(&gf, NULL, "mpt-model.dot");
+    // }
+
+    if (logits_all) {
+        // return result for all tokens
+        embd_w.resize(n_vocab *N);
+        memcpy(embd_w.data(), (float *)ggml_get_data(inpL) , sizeof(float) * n_vocab * N);
+    } else {
+        // return result for just the last token
+        embd_w.resize(n_vocab);
+        memcpy(embd_w.data(), (float *)ggml_get_data(inpL) + (n_vocab * (N - 1)), sizeof(float) * n_vocab);
+    }
+
+    if (mem_per_token == 0) {
+        mem_per_token = ggml_used_mem(ctx0) / N;
+    }
+    // printf("used_mem = %zu\n", ggml_used_mem(ctx0));
+
+    ggml_free(ctx0);
+
+    return true;
+}
+
+std::vector<float> softmax(const std::vector<float> & logits) {
+    std::vector<float> probs(logits.size());
+    float max_logit = logits[0];
+    for (float v : logits) max_logit = std::max(max_logit, v);
+    double sum_exp = 0.0;
+    for (size_t i = 0; i < logits.size(); i++) {
+        // Subtract the maximum logit value from the current logit value for numerical stability
+        const float logit = logits[i] - max_logit;
+        const float exp_logit = expf(logit);
+        sum_exp += exp_logit;
+        probs[i] = exp_logit;
+    }
+    for (size_t i = 0; i < probs.size(); i++) probs[i] /= sum_exp;
+    return probs;
+}
+
+int perplexity(const mpt_params & params) {
+    ggml_time_init();
+
+    const int64_t t_main_start_us = ggml_time_us();
+
+    printf("%s: n_threads = %d\n", __func__, params.n_threads);
+    printf("%s: n_batch   = %d\n", __func__, params.n_batch);
+    printf("%s: n_ctx     = %d\n", __func__, params.n_ctx);
+    printf("\n");
+
+    int64_t t_load_us = 0;
+
+    gpt_vocab vocab;
+    mpt_model model;
+
+    model.hparams.n_ctx = params.n_ctx;
+
+    // load the model
+    {
+        const int64_t t_start_us = ggml_time_us();
+
+        if (!mpt_model_load(params.model, model, vocab)) {
+            fprintf(stderr, "%s: failed to load model from '%s'\n", __func__, params.model.c_str());
+            return 1;
+        }
+
+        t_load_us = ggml_time_us() - t_start_us;
+    }
+
+    int64_t t_predict_us = 0;
+
+    std::vector<float> logits;
+
+    // tokenize the prompt
+    std::vector<int> embd_inp = ::gpt_tokenize(vocab, params.prompt);
+
+    printf("%s: number of tokens in prompt = %zu\n", __func__, embd_inp.size());
+
+    // determine the required inference memory per token:
+    size_t mem_per_token = 0;
+    mpt_eval(model, params.n_threads, 0, {0, 1, 2, 3}, logits, false, mem_per_token);
+
+    int count   = 0;
+
+    const int n_chunk = embd_inp.size() / params.n_ctx;
+
+    const int n_vocab = model.hparams.n_vocab;
+    const int n_batch = params.n_batch;
+
+    double nll = 0.0;
+    fprintf(stderr, "%s: calculating perplexity over %d chunks, batch_size=%d\n", __func__, n_chunk, n_batch);
+
+    for (int i = 0; i < n_chunk; ++i) {
+
+        const int start =     i * params.n_ctx;
+        const int end   = start + params.n_ctx;
+
+        const int num_batches = (params.n_ctx + n_batch - 1) / n_batch;
+
+        std::vector<float> logits;
+
+        const auto t_start = std::chrono::high_resolution_clock::now();
+
+        for (int j = 0; j < num_batches; ++j) {
+
+            const int batch_start = start + j * n_batch;
+            const int batch_size  = std::min(end - batch_start, n_batch);
+
+            std::vector<gpt_vocab::id> embd;
+
+            for(int p=0;p<batch_size;p++) {
+                embd.push_back( embd_inp[batch_start+p]  );
+            }
+
+            std::vector<float> batch_logits;// = llama_get_logits(ctx);
+
+            const int64_t t_start_us = ggml_time_us();
+
+            if (!mpt_eval(model, params.n_threads, j * batch_size, embd, batch_logits, true, mem_per_token)) {
+                printf("%s: failed to evaluate model\n", __func__);
+                return 1;
+            }
+
+            t_predict_us += ggml_time_us() - t_start_us;
+
+            logits.insert(logits.end(), batch_logits.data(), batch_logits.data() + batch_size * n_vocab);
+
+        }
+
+        const auto t_end = std::chrono::high_resolution_clock::now();
+
+        if (i == 0) {
+            const float t_total = std::chrono::duration<float>(t_end - t_start).count();
+            fprintf(stderr, "%s: %.2f seconds per pass - ETA ", __func__, t_total);
+            int total_seconds = (int)(t_total * n_chunk);
+            if (total_seconds >= 60*60) {
+                fprintf(stderr, "%d hours ", total_seconds / (60*60));
+                total_seconds = total_seconds % (60*60);
+            }
+            fprintf(stderr, "%d minutes\n", total_seconds / 60);
+
+            printf("\nChunk\tPPL cumulative\tPPL chunk\n");
+        }
+
+        // We get the logits for all the tokens in the context window (params.n_ctx)
+        // from llama_eval above.  Now, based on https://huggingface.co/docs/transformers/perplexity,
+        // calculate the perplexity over the last half of the window (so the model always has
+        // some context to predict the token).
+        //
+        // We rely on the fact that attention in the forward pass only looks at previous
+        // tokens here, so the logits returned for each token are an accurate representation
+        // of what the model would have predicted at that point.
+        //
+        // Example, we have a context window of 512, we will compute perplexity for each of the
+        // last 256 tokens.  Then, we split the input up into context window size chunks to
+        // process the entire prompt.
+
+        double nllchunk = 0.0;
+        int countchunk = 0;
+
+        for (int j = std::min(512, params.n_ctx / 2); j < params.n_ctx - 1; ++j) {
+            // Calculate probability of next token, given the previous ones.
+            const std::vector<float> tok_logits(
+                logits.begin() + (j + 0) * n_vocab,
+                logits.begin() + (j + 1) * n_vocab);
+
+            const float prob = softmax(tok_logits)[embd_inp[ start+ j + 1]];
+
+            nllchunk += -std::log(prob);
+            ++countchunk;
+        }
+
+		nll += nllchunk;
+		count += countchunk;
+
+        // perplexity is e^(average negative log-likelihood)
+        printf("%d\t%.8lf\t%.8lf\n", i + 1, std::exp(nll / count), std::exp(nllchunk/countchunk) );
+        fflush(stdout);
+    }
+
+    // report timing
+    {
+        const int64_t t_main_end_us = ggml_time_us();
+
+        printf("\n\n");
+        printf("%s: mem per token = %8zu bytes\n", __func__, mem_per_token);
+        printf("%s:     load time = %8.2f ms\n",   __func__, t_load_us / 1000.0f);
+        printf("%s:     eval time = %8.2f ms / %.2f ms per token\n", __func__, t_predict_us / 1000.0f, t_predict_us / 1000.0f / (n_chunk * params.n_ctx));
+        printf("%s:    total time = %8.2f ms\n",   __func__, (t_main_end_us - t_main_start_us) / 1000.0f);
+    }
+
+    ggml_free(model.ctx);
+
+    return 0;
+}
+
+int main(int argc, char ** argv) {
+    mpt_params params;
+
+    if (mpt_params_parse(argc, argv, params) == false) {
+        return 1;
+    }
+
+    if (params.perplexity) {
+        return perplexity(params);
+    }
+
+    ggml_time_init();
+
+    const int64_t t_main_start_us = ggml_time_us();
+
+    if (params.seed < 0) {
+        params.seed = time(NULL);
+    }
+
+    if (params.n_predict < 0) {
+        params.n_predict = 0;
+    }
+
+    printf("%s: seed      = %d\n",   __func__, params.seed);
+    printf("%s: n_threads = %d\n",   __func__, params.n_threads);
+    printf("%s: n_batch   = %d\n",   __func__, params.n_batch);
+    printf("%s: n_ctx     = %d\n",   __func__, params.n_ctx);
+    printf("%s: n_predict = %d\n\n", __func__, params.n_predict);
+
+    std::mt19937 rng(params.seed);
+    if (params.prompt.empty()) {
+        params.prompt = gpt_random_prompt(rng);
+    }
+
+    int64_t t_load_us = 0;
+
+    gpt_vocab vocab;
+    mpt_model model;
+
+    model.hparams.n_ctx = params.n_ctx;
+
+    // load the model
+    {
+        const int64_t t_start_us = ggml_time_us();
+
+        if (!mpt_model_load(params.model, model, vocab)) {
+            fprintf(stderr, "%s: failed to load model from '%s'\n", __func__, params.model.c_str());
+            return 1;
+        }
+
+        t_load_us = ggml_time_us() - t_start_us;
+
+        test_gpt_tokenizer(vocab, params.token_test);
+    }
+
+    if (params.top_k == 0) {
+        params.top_k = model.hparams.n_vocab;
+    }
+
+    if (params.repeat_last_n == -1) {
+        params.repeat_last_n = params.n_ctx;
+    }
+
+    printf("\n");
+    printf("%s: temp           = %.3f\n", __func__, params.temp);
+    printf("%s: top_k          = %d\n",   __func__, params.top_k);
+    printf("%s: top_p          = %.3f\n", __func__, params.top_p);
+    printf("%s: repeat_last_n  = %d\n",   __func__, params.repeat_last_n);
+    printf("%s: repeat_penalty = %.3f\n", __func__, params.repeat_penalty);
+
+    int64_t t_sample_us = 0;
+    int64_t t_predict_us = 0;
+
+    std::vector<int32_t> last_n_tokens(params.n_ctx);
+    std::fill(last_n_tokens.begin(), last_n_tokens.end(), 0);
+
+    // tokenize the prompt
+    std::vector<int> embd_inp = ::gpt_tokenize(vocab, params.prompt);
+
+    printf("\n");
+    printf("%s: number of tokens in prompt = %zu\n", __func__, embd_inp.size());
+
+    for (size_t i = 0; i < embd_inp.size(); i++) {
+        printf("%s: token[%zu] = %6d\n", __func__, i, embd_inp[i]);
+    }
+    printf("\n");
+
+    std::vector<gpt_vocab::id> embd;
+    std::vector<float> logits;
+
+    // determine the required inference memory per token:
+    size_t mem_per_token = 0;
+    mpt_eval(model, params.n_threads, 0, {0, 1, 2, 3}, logits, false, mem_per_token);
+
+    int n_past     = 0;
+    int n_consumed = 0;
+    int n_sampled  = 0;
+
+    while (n_sampled < params.n_predict) {
+        // predict
+        if (embd.size() > 0) {
+            const int64_t t_start_us = ggml_time_us();
+
+            if (!mpt_eval(model, params.n_threads, n_past, embd, logits, false, mem_per_token)) {
+                printf("%s: failed to predict\n", __func__);
+                return 1;
+            }
+
+            t_predict_us += ggml_time_us() - t_start_us;
+
+            n_past += embd.size();
+            embd.clear();
+        }
+
+        if ((int)embd_inp.size() <= n_consumed) {
+            // sample next token
+
+            const int top_k = params.top_k;
+            const float top_p = params.top_p;
+            const float temp = params.temp;
+            const int repeat_last_n = params.repeat_last_n;
+            const float repeat_penalty = params.repeat_penalty;
+
+            gpt_vocab::id id = 0;
+
+            {
+                const int64_t t_start_sample_us = ggml_time_us();
+
+                id = gpt_sample_top_k_top_p_repeat(vocab, logits.data() + (logits.size() - model.hparams.n_vocab), last_n_tokens.data(), last_n_tokens.size(), top_k, top_p, temp, repeat_last_n, repeat_penalty, rng);
+
+                last_n_tokens.erase(last_n_tokens.begin());
+                last_n_tokens.push_back(id);
+
+                t_sample_us += ggml_time_us() - t_start_sample_us;
+            }
+
+            // add it to the context
+            embd.push_back(id);
+            ++n_sampled;
+
+        } else {
+            // if here, it means we are still processing the input prompt
+            while ((int) embd_inp.size() > n_consumed) {
+                embd.push_back(embd_inp[n_consumed]);
+
+                last_n_tokens.erase(last_n_tokens.begin());
+                last_n_tokens.push_back(embd_inp[n_consumed]);
+
+                ++n_consumed;
+                if ((int) embd.size() >= params.n_batch) {
+                    break;
+                }
+            }
+        }
+
+        // display text
+        for (auto id : embd) {
+           printf("%s", vocab.id_to_token[id].c_str());
+        }
+        fflush(stdout);
+
+        // end of text token
+        if (embd.back() == 0) {
+            break;
+        }
+    }
+
+    // report timing
+    {
+        const int64_t t_main_end_us = ggml_time_us();
+
+        printf("\n\n\n");
+        printf("%s: sampled tokens = %8d\n", __func__, n_sampled);
+        printf("%s:  mem per token = %8zu bytes\n", __func__, mem_per_token);
+        printf("%s:      load time = %8.2f ms\n", __func__, t_load_us / 1000.0f);
+        printf("%s:    sample time = %8.2f ms / %.2f ms per token\n", __func__, t_sample_us / 1000.0f, t_sample_us / 1000.0f / n_sampled);
+        printf("%s:      eval time = %8.2f ms / %.2f ms per token\n", __func__, t_predict_us / 1000.0f, t_predict_us / 1000.0f / n_past);
+        printf("%s:     total time = %8.2f ms\n", __func__, (t_main_end_us - t_main_start_us) / 1000.0f);
+    }
+
+    ggml_free(model.ctx);
+
+    return 0;
+}

ggml/examples/mpt/quantize.cpp

@@ -0,0 +1,186 @@
+#include "ggml/ggml.h"
+
+#include "common-ggml.h"
+#include "common.h"
+
+#include <cassert>
+#include <cmath>
+#include <cstdio>
+#include <cstring>
+#include <fstream>
+#include <map>
+#include <regex>
+#include <string>
+#include <vector>
+
+struct mpt_hparams {
+    int32_t d_model      = 0;
+    int32_t max_seq_len  = 0;
+    int32_t n_heads      = 0;
+    int32_t n_layers     = 0;
+    int32_t n_vocab      = 0;
+    float alibi_bias_max = 0;
+    float clip_qkv       = 0;
+    int32_t ftype        = 0;
+};
+
+// quantize a model
+bool mpt_model_quantize(const std::string & fname_inp,
+                        const std::string & fname_out, ggml_ftype ftype) {
+
+    printf("%s: loading model from '%s'\n", __func__, fname_inp.c_str());
+
+    auto finp = std::ifstream(fname_inp, std::ios::binary);
+    if (!finp) {
+        fprintf(stderr, "%s: failed to open '%s' for reading\n", __func__,
+                fname_inp.c_str());
+        return false;
+    }
+
+    auto fout = std::ofstream(fname_out, std::ios::binary);
+    if (!fout) {
+        fprintf(stderr, "%s: failed to open '%s' for writing\n", __func__,
+                fname_out.c_str());
+        return false;
+    }
+
+    // verify magic
+    {
+        uint32_t magic;
+        finp.read((char *)&magic, sizeof(magic));
+        if (magic != GGML_FILE_MAGIC) {
+            fprintf(stderr, "%s: invalid model file '%s' (bad magic)\n",
+                    __func__, fname_inp.c_str());
+            return false;
+        }
+
+        fout.write((char *)&magic, sizeof(magic));
+    }
+
+    mpt_hparams hparams;
+
+    // load hparams
+    {
+        finp.read((char *) &hparams.d_model,        sizeof(hparams.d_model));
+        finp.read((char *) &hparams.max_seq_len,    sizeof(hparams.max_seq_len));
+        finp.read((char *) &hparams.n_heads,        sizeof(hparams.n_heads));
+        finp.read((char *) &hparams.n_layers,       sizeof(hparams.n_layers));
+        finp.read((char *) &hparams.n_vocab,        sizeof(hparams.n_vocab));
+        finp.read((char *) &hparams.alibi_bias_max, sizeof(hparams.alibi_bias_max));
+        finp.read((char *) &hparams.clip_qkv,       sizeof(hparams.clip_qkv));
+        finp.read((char *) &hparams.ftype,          sizeof(hparams.ftype));
+
+        const int32_t qntvr_src =    hparams.ftype / GGML_QNT_VERSION_FACTOR;
+        const int32_t ftype_dst = GGML_QNT_VERSION * GGML_QNT_VERSION_FACTOR + ftype;
+
+        printf("%s: d_model        = %d\n", __func__, hparams.d_model);
+        printf("%s: max_seq_len    = %d\n", __func__, hparams.max_seq_len);
+        printf("%s: n_heads        = %d\n", __func__, hparams.n_heads);
+        printf("%s: n_layers       = %d\n", __func__, hparams.n_layers);
+        printf("%s: n_vocab        = %d\n", __func__, hparams.n_vocab);
+        printf("%s: alibi_bias_max = %f\n", __func__, hparams.alibi_bias_max);
+        printf("%s: clip_qkv       = %f\n", __func__, hparams.clip_qkv);
+        printf("%s: ftype (src) = %d\n", __func__, hparams.ftype);
+        printf("%s: qntvr (src) = %d\n", __func__, qntvr_src);
+        printf("%s: ftype (dst) = %d\n", __func__, ftype_dst);
+        printf("%s: qntvr (dst) = %d\n", __func__, GGML_QNT_VERSION);
+
+        fout.write((char *) &hparams.d_model,        sizeof(hparams.d_model));
+        fout.write((char *) &hparams.max_seq_len,    sizeof(hparams.max_seq_len));
+        fout.write((char *) &hparams.n_heads,        sizeof(hparams.n_heads));
+        fout.write((char *) &hparams.n_layers,       sizeof(hparams.n_layers));
+        fout.write((char *) &hparams.n_vocab,        sizeof(hparams.n_vocab));
+        fout.write((char *) &hparams.alibi_bias_max, sizeof(hparams.alibi_bias_max));
+        fout.write((char *) &hparams.clip_qkv,       sizeof(hparams.clip_qkv));
+        fout.write((char *) &ftype_dst,              sizeof(ftype_dst));
+    }
+
+    // load vocab
+    {
+        const int32_t n_vocab = hparams.n_vocab;
+
+        std::string word;
+        for (int i = 0; i < n_vocab; i++) {
+            uint32_t len;
+            finp.read((char *)&len, sizeof(len));
+            fout.write((char *)&len, sizeof(len));
+
+            word.resize(len);
+            finp.read((char *)word.data(), len);
+            fout.write((char *)word.data(), len);
+        }
+    }
+
+    printf("%s: quantizing tensors\n", __func__);
+
+    // regexes of tensor names to be quantized
+    const std::vector<std::string> to_quant = {
+        ".*weight",
+    };
+
+    if (!ggml_common_quantize_0(finp, fout, ftype, to_quant, {})) {
+        fprintf(stderr, "%s: failed to quantize model '%s'\n", __func__,
+                fname_inp.c_str());
+        return false;
+    }
+
+    finp.close();
+    fout.close();
+
+    return true;
+}
+
+// usage:
+//  ./mpt-quantize models/mpt/ggml-model.bin
+//  models/mpt/ggml-model-quant.bin type
+//
+int main(int argc, char ** argv) {
+    if (argc != 4) {
+        fprintf(stderr, "usage: %s model-f32.bin model-quant.bin type\n",
+                argv[0]);
+        ggml_print_ftypes(stderr);
+        return 1;
+    }
+
+    // needed to initialize f16 tables
+    {
+        struct ggml_init_params params = {0, NULL, false};
+        struct ggml_context * ctx = ggml_init(params);
+        ggml_free(ctx);
+    }
+
+    const std::string fname_inp = argv[1];
+    const std::string fname_out = argv[2];
+
+    const ggml_ftype ftype = ggml_parse_ftype(argv[3]);
+
+    const int64_t t_main_start_us = ggml_time_us();
+
+    int64_t t_quantize_us = 0;
+
+    // load the model
+    {
+        const int64_t t_start_us = ggml_time_us();
+
+        if (!mpt_model_quantize(fname_inp, fname_out, ggml_ftype(ftype))) {
+            fprintf(stderr, "%s: failed to quantize model from '%s'\n",
+                    __func__, fname_inp.c_str());
+            return 1;
+        }
+
+        t_quantize_us = ggml_time_us() - t_start_us;
+    }
+
+    // report timing
+    {
+        const int64_t t_main_end_us = ggml_time_us();
+
+        printf("\n");
+        printf("%s: quantize time = %8.2f ms\n", __func__,
+               t_quantize_us / 1000.0f);
+        printf("%s:    total time = %8.2f ms\n", __func__,
+               (t_main_end_us - t_main_start_us) / 1000.0f);
+    }
+
+    return 0;
+}

ggml/examples/prompts/dolly-v2.txt

@@ -0,0 +1,100 @@
+Hello World! => 12092,3645,2
+I can't believe it's already Friday!" => 42,476,626,2868,352,434,2168,6794,1476
+The URL for the website is https://www.example.com." => 510,10611,323,253,4422,310,5987,1358,2700,15,11667,15,681,449
+"She said, 'I love to travel.'" => 3,2993,753,13,686,42,2389,281,4288,18574
+'The temperature is 25.5°C.' => 8,510,3276,310,2030,15,22,3272,36,2464
+"Let's meet at 2:30 p.m. in the park." => 3,1466,434,2525,387,374,27,1229,268,15,78,15,275,253,5603,449
+The book costs $19.99 => 510,1984,4815,370,746,15,1525
+"John's favorite color is blue." => 3,8732,434,7583,3295,310,4797,449
+Th@nk y0u f0r y0ur h3lp! => 1044,33,30664,340,17,86,269,17,83,340,17,321,288,20,24343,2
+C@n I g3t a c0ffee, pl3@se? => 36,33,79,309,305,20,85,247,260,17,71,6851,13,499,20,33,339,32
+W0w! Th@t's @m@zing! => 56,17,88,2,596,33,85,434,1214,78,33,8537,2
+H0w 4re y0u t0d@y? => 41,17,88,577,250,340,17,86,246,17,69,33,90,32
+I l0ve t0 tr@vel @r0und the w0rld. => 42,298,17,306,246,17,492,33,652,1214,83,17,1504,253,259,17,83,392,15
+Wh@t's y0ur f@v0rite m0vie? => 3152,33,85,434,340,17,321,269,33,87,17,3852,278,17,25858,32
+The cat is sleeping on the mat. => 510,5798,310,14343,327,253,1111,15
+I need to buy some groceries for dinner. => 42,878,281,4489,690,45160,447,323,8955,15
+The sun is shining brightly in the sky. => 510,5101,310,28115,43925,275,253,8467,15
+She is reading a book in the park. => 2993,310,4361,247,1984,275,253,5603,15
+We went for a walk on the beach yesterday. => 1231,2427,323,247,2940,327,253,11600,11066,15
+He plays the guitar like a pro. => 1328,7120,253,12609,751,247,354,15
+They are going to the movies tonight. => 3726,403,1469,281,253,11321,11608,15
+The flowers are blooming in the garden. => 510,12405,403,30601,272,275,253,10329,15
+I enjoy listening to classical music. => 42,4264,11298,281,8946,3440,15
+We need to buy groceries for the week. => 1231,878,281,4489,45160,447,323,253,2129,15
+The dog is chasing its tail in circles. => 510,4370,310,31702,697,8105,275,14240,15
+She is wearing a beautiful red dress. => 2993,310,9398,247,5389,2502,7619,15
+He is a talented actor in Hollywood. => 1328,310,247,21220,12353,275,14759,15
+The children are playing in the playground. => 510,2151,403,4882,275,253,41008,15
+I'm going to visit my grandparents this weekend. => 42,1353,1469,281,4143,619,37186,436,8849,15
+The coffee tastes bitter without sugar. => 510,8574,27491,17123,1293,8618,15
+They are planning a surprise party for her. => 3726,403,7219,247,9326,3128,323,617,15
+She sings like an angel on stage. => 2993,44718,751,271,23087,327,3924,15
+We should take a vacation to relax. => 1231,943,1379,247,18125,281,7921,15
+He is studying medicine at the university. => 1328,310,12392,9921,387,253,9835,15
+The rain is pouring heavily outside. => 510,9313,310,31226,11306,3345,15
+I enjoy watching romantic movies. => 42,4264,7487,18109,11321,15
+They are celebrating their anniversary today. => 3726,403,28765,616,19054,3063,15
+She dances gracefully to the music. => 2993,47078,14426,2920,281,253,3440,15
+He is an excellent basketball player. => 1328,310,271,7126,14648,4760,15
+The baby is sleeping soundly in the crib. => 510,6858,310,14343,3590,314,275,253,260,725,15
+I need to finish my homework before dinner. => 42,878,281,8416,619,32110,1078,8955,15
+They are organizing a charity event next month. => 3726,403,26169,247,19489,2362,1735,1770,15
+She is cooking a delicious meal for us. => 2993,310,12398,247,17319,11484,323,441,15
+We should go hiking in the mountains. => 1231,943,564,33061,275,253,14700,15
+The car broke down on the way to work. => 510,1113,9377,1066,327,253,1039,281,789,15
+He loves playing video games in his free time. => 1328,14528,4882,3492,3958,275,521,1959,673,15
+The birds are chirping in the trees. => 510,11260,403,36494,14650,275,253,7139,15
+I want to learn how to play the piano. => 42,971,281,3037,849,281,1132,253,18542,15
+They are building a new shopping mall in the city. => 3726,403,3652,247,747,12701,28974,275,253,2846,15
+She is writing a novel in her spare time. => 2993,310,4028,247,4460,275,617,18345,673,15
+We are going to the zoo this Saturday. => 1231,403,1469,281,253,41089,436,7814,15
+The cake looks delicious with chocolate frosting. => 510,15221,4453,17319,342,14354,34724,272,15
+He is a talented painter who sells his artwork. => 1328,310,247,21220,27343,665,27924,521,28227,15
+The students are studying for their exams. => 510,3484,403,12392,323,616,34666,15
+I enjoy swimming in the ocean. => 42,4264,17120,275,253,12927,15
+They are renovating their house. => 3726,403,30074,839,616,2419,15
+She is practicing yoga to stay healthy. => 2993,310,25815,25551,281,3297,5875,15
+We should plant flowers in the garden. => 1231,943,4444,12405,275,253,10329,15
+The traffic is heavy during rush hour. => 510,7137,310,5536,1309,16949,4964,15
+He is a skilled chef who creates amazing dishes. => 1328,310,247,18024,26540,665,10513,8644,17114,15
+The baby is crawling on the floor. => 510,6858,310,44922,327,253,5254,15
+I need to buy a new pair of shoes. => 42,878,281,4489,247,747,4667,273,12682,15
+They are going on a road trip across the country. => 3726,403,1469,327,247,3971,7408,2439,253,2586,15
+She is playing the piano beautifully. => 2993,310,4882,253,18542,27839,15
+We are going to a concert tomorrow night. => 1231,403,1469,281,247,12699,10873,2360,15
+The cake tastes delicious with vanilla frosting. => 510,15221,27491,17319,342,26724,34724,272,15
+He is a dedicated teacher who inspires his students. => 1328,310,247,9940,9732,665,6381,2731,521,3484,15
+The students are participating in a science fair. => 510,3484,403,15299,275,247,5859,4344,15
+I enjoy hiking in the mountains. => 42,4264,33061,275,253,14700,15
+They are organizing a beach cleanup next weekend. => 3726,403,26169,247,11600,34709,1735,8849,15
+She is taking photographs of nature. => 2993,310,3192,15928,273,3753,15
+We should try a new restaurant in town. => 1231,943,1611,247,747,10301,275,3874,15
+The traffic is moving slowly on the highway. => 510,7137,310,4886,7808,327,253,17657,15
+He is a talented singer with a beautiful voice. => 1328,310,247,21220,16057,342,247,5389,4318,15
+The baby is laughing and giggling. => 510,6858,310,17053,285,41542,1981,15
+I need to do laundry and wash my clothes. => 42,878,281,513,29023,285,14841,619,10015,15
+They are planning a trip to Europe. => 3726,403,7219,247,7408,281,3060,15
+She is learning how to play the guitar. => 2993,310,4715,849,281,1132,253,12609,15
+We are going to a museum this Sunday. => 1231,403,1469,281,247,16064,436,6926,15
+The coffee smells amazing in the morning. => 510,8574,34247,8644,275,253,4131,15
+He is a hardworking farmer who grows crops. => 1328,310,247,1892,21107,24718,665,17202,19492,15
+The students are presenting their research projects. => 510,3484,403,15250,616,2561,6493,15
+I enjoy playing soccer with my friends. => 42,4264,4882,20391,342,619,3858,15
+They are volunteering at a local shelter. => 3726,403,10057,2158,387,247,1980,17824,15
+She is practicing martial arts for self-defense. => 2993,310,25815,29731,14635,323,1881,14,29337,15
+We should try a new recipe for dinner. => 1231,943,1611,247,747,13612,323,8955,15
+The traffic is congest => 510,7137,310,25801
+The sun is shining brightly today. => 510,5101,310,28115,43925,3063,15
+I enjoy reading books in my free time. => 42,4264,4361,5098,275,619,1959,673,15
+She plays the piano beautifully. => 2993,7120,253,18542,27839,15
+The cat chased the mouse around the room. => 510,5798,40754,253,6521,1475,253,2316,15
+I love eating pizza with extra cheese. => 42,2389,9123,22534,342,4465,12173,15
+He always wears a hat wherever he goes. => 1328,1900,31394,247,7856,20312,344,4566,15
+The flowers in the garden are blooming. => 510,12405,275,253,10329,403,30601,272,15
+She danced gracefully on the stage. => 2993,39860,14426,2920,327,253,3924,15
+The dog barked loudly in the park. => 510,4370,21939,264,31311,275,253,5603,15
+We went swimming in the ocean yesterday. => 1231,2427,17120,275,253,12927,11066,15
+He speaks fluent French and Spanish. => 1328,16544,2938,290,5112,285,9883,15
+The train arrived at the station on time. => 510,6194,7244,387,253,4660,327,673,15
+She cooked a delicious meal for her family. => 2993,18621,247,17319,11484,323,617,2021,15

ggml/examples/prompts/gpt-2-chinese.txt

@@ -0,0 +1 @@
+请问洗手间在哪里？ => 6435,7309,3819,2797,7313,1762,1525,7027,8043

ggml/examples/prompts/gpt-2.txt

@@ -0,0 +1,100 @@
+Hello World! => 15496,2159,0
+I can't believe it's already Friday!" => 40,460,470,1975,340,338,1541,3217,2474
+The URL for the website is https://www.example.com." => 464,10289,329,262,3052,318,3740,1378,2503,13,20688,13,785,526
+"She said, 'I love to travel.'" => 1,3347,531,11,705,40,1842,284,3067,11496
+'The temperature is 25.5°C.' => 6,464,5951,318,1679,13,20,7200,34,2637
+"Let's meet at 2:30 p.m. in the park." => 1,5756,338,1826,379,362,25,1270,279,13,76,13,287,262,3952,526
+The book costs $19.99 => 464,1492,3484,720,1129,13,2079
+"John's favorite color is blue." => 1,7554,338,4004,3124,318,4171,526
+Th@nk y0u f0r y0ur h3lp! => 817,31,77,74,331,15,84,277,15,81,331,15,333,289,18,34431,0
+C@n I g3t a c0ffee, pl3@se? => 34,31,77,314,308,18,83,257,269,15,5853,11,458,18,31,325,30
+W0w! Th@t's @m@zing! => 54,15,86,0,536,31,83,338,2488,76,31,9510,0
+H0w 4re y0u t0d@y? => 39,15,86,604,260,331,15,84,256,15,67,31,88,30
+I l0ve t0 tr@vel @r0und the w0rld. => 40,300,15,303,256,15,491,31,626,2488,81,15,917,262,266,15,81,335,13
+Wh@t's y0ur f@v0rite m0vie? => 1199,31,83,338,331,15,333,277,31,85,15,6525,285,15,85,494,30
+The cat is sleeping on the mat. => 464,3797,318,11029,319,262,2603,13
+I need to buy some groceries for dinner. => 40,761,284,2822,617,38464,329,8073,13
+The sun is shining brightly in the sky. => 464,4252,318,22751,35254,287,262,6766,13
+She is reading a book in the park. => 3347,318,3555,257,1492,287,262,3952,13
+We went for a walk on the beach yesterday. => 1135,1816,329,257,2513,319,262,10481,7415,13
+He plays the guitar like a pro. => 1544,5341,262,10047,588,257,386,13
+They are going to the movies tonight. => 2990,389,1016,284,262,6918,9975,13
+The flowers are blooming in the garden. => 464,12734,389,24924,3383,287,262,11376,13
+I enjoy listening to classical music. => 40,2883,8680,284,15993,2647,13
+We need to buy groceries for the week. => 1135,761,284,2822,38464,329,262,1285,13
+The dog is chasing its tail in circles. => 464,3290,318,20023,663,7894,287,13332,13
+She is wearing a beautiful red dress. => 3347,318,5762,257,4950,2266,6576,13
+He is a talented actor in Hollywood. => 1544,318,257,12356,8674,287,8502,13
+The children are playing in the playground. => 464,1751,389,2712,287,262,24817,13
+I'm going to visit my grandparents this weekend. => 40,1101,1016,284,3187,616,28571,428,5041,13
+The coffee tastes bitter without sugar. => 464,6891,18221,12922,1231,7543,13
+They are planning a surprise party for her. => 2990,389,5410,257,5975,2151,329,607,13
+She sings like an angel on stage. => 3347,33041,588,281,18304,319,3800,13
+We should take a vacation to relax. => 1135,815,1011,257,14600,284,8960,13
+He is studying medicine at the university. => 1544,318,11065,9007,379,262,6403,13
+The rain is pouring heavily outside. => 464,6290,318,23147,7272,2354,13
+I enjoy watching romantic movies. => 40,2883,4964,14348,6918,13
+They are celebrating their anniversary today. => 2990,389,17499,511,11162,1909,13
+She dances gracefully to the music. => 3347,38207,11542,2759,284,262,2647,13
+He is an excellent basketball player. => 1544,318,281,6275,9669,2137,13
+The baby is sleeping soundly in the crib. => 464,5156,318,11029,2128,306,287,262,48083,13
+I need to finish my homework before dinner. => 40,761,284,5461,616,26131,878,8073,13
+They are organizing a charity event next month. => 2990,389,16924,257,11016,1785,1306,1227,13
+She is cooking a delicious meal for us. => 3347,318,10801,257,12625,9799,329,514,13
+We should go hiking in the mountains. => 1135,815,467,24522,287,262,12269,13
+The car broke down on the way to work. => 464,1097,6265,866,319,262,835,284,670,13
+He loves playing video games in his free time. => 1544,10408,2712,2008,1830,287,465,1479,640,13
+The birds are chirping in the trees. => 464,10087,389,442,343,13886,287,262,7150,13
+I want to learn how to play the piano. => 40,765,284,2193,703,284,711,262,19132,13
+They are building a new shopping mall in the city. => 2990,389,2615,257,649,9735,17374,287,262,1748,13
+She is writing a novel in her spare time. => 3347,318,3597,257,5337,287,607,13952,640,13
+We are going to the zoo this Saturday. => 1135,389,1016,284,262,26626,428,3909,13
+The cake looks delicious with chocolate frosting. => 464,12187,3073,12625,351,11311,21682,278,13
+He is a talented painter who sells his artwork. => 1544,318,257,12356,34537,508,16015,465,16257,13
+The students are studying for their exams. => 464,2444,389,11065,329,511,26420,13
+I enjoy swimming in the ocean. => 40,2883,14899,287,262,9151,13
+They are renovating their house. => 2990,389,24317,803,511,2156,13
+She is practicing yoga to stay healthy. => 3347,318,18207,20351,284,2652,5448,13
+We should plant flowers in the garden. => 1135,815,4618,12734,287,262,11376,13
+The traffic is heavy during rush hour. => 464,4979,318,4334,1141,10484,1711,13
+He is a skilled chef who creates amazing dishes. => 1544,318,257,14297,21221,508,8075,4998,16759,13
+The baby is crawling on the floor. => 464,5156,318,34499,319,262,4314,13
+I need to buy a new pair of shoes. => 40,761,284,2822,257,649,5166,286,10012,13
+They are going on a road trip across the country. => 2990,389,1016,319,257,2975,5296,1973,262,1499,13
+She is playing the piano beautifully. => 3347,318,2712,262,19132,21104,13
+We are going to a concert tomorrow night. => 1135,389,1016,284,257,10010,9439,1755,13
+The cake tastes delicious with vanilla frosting. => 464,12187,18221,12625,351,16858,21682,278,13
+He is a dedicated teacher who inspires his students. => 1544,318,257,7256,4701,508,38934,465,2444,13
+The students are participating in a science fair. => 464,2444,389,11983,287,257,3783,3148,13
+I enjoy hiking in the mountains. => 40,2883,24522,287,262,12269,13
+They are organizing a beach cleanup next weekend. => 2990,389,16924,257,10481,27425,1306,5041,13
+She is taking photographs of nature. => 3347,318,2263,12566,286,3450,13
+We should try a new restaurant in town. => 1135,815,1949,257,649,7072,287,3240,13
+The traffic is moving slowly on the highway. => 464,4979,318,3867,6364,319,262,12763,13
+He is a talented singer with a beautiful voice. => 1544,318,257,12356,14015,351,257,4950,3809,13
+The baby is laughing and giggling. => 464,5156,318,14376,290,30442,1359,13
+I need to do laundry and wash my clothes. => 40,761,284,466,25724,290,13502,616,8242,13
+They are planning a trip to Europe. => 2990,389,5410,257,5296,284,2031,13
+She is learning how to play the guitar. => 3347,318,4673,703,284,711,262,10047,13
+We are going to a museum this Sunday. => 1135,389,1016,284,257,13257,428,3502,13
+The coffee smells amazing in the morning. => 464,6891,25760,4998,287,262,3329,13
+He is a hardworking farmer who grows crops. => 1544,318,257,1327,16090,18739,508,13676,14450,13
+The students are presenting their research projects. => 464,2444,389,17728,511,2267,4493,13
+I enjoy playing soccer with my friends. => 40,2883,2712,11783,351,616,2460,13
+They are volunteering at a local shelter. => 2990,389,41434,379,257,1957,11772,13
+She is practicing martial arts for self-defense. => 3347,318,18207,15618,10848,329,2116,12,19774,13
+We should try a new recipe for dinner. => 1135,815,1949,257,649,8364,329,8073,13
+The traffic is congest => 464,4979,318,22791
+The sun is shining brightly today. => 464,4252,318,22751,35254,1909,13
+I enjoy reading books in my free time. => 40,2883,3555,3835,287,616,1479,640,13
+She plays the piano beautifully. => 3347,5341,262,19132,21104,13
+The cat chased the mouse around the room. => 464,3797,26172,262,10211,1088,262,2119,13
+I love eating pizza with extra cheese. => 40,1842,6600,14256,351,3131,9891,13
+He always wears a hat wherever he goes. => 1544,1464,17326,257,6877,14530,339,2925,13
+The flowers in the garden are blooming. => 464,12734,287,262,11376,389,24924,3383,13
+She danced gracefully on the stage. => 3347,39480,11542,2759,319,262,3800,13
+The dog barked loudly in the park. => 464,3290,21405,276,23112,287,262,3952,13
+We went swimming in the ocean yesterday. => 1135,1816,14899,287,262,9151,7415,13
+He speaks fluent French and Spanish. => 1544,9209,43472,4141,290,7897,13
+The train arrived at the station on time. => 464,4512,5284,379,262,4429,319,640,13
+She cooked a delicious meal for her family. => 3347,15847,257,12625,9799,329,607,1641,13

ggml/examples/prompts/gpt-j.txt

@@ -0,0 +1,100 @@
+Hello World! => 15496,2159,0
+I can't believe it's already Friday!" => 40,460,470,1975,340,338,1541,3217,2474
+The URL for the website is https://www.example.com." => 464,10289,329,262,3052,318,3740,1378,2503,13,20688,13,785,526
+"She said, 'I love to travel.'" => 1,3347,531,11,705,40,1842,284,3067,11496
+'The temperature is 25.5°C.' => 6,464,5951,318,1679,13,20,7200,34,2637
+"Let's meet at 2:30 p.m. in the park." => 1,5756,338,1826,379,362,25,1270,279,13,76,13,287,262,3952,526
+The book costs $19.99 => 464,1492,3484,720,1129,13,2079
+"John's favorite color is blue." => 1,7554,338,4004,3124,318,4171,526
+Th@nk y0u f0r y0ur h3lp! => 817,31,77,74,331,15,84,277,15,81,331,15,333,289,18,34431,0
+C@n I g3t a c0ffee, pl3@se? => 34,31,77,314,308,18,83,257,269,15,5853,11,458,18,31,325,30
+W0w! Th@t's @m@zing! => 54,15,86,0,536,31,83,338,2488,76,31,9510,0
+H0w 4re y0u t0d@y? => 39,15,86,604,260,331,15,84,256,15,67,31,88,30
+I l0ve t0 tr@vel @r0und the w0rld. => 40,300,15,303,256,15,491,31,626,2488,81,15,917,262,266,15,81,335,13
+Wh@t's y0ur f@v0rite m0vie? => 1199,31,83,338,331,15,333,277,31,85,15,6525,285,15,85,494,30
+The cat is sleeping on the mat. => 464,3797,318,11029,319,262,2603,13
+I need to buy some groceries for dinner. => 40,761,284,2822,617,38464,329,8073,13
+The sun is shining brightly in the sky. => 464,4252,318,22751,35254,287,262,6766,13
+She is reading a book in the park. => 3347,318,3555,257,1492,287,262,3952,13
+We went for a walk on the beach yesterday. => 1135,1816,329,257,2513,319,262,10481,7415,13
+He plays the guitar like a pro. => 1544,5341,262,10047,588,257,386,13
+They are going to the movies tonight. => 2990,389,1016,284,262,6918,9975,13
+The flowers are blooming in the garden. => 464,12734,389,24924,3383,287,262,11376,13
+I enjoy listening to classical music. => 40,2883,8680,284,15993,2647,13
+We need to buy groceries for the week. => 1135,761,284,2822,38464,329,262,1285,13
+The dog is chasing its tail in circles. => 464,3290,318,20023,663,7894,287,13332,13
+She is wearing a beautiful red dress. => 3347,318,5762,257,4950,2266,6576,13
+He is a talented actor in Hollywood. => 1544,318,257,12356,8674,287,8502,13
+The children are playing in the playground. => 464,1751,389,2712,287,262,24817,13
+I'm going to visit my grandparents this weekend. => 40,1101,1016,284,3187,616,28571,428,5041,13
+The coffee tastes bitter without sugar. => 464,6891,18221,12922,1231,7543,13
+They are planning a surprise party for her. => 2990,389,5410,257,5975,2151,329,607,13
+She sings like an angel on stage. => 3347,33041,588,281,18304,319,3800,13
+We should take a vacation to relax. => 1135,815,1011,257,14600,284,8960,13
+He is studying medicine at the university. => 1544,318,11065,9007,379,262,6403,13
+The rain is pouring heavily outside. => 464,6290,318,23147,7272,2354,13
+I enjoy watching romantic movies. => 40,2883,4964,14348,6918,13
+They are celebrating their anniversary today. => 2990,389,17499,511,11162,1909,13
+She dances gracefully to the music. => 3347,38207,11542,2759,284,262,2647,13
+He is an excellent basketball player. => 1544,318,281,6275,9669,2137,13
+The baby is sleeping soundly in the crib. => 464,5156,318,11029,2128,306,287,262,48083,13
+I need to finish my homework before dinner. => 40,761,284,5461,616,26131,878,8073,13
+They are organizing a charity event next month. => 2990,389,16924,257,11016,1785,1306,1227,13
+She is cooking a delicious meal for us. => 3347,318,10801,257,12625,9799,329,514,13
+We should go hiking in the mountains. => 1135,815,467,24522,287,262,12269,13
+The car broke down on the way to work. => 464,1097,6265,866,319,262,835,284,670,13
+He loves playing video games in his free time. => 1544,10408,2712,2008,1830,287,465,1479,640,13
+The birds are chirping in the trees. => 464,10087,389,442,343,13886,287,262,7150,13
+I want to learn how to play the piano. => 40,765,284,2193,703,284,711,262,19132,13
+They are building a new shopping mall in the city. => 2990,389,2615,257,649,9735,17374,287,262,1748,13
+She is writing a novel in her spare time. => 3347,318,3597,257,5337,287,607,13952,640,13
+We are going to the zoo this Saturday. => 1135,389,1016,284,262,26626,428,3909,13
+The cake looks delicious with chocolate frosting. => 464,12187,3073,12625,351,11311,21682,278,13
+He is a talented painter who sells his artwork. => 1544,318,257,12356,34537,508,16015,465,16257,13
+The students are studying for their exams. => 464,2444,389,11065,329,511,26420,13
+I enjoy swimming in the ocean. => 40,2883,14899,287,262,9151,13
+They are renovating their house. => 2990,389,24317,803,511,2156,13
+She is practicing yoga to stay healthy. => 3347,318,18207,20351,284,2652,5448,13
+We should plant flowers in the garden. => 1135,815,4618,12734,287,262,11376,13
+The traffic is heavy during rush hour. => 464,4979,318,4334,1141,10484,1711,13
+He is a skilled chef who creates amazing dishes. => 1544,318,257,14297,21221,508,8075,4998,16759,13
+The baby is crawling on the floor. => 464,5156,318,34499,319,262,4314,13
+I need to buy a new pair of shoes. => 40,761,284,2822,257,649,5166,286,10012,13
+They are going on a road trip across the country. => 2990,389,1016,319,257,2975,5296,1973,262,1499,13
+She is playing the piano beautifully. => 3347,318,2712,262,19132,21104,13
+We are going to a concert tomorrow night. => 1135,389,1016,284,257,10010,9439,1755,13
+The cake tastes delicious with vanilla frosting. => 464,12187,18221,12625,351,16858,21682,278,13
+He is a dedicated teacher who inspires his students. => 1544,318,257,7256,4701,508,38934,465,2444,13
+The students are participating in a science fair. => 464,2444,389,11983,287,257,3783,3148,13
+I enjoy hiking in the mountains. => 40,2883,24522,287,262,12269,13
+They are organizing a beach cleanup next weekend. => 2990,389,16924,257,10481,27425,1306,5041,13
+She is taking photographs of nature. => 3347,318,2263,12566,286,3450,13
+We should try a new restaurant in town. => 1135,815,1949,257,649,7072,287,3240,13
+The traffic is moving slowly on the highway. => 464,4979,318,3867,6364,319,262,12763,13
+He is a talented singer with a beautiful voice. => 1544,318,257,12356,14015,351,257,4950,3809,13
+The baby is laughing and giggling. => 464,5156,318,14376,290,30442,1359,13
+I need to do laundry and wash my clothes. => 40,761,284,466,25724,290,13502,616,8242,13
+They are planning a trip to Europe. => 2990,389,5410,257,5296,284,2031,13
+She is learning how to play the guitar. => 3347,318,4673,703,284,711,262,10047,13
+We are going to a museum this Sunday. => 1135,389,1016,284,257,13257,428,3502,13
+The coffee smells amazing in the morning. => 464,6891,25760,4998,287,262,3329,13
+He is a hardworking farmer who grows crops. => 1544,318,257,1327,16090,18739,508,13676,14450,13
+The students are presenting their research projects. => 464,2444,389,17728,511,2267,4493,13
+I enjoy playing soccer with my friends. => 40,2883,2712,11783,351,616,2460,13
+They are volunteering at a local shelter. => 2990,389,41434,379,257,1957,11772,13
+She is practicing martial arts for self-defense. => 3347,318,18207,15618,10848,329,2116,12,19774,13
+We should try a new recipe for dinner. => 1135,815,1949,257,649,8364,329,8073,13
+The traffic is congest => 464,4979,318,22791
+The sun is shining brightly today. => 464,4252,318,22751,35254,1909,13
+I enjoy reading books in my free time. => 40,2883,3555,3835,287,616,1479,640,13
+She plays the piano beautifully. => 3347,5341,262,19132,21104,13
+The cat chased the mouse around the room. => 464,3797,26172,262,10211,1088,262,2119,13
+I love eating pizza with extra cheese. => 40,1842,6600,14256,351,3131,9891,13
+He always wears a hat wherever he goes. => 1544,1464,17326,257,6877,14530,339,2925,13
+The flowers in the garden are blooming. => 464,12734,287,262,11376,389,24924,3383,13
+She danced gracefully on the stage. => 3347,39480,11542,2759,319,262,3800,13
+The dog barked loudly in the park. => 464,3290,21405,276,23112,287,262,3952,13
+We went swimming in the ocean yesterday. => 1135,1816,14899,287,262,9151,7415,13
+He speaks fluent French and Spanish. => 1544,9209,43472,4141,290,7897,13
+The train arrived at the station on time. => 464,4512,5284,379,262,4429,319,640,13
+She cooked a delicious meal for her family. => 3347,15847,257,12625,9799,329,607,1641,13

ggml/examples/prompts/gpt-neox-japanese.txt

@@ -0,0 +1 @@
+明日の天気はどうですか。 => 263,7353,268,18461,271,1722,18405,265

ggml/examples/prompts/gpt-neox.txt

@@ -0,0 +1,100 @@
+Hello World! => 12092,3645,2
+I can't believe it's already Friday!" => 42,476,626,2868,352,434,2168,6794,1476
+The URL for the website is https://www.example.com." => 510,10611,323,253,4422,310,5987,1358,2700,15,11667,15,681,449
+"She said, 'I love to travel.'" => 3,2993,753,13,686,42,2389,281,4288,18574
+'The temperature is 25.5°C.' => 8,510,3276,310,2030,15,22,3272,36,2464
+"Let's meet at 2:30 p.m. in the park." => 3,1466,434,2525,387,374,27,1229,268,15,78,15,275,253,5603,449
+The book costs $19.99 => 510,1984,4815,370,746,15,1525
+"John's favorite color is blue." => 3,8732,434,7583,3295,310,4797,449
+Th@nk y0u f0r y0ur h3lp! => 1044,33,30664,340,17,86,269,17,83,340,17,321,288,20,24343,2
+C@n I g3t a c0ffee, pl3@se? => 36,33,79,309,305,20,85,247,260,17,71,6851,13,499,20,33,339,32
+W0w! Th@t's @m@zing! => 56,17,88,2,596,33,85,434,1214,78,33,8537,2
+H0w 4re y0u t0d@y? => 41,17,88,577,250,340,17,86,246,17,69,33,90,32
+I l0ve t0 tr@vel @r0und the w0rld. => 42,298,17,306,246,17,492,33,652,1214,83,17,1504,253,259,17,83,392,15
+Wh@t's y0ur f@v0rite m0vie? => 3152,33,85,434,340,17,321,269,33,87,17,3852,278,17,25858,32
+The cat is sleeping on the mat. => 510,5798,310,14343,327,253,1111,15
+I need to buy some groceries for dinner. => 42,878,281,4489,690,45160,447,323,8955,15
+The sun is shining brightly in the sky. => 510,5101,310,28115,43925,275,253,8467,15
+She is reading a book in the park. => 2993,310,4361,247,1984,275,253,5603,15
+We went for a walk on the beach yesterday. => 1231,2427,323,247,2940,327,253,11600,11066,15
+He plays the guitar like a pro. => 1328,7120,253,12609,751,247,354,15
+They are going to the movies tonight. => 3726,403,1469,281,253,11321,11608,15
+The flowers are blooming in the garden. => 510,12405,403,30601,272,275,253,10329,15
+I enjoy listening to classical music. => 42,4264,11298,281,8946,3440,15
+We need to buy groceries for the week. => 1231,878,281,4489,45160,447,323,253,2129,15
+The dog is chasing its tail in circles. => 510,4370,310,31702,697,8105,275,14240,15
+She is wearing a beautiful red dress. => 2993,310,9398,247,5389,2502,7619,15
+He is a talented actor in Hollywood. => 1328,310,247,21220,12353,275,14759,15
+The children are playing in the playground. => 510,2151,403,4882,275,253,41008,15
+I'm going to visit my grandparents this weekend. => 42,1353,1469,281,4143,619,37186,436,8849,15
+The coffee tastes bitter without sugar. => 510,8574,27491,17123,1293,8618,15
+They are planning a surprise party for her. => 3726,403,7219,247,9326,3128,323,617,15
+She sings like an angel on stage. => 2993,44718,751,271,23087,327,3924,15
+We should take a vacation to relax. => 1231,943,1379,247,18125,281,7921,15
+He is studying medicine at the university. => 1328,310,12392,9921,387,253,9835,15
+The rain is pouring heavily outside. => 510,9313,310,31226,11306,3345,15
+I enjoy watching romantic movies. => 42,4264,7487,18109,11321,15
+They are celebrating their anniversary today. => 3726,403,28765,616,19054,3063,15
+She dances gracefully to the music. => 2993,47078,14426,2920,281,253,3440,15
+He is an excellent basketball player. => 1328,310,271,7126,14648,4760,15
+The baby is sleeping soundly in the crib. => 510,6858,310,14343,3590,314,275,253,260,725,15
+I need to finish my homework before dinner. => 42,878,281,8416,619,32110,1078,8955,15
+They are organizing a charity event next month. => 3726,403,26169,247,19489,2362,1735,1770,15
+She is cooking a delicious meal for us. => 2993,310,12398,247,17319,11484,323,441,15
+We should go hiking in the mountains. => 1231,943,564,33061,275,253,14700,15
+The car broke down on the way to work. => 510,1113,9377,1066,327,253,1039,281,789,15
+He loves playing video games in his free time. => 1328,14528,4882,3492,3958,275,521,1959,673,15
+The birds are chirping in the trees. => 510,11260,403,36494,14650,275,253,7139,15
+I want to learn how to play the piano. => 42,971,281,3037,849,281,1132,253,18542,15
+They are building a new shopping mall in the city. => 3726,403,3652,247,747,12701,28974,275,253,2846,15
+She is writing a novel in her spare time. => 2993,310,4028,247,4460,275,617,18345,673,15
+We are going to the zoo this Saturday. => 1231,403,1469,281,253,41089,436,7814,15
+The cake looks delicious with chocolate frosting. => 510,15221,4453,17319,342,14354,34724,272,15
+He is a talented painter who sells his artwork. => 1328,310,247,21220,27343,665,27924,521,28227,15
+The students are studying for their exams. => 510,3484,403,12392,323,616,34666,15
+I enjoy swimming in the ocean. => 42,4264,17120,275,253,12927,15
+They are renovating their house. => 3726,403,30074,839,616,2419,15
+She is practicing yoga to stay healthy. => 2993,310,25815,25551,281,3297,5875,15
+We should plant flowers in the garden. => 1231,943,4444,12405,275,253,10329,15
+The traffic is heavy during rush hour. => 510,7137,310,5536,1309,16949,4964,15
+He is a skilled chef who creates amazing dishes. => 1328,310,247,18024,26540,665,10513,8644,17114,15
+The baby is crawling on the floor. => 510,6858,310,44922,327,253,5254,15
+I need to buy a new pair of shoes. => 42,878,281,4489,247,747,4667,273,12682,15
+They are going on a road trip across the country. => 3726,403,1469,327,247,3971,7408,2439,253,2586,15
+She is playing the piano beautifully. => 2993,310,4882,253,18542,27839,15
+We are going to a concert tomorrow night. => 1231,403,1469,281,247,12699,10873,2360,15
+The cake tastes delicious with vanilla frosting. => 510,15221,27491,17319,342,26724,34724,272,15
+He is a dedicated teacher who inspires his students. => 1328,310,247,9940,9732,665,6381,2731,521,3484,15
+The students are participating in a science fair. => 510,3484,403,15299,275,247,5859,4344,15
+I enjoy hiking in the mountains. => 42,4264,33061,275,253,14700,15
+They are organizing a beach cleanup next weekend. => 3726,403,26169,247,11600,34709,1735,8849,15
+She is taking photographs of nature. => 2993,310,3192,15928,273,3753,15
+We should try a new restaurant in town. => 1231,943,1611,247,747,10301,275,3874,15
+The traffic is moving slowly on the highway. => 510,7137,310,4886,7808,327,253,17657,15
+He is a talented singer with a beautiful voice. => 1328,310,247,21220,16057,342,247,5389,4318,15
+The baby is laughing and giggling. => 510,6858,310,17053,285,41542,1981,15
+I need to do laundry and wash my clothes. => 42,878,281,513,29023,285,14841,619,10015,15
+They are planning a trip to Europe. => 3726,403,7219,247,7408,281,3060,15
+She is learning how to play the guitar. => 2993,310,4715,849,281,1132,253,12609,15
+We are going to a museum this Sunday. => 1231,403,1469,281,247,16064,436,6926,15
+The coffee smells amazing in the morning. => 510,8574,34247,8644,275,253,4131,15
+He is a hardworking farmer who grows crops. => 1328,310,247,1892,21107,24718,665,17202,19492,15
+The students are presenting their research projects. => 510,3484,403,15250,616,2561,6493,15
+I enjoy playing soccer with my friends. => 42,4264,4882,20391,342,619,3858,15
+They are volunteering at a local shelter. => 3726,403,10057,2158,387,247,1980,17824,15
+She is practicing martial arts for self-defense. => 2993,310,25815,29731,14635,323,1881,14,29337,15
+We should try a new recipe for dinner. => 1231,943,1611,247,747,13612,323,8955,15
+The traffic is congest => 510,7137,310,25801
+The sun is shining brightly today. => 510,5101,310,28115,43925,3063,15
+I enjoy reading books in my free time. => 42,4264,4361,5098,275,619,1959,673,15
+She plays the piano beautifully. => 2993,7120,253,18542,27839,15
+The cat chased the mouse around the room. => 510,5798,40754,253,6521,1475,253,2316,15
+I love eating pizza with extra cheese. => 42,2389,9123,22534,342,4465,12173,15
+He always wears a hat wherever he goes. => 1328,1900,31394,247,7856,20312,344,4566,15
+The flowers in the garden are blooming. => 510,12405,275,253,10329,403,30601,272,15
+She danced gracefully on the stage. => 2993,39860,14426,2920,327,253,3924,15
+The dog barked loudly in the park. => 510,4370,21939,264,31311,275,253,5603,15
+We went swimming in the ocean yesterday. => 1231,2427,17120,275,253,12927,11066,15
+He speaks fluent French and Spanish. => 1328,16544,2938,290,5112,285,9883,15
+The train arrived at the station on time. => 510,6194,7244,387,253,4660,327,673,15
+She cooked a delicious meal for her family. => 2993,18621,247,17319,11484,323,617,2021,15

ggml/examples/prompts/polyglot-ko.txt

@@ -0,0 +1,3 @@
+이것은 테스트 이다. => 12271,296,6474,28037,17
+걱정할 필요 없다. => 18311,482,1062,550,267,17
+버그는 언젠가 고쳐진다. => 6904,272,8575,10381,1765,17

ggml/examples/prompts/replit.txt

@@ -0,0 +1,100 @@
+Hello World! => 6466,147,2317,350
+I can't believe it's already Friday!" => 286,512,172,185,13392,393,172,155,3239,147,29249,8537
+The URL for the website is https://www.example.com." => 505,5635,250,170,11745,235,147,303,262,552,148,811,148,241,148,161
+"She said, 'I love to travel.'" => 161,10386,4089,150,206,286,8440,194,147,12363,148,172,161
+'The temperature is 25.5°C.' => 172,505,147,9502,235,147,20022,8516,228,148,172
+"Let's meet at 2:30 p.m. in the park." => 161,8997,172,155,17120,536,147,162,5245,147,207,148,204,148,219,170,147,17664,148,161
+The book costs $19.99 => 505,147,2277,17494,236,166,11824
+"John's favorite color is blue." => 161,7475,172,155,147,11105,147,349,235,17046,148,161
+Th@nk y0u f0r y0ur h3lp! => 6309,240,9019,147,237,159,247,147,202,159,223,147,237,159,2458,147,226,171,3899,350
+C@n I g3t a c0ffee, pl3@se? => 228,240,211,398,147,267,171,185,216,147,196,159,13360,163,150,147,1287,171,240,155,163,272
+W0w! Th@t's @m@zing! => 450,159,274,350,147,6309,240,185,172,155,268,204,240,301,248,350
+H0w 4re y0u t0d@y? => 304,159,274,320,440,147,237,159,247,147,185,159,182,240,237,272
+I l0ve t0 tr@vel @r0und the w0rld. => 286,997,159,1290,147,185,159,147,490,240,3893,268,223,159,3981,170,147,274,159,223,2833,148
+Wh@t's y0ur f@v0rite m0vie? => 450,226,240,185,172,155,147,237,159,2458,147,202,240,252,159,5961,163,147,204,159,24373,272
+The cat is sleeping on the mat. => 505,147,1604,235,147,3987,248,347,170,147,1297,148
+I need to buy some groceries for dinner. => 286,1645,194,147,8068,1499,147,10022,1037,10023,250,147,182,2749,148
+The sun is shining brightly in the sky. => 505,147,5852,235,147,7304,2967,147,215,649,391,219,170,147,7310,148
+She is reading a book in the park. => 10386,235,9838,216,147,2277,219,170,147,17664,148
+We went for a walk on the beach yesterday. => 3250,10825,250,216,147,8156,347,170,294,5371,147,28830,148
+He plays the guitar like a pro. => 5301,7084,155,170,147,4604,2214,1425,216,3474,148
+They are going to the movies tonight. => 18815,429,6552,194,170,147,15877,194,7907,148
+The flowers are blooming in the garden. => 505,147,22953,155,429,147,10411,2799,248,219,170,147,22140,148
+I enjoy listening to classical music. => 286,23162,15876,248,194,239,4251,147,7395,148
+We need to buy groceries for the week. => 3250,1645,194,147,8068,147,10022,1037,10023,250,170,9238,148
+The dog is chasing its tail in circles. => 505,147,6540,235,147,196,916,248,1602,147,5129,219,147,4095,155,148
+She is wearing a beautiful red dress. => 10386,235,147,16427,248,216,147,23447,147,1160,147,14592,148
+He is a talented actor in Hollywood. => 5301,235,216,147,29750,246,147,5112,219,147,16924,391,10477,148
+The children are playing in the playground. => 505,7934,429,7084,248,219,170,7084,12055,148
+I'm going to visit my grandparents this weekend. => 286,172,204,6552,194,9939,1247,147,11806,12019,291,9238,314,148
+The coffee tastes bitter without sugar. => 505,147,21526,147,20931,155,5145,1430,1988,147,28759,148
+They are planning a surprise party for her. => 18815,429,147,23661,216,147,29240,147,7344,250,1869,148
+She sings like an angel on stage. => 10386,147,155,6502,1425,426,147,26028,347,12685,148
+We should take a vacation to relax. => 3250,936,4654,216,147,15388,946,194,1998,2744,148
+He is studying medicine at the university. => 5301,235,7959,248,147,20742,1668,536,170,147,8025,148
+The rain is pouring heavily outside. => 505,147,6885,235,5306,248,1189,5451,391,8096,148
+I enjoy watching romantic movies. => 286,23162,147,3355,248,147,26080,4140,147,15877,148
+They are celebrating their anniversary today. => 18815,429,147,30000,5841,1669,147,24734,5464,1770,13386,148
+She dances gracefully to the music. => 10386,147,182,1626,155,147,267,8771,8001,194,170,147,7395,148
+He is an excellent basketball player. => 5301,235,426,147,12300,675,185,147,26646,5132,6294,148
+The baby is sleeping soundly in the crib. => 505,147,23597,235,147,3987,248,12642,391,219,170,147,7696,215,148
+I need to finish my homework before dinner. => 286,1645,194,147,6717,1247,147,1071,2722,2643,147,182,2749,148
+They are organizing a charity event next month. => 18815,429,147,16442,248,216,1054,1511,1663,2399,12821,148
+She is cooking a delicious meal for us. => 10386,235,147,20453,248,216,3936,23455,147,26658,250,147,539,148
+We should go hiking in the mountains. => 3250,936,4242,147,2254,5357,219,170,147,204,18028,155,148
+The car broke down on the way to work. => 505,7553,147,510,10036,4288,347,170,3699,194,1916,148
+He loves playing video games in his free time. => 5301,8440,155,7084,248,8722,147,11281,219,1439,4002,801,148
+The birds are chirping in the trees. => 505,147,13043,155,429,147,3904,223,4639,219,170,5311,155,148
+I want to learn how to play the piano. => 286,1857,194,14167,2496,194,7084,170,147,207,23635,148
+They are building a new shopping mall in the city. => 18815,429,11038,216,277,147,22184,147,204,609,219,170,147,2416,148
+She is writing a novel in her spare time. => 10386,235,3242,216,147,25814,219,1869,6772,2382,801,148
+We are going to the zoo this Saturday. => 3250,429,6552,194,170,147,25101,291,147,31426,148
+The cake looks delicious with chocolate frosting. => 505,147,24422,16303,3936,23455,312,147,5619,533,2239,147,202,3973,3431,148
+He is a talented painter who sells his artwork. => 5301,235,216,147,29750,246,147,9226,279,2888,13004,155,1439,12234,2722,148
+The students are studying for their exams. => 505,15707,429,7959,248,250,1669,147,12398,155,148
+I enjoy swimming in the ocean. => 286,23162,147,4729,8528,248,219,170,147,26193,148
+They are renovating their house. => 18815,429,991,10724,3643,1669,13788,148
+She is practicing yoga to stay healthy. => 10386,235,147,18453,248,147,5063,1186,194,15344,147,28550,148
+We should plant flowers in the garden. => 3250,936,147,9212,147,22953,155,219,170,147,22140,148
+The traffic is heavy during rush hour. => 505,147,11097,235,147,22232,4340,147,22319,147,5686,148
+He is a skilled chef who creates amazing dishes. => 5301,235,216,147,8891,246,9784,202,2888,13720,147,28880,147,23852,383,148
+The baby is crawling on the floor. => 505,147,23597,235,147,22120,248,347,170,147,5895,148
+I need to buy a new pair of shoes. => 286,1645,194,147,8068,216,277,12632,210,147,155,21953,155,148
+They are going on a road trip across the country. => 18815,429,6552,347,216,147,6362,147,11395,9762,170,11305,148
+She is playing the piano beautifully. => 10386,235,7084,248,170,147,207,23635,147,23447,391,148
+We are going to a concert tomorrow night. => 3250,429,6552,194,216,1710,4391,29524,12716,148
+The cake tastes delicious with vanilla frosting. => 505,147,24422,147,20931,155,3936,23455,312,5535,7476,147,202,3973,3431,148
+He is a dedicated teacher who inspires his students. => 5301,235,216,326,8298,3460,147,9675,2888,147,28801,155,1439,15707,148
+The students are participating in a science fair. => 505,15707,429,147,30961,3643,219,216,147,10587,147,7636,148
+I enjoy hiking in the mountains. => 286,23162,147,2254,5357,219,170,147,204,18028,155,148
+They are organizing a beach cleanup next weekend. => 18815,429,147,16442,248,216,294,5371,147,10401,2399,9238,314,148
+She is taking photographs of nature. => 10386,235,147,12345,147,4709,1547,155,210,147,211,8603,148
+We should try a new restaurant in town. => 3250,936,147,746,216,277,147,11007,219,147,10200,148
+The traffic is moving slowly on the highway. => 505,147,11097,235,147,8601,147,9880,391,347,170,5976,3330,148
+He is a talented singer with a beautiful voice. => 5301,235,216,147,29750,246,147,155,248,279,312,216,147,23447,147,9316,148
+The baby is laughing and giggling. => 505,147,23597,235,147,23066,248,221,147,2341,3631,2869,148
+I need to do laundry and wash my clothes. => 286,1645,194,543,960,3981,2154,221,147,27589,1247,147,22141,383,148
+They are planning a trip to Europe. => 18815,429,147,23661,216,147,11395,194,13131,148
+She is learning how to play the guitar. => 10386,235,11754,2496,194,7084,170,147,4604,2214,148
+We are going to a museum this Sunday. => 3250,429,6552,194,216,147,204,433,1177,291,147,29111,148
+The coffee smells amazing in the morning. => 505,147,21526,31454,155,147,28880,219,170,20701,148
+He is a hardworking farmer who grows crops. => 5301,235,216,8524,14992,147,16679,279,2888,147,6044,155,147,8650,155,148
+The students are presenting their research projects. => 505,15707,429,5130,248,1669,13217,14235,148
+I enjoy playing soccer with my friends. => 286,23162,7084,248,147,9351,5318,312,1247,147,5347,155,148
+They are volunteering at a local shelter. => 18815,429,147,5238,7478,163,12798,536,216,2491,2905,1359,279,148
+She is practicing martial arts for self-defense. => 10386,235,147,18453,248,147,3261,185,4381,12234,155,250,623,153,29896,148
+We should try a new recipe for dinner. => 3250,936,147,746,216,277,147,9851,250,147,182,2749,148
+The traffic is congest => 505,147,11097,235,1710,14169
+The sun is shining brightly today. => 505,147,5852,235,147,7304,2967,147,215,649,391,13386,148
+I enjoy reading books in my free time. => 286,23162,9838,147,9670,219,1247,4002,801,148
+She plays the piano beautifully. => 10386,7084,155,170,147,207,23635,147,23447,391,148
+The cat chased the mouse around the room. => 505,147,1604,147,196,916,246,170,12551,6890,170,9654,148
+I love eating pizza with extra cheese. => 286,8440,147,163,3643,147,207,8403,312,8230,9784,383,163,148
+He always wears a hat wherever he goes. => 5301,5418,147,16427,155,216,147,4879,2171,2433,1189,16177,148
+The flowers in the garden are blooming. => 505,147,22953,155,219,170,147,22140,429,147,10411,2799,248,148
+She danced gracefully on the stage. => 10386,13378,12408,147,267,8771,8001,347,170,12685,148
+The dog barked loudly in the park. => 505,147,6540,147,973,293,246,147,30182,391,219,170,147,17664,148
+We went swimming in the ocean yesterday. => 3250,10825,147,4729,8528,248,219,170,147,26193,147,28830,148
+He speaks fluent French and Spanish. => 5301,147,13285,155,147,21677,147,254,17590,221,147,31519,148
+The train arrived at the station on time. => 505,147,872,147,20712,182,536,170,147,7184,347,801,148
+She cooked a delicious meal for her family. => 10386,147,20453,246,216,3936,23455,147,26658,250,1869,147,2002,148

ggml/examples/prompts/starcoder.txt

@@ -0,0 +1,100 @@
+Hello World! => 8279,10896,19
+I can't believe it's already Friday!" => 59,883,1330,13710,561,1182,3425,506,25674,11555
+The URL for the website is https://www.example.com." => 1318,3834,436,322,9575,438,1678,555,1499,32,2763,32,508,3107
+"She said, 'I love to travel.'" => 20,25387,9884,30,330,59,14290,372,25283,29329
+'The temperature is 25.5°C.' => 25,1318,13587,438,225,36,39,32,39,23767,53,4564
+"Let's meet at 2:30 p.m. in the park." => 20,9809,1182,18450,821,225,36,44,37,34,298,32,95,32,328,322,880,93,3107
+The book costs $19.99 => 1318,7618,25950,398,35,43,32,43,43
+"John's favorite color is blue." => 20,19693,1182,27448,1963,438,10087,3107
+Th@nk y0u f0r y0ur h3lp! => 1027,50,19877,533,34,103,296,34,100,533,34,305,420,37,1915,19
+C@n I g3t a c0ffee, pl3@se? => 53,50,96,439,485,37,102,312,281,34,21298,30,1278,37,50,277,49
+W0w! Th@t's @m@zing! => 73,34,105,19,947,50,102,1182,477,95,50,26768,19
+H0w 4re y0u t0d@y? => 58,34,105,225,38,268,533,34,103,273,34,86,50,107,49
+I l0ve t0 tr@vel @r0und the w0rld. => 59,456,34,587,273,34,554,50,1203,477,100,34,642,322,341,34,100,1381,32
+Wh@t's y0ur f@v0rite m0vie? => 2444,50,102,1182,533,34,305,296,50,104,34,1049,345,34,104,1075,49
+The cat is sleeping on the mat. => 1318,10501,438,9368,299,544,322,2491,32
+I need to buy some groceries for dinner. => 59,1849,372,16968,1629,20234,85,6958,436,343,3369,32
+The sun is shining brightly in the sky. => 1318,15323,438,787,19068,38231,631,328,322,26718,32
+She is reading a book in the park. => 25387,438,9175,312,7618,328,322,880,93,32
+We went for a walk on the beach yesterday. => 3122,14236,436,312,13503,544,322,526,867,39485,32
+He plays the guitar like a pro. => 1331,41271,322,3932,19931,2124,312,534,32
+They are going to the movies tonight. => 31805,884,6783,372,322,27889,26076,694,32
+The flowers are blooming in the garden. => 1318,7290,483,884,323,18466,299,328,322,485,22461,32
+I enjoy listening to classical music. => 59,31567,20498,372,443,1578,17522,32
+We need to buy groceries for the week. => 3122,1849,372,16968,20234,85,6958,436,322,8209,32
+The dog is chasing its tail in circles. => 1318,27435,438,663,9949,2819,13203,328,46428,32
+She is wearing a beautiful red dress. => 25387,438,996,6992,312,36493,3346,343,714,32
+He is a talented actor in Hollywood. => 1331,438,312,273,9556,318,16038,328,48228,631,21118,32
+The children are playing in the playground. => 1318,5713,884,19788,328,322,4654,1749,32
+I'm going to visit my grandparents this weekend. => 59,3464,6783,372,7725,1672,33162,19277,458,40618,32
+The coffee tastes bitter without sugar. => 1318,36917,273,633,307,3493,391,2876,309,18628,32
+They are planning a surprise party for her. => 31805,884,26116,312,6178,9251,15270,436,7791,32
+She sings like an angel on stage. => 25387,309,2052,2124,600,600,17691,544,10019,32
+We should take a vacation to relax. => 3122,1395,4818,312,29164,367,372,41972,32
+He is studying medicine at the university. => 1331,438,14866,299,32388,482,821,322,707,9190,32
+The rain is pouring heavily outside. => 1318,36987,438,9202,299,46003,2801,11127,32
+I enjoy watching romantic movies. => 59,31567,37652,26045,7268,27889,32
+They are celebrating their anniversary today. => 31805,884,48278,839,1741,3623,23921,5810,672,11610,32
+She dances gracefully to the music. => 25387,343,3151,31376,4938,372,322,17522,32
+He is an excellent basketball player. => 1331,438,600,39203,48400,11653,4362,32
+The baby is sleeping soundly in the crib. => 1318,323,17156,438,9368,299,9934,631,328,322,281,7972,32
+I need to finish my homework before dinner. => 59,1849,372,11361,1672,6765,1007,2670,343,3369,32
+They are organizing a charity event next month. => 31805,884,10558,6183,312,1351,543,1692,2354,6811,32
+She is cooking a delicious meal for us. => 25387,438,23682,299,312,409,406,2406,597,279,436,1770,32
+We should go hiking in the mountains. => 3122,1395,1983,420,1546,299,328,322,10874,1907,32
+The car broke down on the way to work. => 1318,6346,43289,2835,544,322,3352,372,1389,32
+He loves playing video games in his free time. => 1331,598,4954,19788,6027,19705,328,6697,3741,1133,32
+The birds are chirping in the trees. => 1318,8424,3210,884,663,476,7075,328,322,23453,32
+I want to learn how to play the piano. => 59,2637,372,7350,2624,372,4654,322,298,25757,32
+They are building a new shopping mall in the city. => 31805,884,9038,312,537,40692,345,464,328,322,11297,32
+She is writing a novel in her spare time. => 25387,438,4127,312,32913,328,7791,1869,586,1133,32
+We are going to the zoo this Saturday. => 3122,884,6783,372,322,1288,604,458,358,30288,32
+The cake looks delicious with chocolate frosting. => 1318,281,1062,7780,409,406,2406,623,10408,27589,296,20932,299,32
+He is a talented painter who sells his artwork. => 1331,438,312,273,9556,318,42300,6560,10800,101,6697,5549,1007,32
+The students are studying for their exams. => 1318,16512,884,14866,299,436,3623,538,1462,32
+I enjoy swimming in the ocean. => 59,31567,2535,449,6714,328,322,337,18857,32
+They are renovating their house. => 31805,884,316,15007,1741,3623,17075,32
+She is practicing yoga to stay healthy. => 25387,438,11808,11636,533,40067,372,20005,44538,32
+We should plant flowers in the garden. => 3122,1395,26795,7290,483,328,322,485,22461,32
+The traffic is heavy during rush hour. => 1318,16391,438,32389,5929,540,1372,12021,32
+He is a skilled chef who creates amazing dishes. => 1331,438,312,3001,12088,44051,6560,9585,36986,1214,4279,32
+The baby is crawling on the floor. => 1318,323,17156,438,281,1294,2920,544,322,17648,32
+I need to buy a new pair of shoes. => 59,1849,372,16968,312,537,6092,432,787,37764,32
+They are going on a road trip across the country. => 31805,884,6783,544,312,24122,19337,10160,322,10769,32
+She is playing the piano beautifully. => 25387,438,19788,322,298,25757,526,4846,325,514,107,32
+We are going to a concert tomorrow night. => 3122,884,6783,372,312,457,6989,31841,19212,32
+The cake tastes delicious with vanilla frosting. => 1318,281,1062,273,633,307,409,406,2406,623,44653,296,20932,299,32
+He is a dedicated teacher who inspires his students. => 1331,438,312,23112,30877,6560,26194,8017,6697,16512,32
+The students are participating in a science fair. => 1318,16512,884,24623,1741,328,312,27536,19375,32
+I enjoy hiking in the mountains. => 59,31567,420,1546,299,328,322,10874,1907,32
+They are organizing a beach cleanup next weekend. => 31805,884,10558,6183,312,526,867,13144,2354,40618,32
+She is taking photographs of nature. => 25387,438,15137,15110,23626,432,24406,32
+We should try a new restaurant in town. => 3122,1395,1596,312,537,43719,328,38212,32
+The traffic is moving slowly on the highway. => 1318,16391,438,14089,12899,631,544,322,3857,3073,32
+He is a talented singer with a beautiful voice. => 1331,438,312,273,9556,318,309,10118,623,312,36493,20309,32
+The baby is laughing and giggling. => 1318,323,17156,438,2317,2943,299,461,485,365,36088,32
+I need to do laundry and wash my clothes. => 59,1849,372,745,2317,642,994,461,341,917,1672,7375,46948,32
+They are planning a trip to Europe. => 31805,884,26116,312,19337,372,27268,32
+She is learning how to play the guitar. => 25387,438,9608,2624,372,4654,322,3932,19931,32
+We are going to a museum this Sunday. => 3122,884,6783,372,312,345,539,378,458,358,28036,32
+The coffee smells amazing in the morning. => 1318,36917,309,42153,101,36986,328,322,33768,32
+He is a hardworking farmer who grows crops. => 1331,438,312,6784,13578,9019,2302,6560,485,2138,25170,1069,32
+The students are presenting their research projects. => 1318,16512,884,5024,299,3623,13234,8528,32
+I enjoy playing soccer with my friends. => 59,31567,19788,22682,10035,623,1672,22523,32
+They are volunteering at a local shelter. => 31805,884,3920,45585,8637,821,312,2196,309,2542,391,32
+She is practicing martial arts for self-defense. => 25387,438,11808,11636,345,502,564,5549,101,436,630,31,43694,32
+We should try a new recipe for dinner. => 3122,1395,1596,312,537,15233,436,343,3369,32
+The traffic is congest => 1318,16391,438,457,2776
+The sun is shining brightly today. => 1318,15323,438,787,19068,38231,631,11610,32
+I enjoy reading books in my free time. => 59,31567,9175,21739,328,1672,3741,1133,32
+She plays the piano beautifully. => 25387,41271,322,298,25757,526,4846,325,514,107,32
+The cat chased the mouse around the room. => 1318,10501,663,16109,322,8459,6835,322,8355,32
+I love eating pizza with extra cheese. => 59,14290,484,1741,47630,623,6717,8277,30315,32
+He always wears a hat wherever he goes. => 1331,5182,996,4177,312,25793,2154,424,938,13107,32
+The flowers in the garden are blooming. => 1318,7290,483,328,322,485,22461,884,323,18466,299,32
+She danced gracefully on the stage. => 25387,343,6087,31376,4938,544,322,10019,32
+The dog barked loudly in the park. => 1318,27435,323,1087,318,598,836,631,328,322,880,93,32
+We went swimming in the ocean yesterday. => 3122,14236,2535,449,6714,328,322,337,18857,39485,32
+He speaks fluent French and Spanish. => 1331,24498,101,38055,43652,461,14911,1708,32
+The train arrived at the station on time. => 1318,5683,2099,32114,821,322,18662,544,1133,32
+She cooked a delicious meal for her family. => 25387,23682,318,312,409,406,2406,597,279,436,7791,13872,32

ggml/examples/prompts/test-cases.txt

@@ -0,0 +1,110 @@
+# test case format
+# <language>: <sentence>
+
+English: Hello World!
+English: I can't believe it's already Friday!"
+English: The URL for the website is https://www.example.com."
+English: "She said, 'I love to travel.'"
+English: 'The temperature is 25.5°C.'
+English: "Let's meet at 2:30 p.m. in the park."
+English: The book costs $19.99
+English: "John's favorite color is blue."
+English: Th@nk y0u f0r y0ur h3lp!
+English: C@n I g3t a c0ffee, pl3@se?
+English: W0w! Th@t's @m@zing!
+English: H0w 4re y0u t0d@y?
+English: I l0ve t0 tr@vel @r0und the w0rld.
+English: Wh@t's y0ur f@v0rite m0vie?
+English: The cat is sleeping on the mat.
+English: I need to buy some groceries for dinner.
+English: The sun is shining brightly in the sky.
+English: She is reading a book in the park.
+English: We went for a walk on the beach yesterday.
+English: He plays the guitar like a pro.
+English: They are going to the movies tonight.
+English: The flowers are blooming in the garden.
+English: I enjoy listening to classical music.
+English: We need to buy groceries for the week.
+English: The dog is chasing its tail in circles.
+English: She is wearing a beautiful red dress.
+English: He is a talented actor in Hollywood.
+English: The children are playing in the playground.
+English: I'm going to visit my grandparents this weekend.
+English: The coffee tastes bitter without sugar.
+English: They are planning a surprise party for her.
+English: She sings like an angel on stage.
+English: We should take a vacation to relax.
+English: He is studying medicine at the university.
+English: The rain is pouring heavily outside.
+English: I enjoy watching romantic movies.
+English: They are celebrating their anniversary today.
+English: She dances gracefully to the music.
+English: He is an excellent basketball player.
+English: The baby is sleeping soundly in the crib.
+English: I need to finish my homework before dinner.
+English: They are organizing a charity event next month.
+English: She is cooking a delicious meal for us.
+English: We should go hiking in the mountains.
+English: The car broke down on the way to work.
+English: He loves playing video games in his free time.
+English: The birds are chirping in the trees.
+English: I want to learn how to play the piano.
+English: They are building a new shopping mall in the city.
+English: She is writing a novel in her spare time.
+English: We are going to the zoo this Saturday.
+English: The cake looks delicious with chocolate frosting.
+English: He is a talented painter who sells his artwork.
+English: The students are studying for their exams.
+English: I enjoy swimming in the ocean.
+English: They are renovating their house.
+English: She is practicing yoga to stay healthy.
+English: We should plant flowers in the garden.
+English: The traffic is heavy during rush hour.
+English: He is a skilled chef who creates amazing dishes.
+English: The baby is crawling on the floor.
+English: I need to buy a new pair of shoes.
+English: They are going on a road trip across the country.
+English: She is playing the piano beautifully.
+English: We are going to a concert tomorrow night.
+English: The cake tastes delicious with vanilla frosting.
+English: He is a dedicated teacher who inspires his students.
+English: The students are participating in a science fair.
+English: I enjoy hiking in the mountains.
+English: They are organizing a beach cleanup next weekend.
+English: She is taking photographs of nature.
+English: We should try a new restaurant in town.
+English: The traffic is moving slowly on the highway.
+English: He is a talented singer with a beautiful voice.
+English: The baby is laughing and giggling.
+English: I need to do laundry and wash my clothes.
+English: They are planning a trip to Europe.
+English: She is learning how to play the guitar.
+English: We are going to a museum this Sunday.
+English: The coffee smells amazing in the morning.
+English: He is a hardworking farmer who grows crops.
+English: The students are presenting their research projects.
+English: I enjoy playing soccer with my friends.
+English: They are volunteering at a local shelter.
+English: She is practicing martial arts for self-defense.
+English: We should try a new recipe for dinner.
+English: The traffic is congest
+English: The sun is shining brightly today.
+English: I enjoy reading books in my free time.
+English: She plays the piano beautifully.
+English: The cat chased the mouse around the room.
+English: I love eating pizza with extra cheese.
+English: He always wears a hat wherever he goes.
+English: The flowers in the garden are blooming.
+English: She danced gracefully on the stage.
+English: The dog barked loudly in the park.
+English: We went swimming in the ocean yesterday.
+English: He speaks fluent French and Spanish.
+English: The train arrived at the station on time.
+English: She cooked a delicious meal for her family.
+Korean: 이것은 테스트 이다.
+Korean: 걱정할 필요 없다.
+Korean: 버그는 언젠가 고쳐진다.
+Japanese: 明日の天気はどうですか。
+Chinese: 请问洗手间在哪里？
+Emoji: I'm feeling 😄 today!
+Unicode: ◑ ▢ ▣ ◱

ggml/examples/prompts/tokenize_huggingface.py

@@ -0,0 +1,65 @@
+import os
+from transformers import AutoTokenizer
+
+os.environ['TOKENIZERS_PARALLELISM'] = "false"
+
+list_repo_hf  = ["databricks/dolly-v2-3b",           # dolly-v2 (3b, 7b, 12b models share the same tokenizer)
+                 "gpt2",                             # gpt-2 (gpt2-xl, gpt2-large share the same tokenizer)
+                 "uer/gpt2-chinese-cluecorpussmall", # gpt-2-chinese
+                 "EleutherAI/gpt-j-6b",              # gpt-j
+                 "EleutherAI/gpt-neox-20b",          # gpt-neox
+                 "EleutherAI/polyglot-ko-1.3b",      # gpt-neox (polyglot-ko 5.8b and 12.8b share the same tokenizer")
+                 "rinna/japanese-gpt-neox-3.6b",     # gpt-neox
+                 # mpt-7b (uses gpt-neox-20b tokenizer)
+                 "replit/replit-code-v1-3b",         # replit
+                 "bigcode/starcoder",                # starcoder (huggingface-cli login required)
+                 "openai/whisper-tiny"               # whisper (base, large, large-v2 share the same tokenizer)
+                 ]
+
+repo2ggml     = {"databricks/dolly-v2-3b"           : "dolly-v2",
+                 "gpt2"                             : "gpt-2",
+                 "uer/gpt2-chinese-cluecorpussmall" : "gpt-2-chinese",
+                 "EleutherAI/gpt-j-6b"              : "gpt-j",
+                 "EleutherAI/gpt-neox-20b"          : "gpt-neox",
+                 "EleutherAI/polyglot-ko-1.3b"      : "polyglot-ko",
+                 "rinna/japanese-gpt-neox-3.6b"     : "gpt-neox-japanese",
+                 "replit/replit-code-v1-3b"         : "replit",
+                 "bigcode/starcoder"                : "starcoder",
+                 "openai/whisper-tiny"              : "whisper"}
+
+repo2language = {"databricks/dolly-v2-3b"           : "english",
+                 "gpt2"                             : "english",
+                 "uer/gpt2-chinese-cluecorpussmall" : "chinese",
+                 "EleutherAI/gpt-j-6b"              : "english",
+                 "EleutherAI/gpt-neox-20b"          : "english",
+                 "EleutherAI/polyglot-ko-1.3b"      : "korean",
+                 "rinna/japanese-gpt-neox-3.6b"     : "japanese",
+                 "replit/replit-code-v1-3b"         : "english",
+                 "bigcode/starcoder"                : "english",
+                 "openai/whisper-tiny"              : "english"}
+
+delimeter = ": "
+test_sentences = []
+with open("test-cases.txt", "r") as f:
+    lines = [l.rstrip() for l in f.readlines()]
+    for l in lines:
+        if delimeter in l:
+            language = l[:l.index(delimeter)]
+            sentence = l[l.index(delimeter) + len(delimeter):]
+            test_sentences.append((language.lower(), sentence))
+
+for repo in list_repo_hf:
+
+    target_language = repo2language[repo]
+
+    tokenizer = AutoTokenizer.from_pretrained(repo, trust_remote_code=True)
+
+    tokens_hf = []
+    for language, sentence in test_sentences:
+        if language == target_language:
+            tokens = tokenizer.convert_tokens_to_ids(tokenizer.tokenize(sentence))
+            tokens_hf.append((sentence, tokens))
+
+    save_txt = repo2ggml[repo] + ".txt"
+    with open(save_txt, "w") as f:
+        f.writelines([sentence + " => " + ",".join(str(t) for t in tokens) + "\n" for sentence, tokens in tokens_hf])

ggml/examples/prompts/whisper.txt

@@ -0,0 +1,100 @@
+Hello World! => 15947,3937,0
+I can't believe it's already Friday!" => 40,393,380,1697,309,311,1217,6984,2963
+The URL for the website is https://www.example.com." => 2278,12905,337,220,3322,3144,307,34426,21492,17919,13,3121,335,781,13,1112,889
+"She said, 'I love to travel.'" => 1,9526,848,11,922,40,959,220,1353,220,17227,779,28763
+'The temperature is 25.5°C.' => 6,2278,220,18275,610,1503,307,3552,13,20,11782,34,4443
+"Let's meet at 2:30 p.m. in the park." => 1,8373,311,1677,412,568,25,3446,280,13,76,13,294,220,3322,3884,889
+The book costs $19.99 => 2278,1446,5497,1848,3405,13,8494
+"John's favorite color is blue." => 1,16938,311,2954,2017,307,3344,889
+Th@nk y0u f0r y0ur h3lp! => 2434,31,77,74,288,15,84,283,15,81,288,15,374,276,18,75,79,0
+C@n I g3t a c0ffee, pl3@se? => 34,31,77,286,290,18,83,257,269,15,4617,11,499,18,31,405,30
+W0w! Th@t's @m@zing! => 54,15,86,0,334,31,83,311,10428,76,31,8781,0
+H0w 4re y0u t0d@y? => 39,15,86,1017,265,288,15,84,220,83,15,67,31,88,30
+I l0ve t0 tr@vel @r0und the w0rld. => 40,287,15,303,220,83,15,220,6903,31,779,10428,81,15,997,220,3322,261,15,81,348,13
+Wh@t's y0ur f@v0rite m0vie? => 2471,31,83,311,288,15,374,283,31,85,15,35002,275,15,12702,30
+The cat is sleeping on the mat. => 2278,3857,307,8296,322,220,3322,3803,13
+I need to buy some groceries for dinner. => 40,643,220,1353,2256,512,31391,337,6148,13
+The sun is shining brightly in the sky. => 2278,3295,307,18269,47418,294,220,3322,5443,13
+She is reading a book in the park. => 9526,307,3760,257,1446,294,220,3322,3884,13
+We went for a walk on the beach yesterday. => 4360,1437,337,257,1792,322,220,3322,7534,5186,13
+He plays the guitar like a pro. => 5205,5749,220,3322,7531,411,257,447,13
+They are going to the movies tonight. => 8829,366,516,220,1353,220,3322,6233,220,1756,397,13
+The flowers are blooming in the garden. => 2278,8085,366,45294,294,220,3322,7431,13
+I enjoy listening to classical music. => 40,2103,4764,220,1353,13735,1318,13
+We need to buy groceries for the week. => 4360,643,220,1353,2256,31391,337,220,3322,1243,13
+The dog is chasing its tail in circles. => 2278,3000,307,17876,1080,220,14430,294,13040,13
+She is wearing a beautiful red dress. => 9526,307,4769,257,2238,2182,5231,13
+He is a talented actor in Hollywood. => 5205,307,257,220,32831,6003,8747,294,11628,13
+The children are playing in the playground. => 2278,2227,366,2433,294,220,3322,24646,13
+I'm going to visit my grandparents this weekend. => 40,478,516,220,1353,3441,452,21876,220,11176,6711,13
+The coffee tastes bitter without sugar. => 2278,4982,220,83,40246,13871,1553,5076,13
+They are planning a surprise party for her. => 8829,366,5038,257,6365,3595,337,720,13
+She sings like an angel on stage. => 9526,23250,411,364,14250,322,3233,13
+We should take a vacation to relax. => 4360,820,220,27612,257,12830,220,1353,5789,13
+He is studying medicine at the university. => 5205,307,7601,7195,412,220,3322,5454,13
+The rain is pouring heavily outside. => 2278,4830,307,20450,10950,2380,13
+I enjoy watching romantic movies. => 40,2103,1976,13590,6233,13
+They are celebrating their anniversary today. => 8829,366,15252,220,3322,347,12962,220,83,378,320,13
+She dances gracefully to the music. => 9526,28322,10042,2277,220,1353,220,3322,1318,13
+He is an excellent basketball player. => 5205,307,364,7103,11767,4256,13
+The baby is sleeping soundly in the crib. => 2278,3186,307,8296,1626,356,294,220,3322,47163,13
+I need to finish my homework before dinner. => 40,643,220,1353,2413,452,14578,949,6148,13
+They are organizing a charity event next month. => 8829,366,17608,257,16863,2280,958,1618,13
+She is cooking a delicious meal for us. => 9526,307,6361,257,4809,6791,337,505,13
+We should go hiking in the mountains. => 4360,820,352,23784,294,220,3322,10233,13
+The car broke down on the way to work. => 2278,1032,6902,760,322,220,3322,636,220,1353,589,13
+He loves playing video games in his free time. => 5205,6752,2433,960,2813,294,702,1737,220,3766,13
+The birds are chirping in the trees. => 2278,9009,366,36682,294,220,3322,220,3599,279,13
+I want to learn how to play the piano. => 40,528,220,1353,1466,577,220,1353,862,220,3322,9211,13
+They are building a new shopping mall in the city. => 8829,366,2390,257,777,8688,16026,294,220,3322,2307,13
+She is writing a novel in her spare time. => 9526,307,3579,257,7613,294,720,13798,220,3766,13
+We are going to the zoo this Saturday. => 4360,366,516,220,1353,220,3322,25347,220,11176,8803,13
+The cake looks delicious with chocolate frosting. => 2278,5908,1542,4809,365,6215,37048,13
+He is a talented painter who sells his artwork. => 5205,307,257,220,32831,6003,26619,567,20897,702,15829,13
+The students are studying for their exams. => 2278,1731,366,7601,337,220,3322,347,20514,13
+I enjoy swimming in the ocean. => 40,2103,11989,294,220,3322,7810,13
+They are renovating their house. => 8829,366,18845,990,220,3322,347,1782,13
+She is practicing yoga to stay healthy. => 9526,307,11350,15128,220,1353,1754,4627,13
+We should plant flowers in the garden. => 4360,820,3709,8085,294,220,3322,7431,13
+The traffic is heavy during rush hour. => 2278,220,17227,3341,307,4676,1830,9300,1773,13
+He is a skilled chef who creates amazing dishes. => 5205,307,257,19690,10530,567,7829,2243,10814,13
+The baby is crawling on the floor. => 2278,3186,307,32979,322,220,3322,4123,13
+I need to buy a new pair of shoes. => 40,643,220,1353,2256,257,777,6119,295,6654,13
+They are going on a road trip across the country. => 8829,366,516,322,257,3060,220,83,8400,2108,220,3322,1941,13
+She is playing the piano beautifully. => 9526,307,2433,220,3322,9211,16525,13
+We are going to a concert tomorrow night. => 4360,366,516,220,1353,257,8543,220,83,298,3162,1818,13
+The cake tastes delicious with vanilla frosting. => 2278,5908,220,83,40246,4809,365,17528,37048,13
+He is a dedicated teacher who inspires his students. => 5205,307,257,8374,220,975,4062,567,32566,702,1731,13
+The students are participating in a science fair. => 2278,1731,366,13950,294,257,3497,3143,13
+I enjoy hiking in the mountains. => 40,2103,23784,294,220,3322,10233,13
+They are organizing a beach cleanup next weekend. => 8829,366,17608,257,7534,40991,958,6711,13
+She is taking photographs of nature. => 9526,307,220,48625,17649,295,3687,13
+We should try a new restaurant in town. => 4360,820,220,83,627,257,777,6383,294,220,30401,13
+The traffic is moving slowly on the highway. => 2278,220,17227,3341,307,2684,5692,322,220,3322,17205,13
+He is a talented singer with a beautiful voice. => 5205,307,257,220,32831,6003,11564,365,257,2238,3177,13
+The baby is laughing and giggling. => 2278,3186,307,5059,293,290,24542,13
+I need to do laundry and wash my clothes. => 40,643,220,1353,360,19811,293,5675,452,5534,13
+They are planning a trip to Europe. => 8829,366,5038,257,220,83,8400,220,1353,3315,13
+She is learning how to play the guitar. => 9526,307,2539,577,220,1353,862,220,3322,7531,13
+We are going to a museum this Sunday. => 4360,366,516,220,1353,257,8441,220,11176,7776,13
+The coffee smells amazing in the morning. => 2278,4982,10036,2243,294,220,3322,2446,13
+He is a hardworking farmer who grows crops. => 5205,307,257,1152,22475,17891,567,13156,16829,13
+The students are presenting their research projects. => 2278,1731,366,15578,220,3322,347,2132,4455,13
+I enjoy playing soccer with my friends. => 40,2103,2433,15469,365,452,1855,13
+They are volunteering at a local shelter. => 8829,366,33237,412,257,2654,13341,13
+She is practicing martial arts for self-defense. => 9526,307,11350,20755,8609,337,2698,12,49268,13
+We should try a new recipe for dinner. => 4360,820,220,83,627,257,777,6782,337,6148,13
+The traffic is congest => 2278,220,17227,3341,307,31871
+The sun is shining brightly today. => 2278,3295,307,18269,47418,220,83,378,320,13
+I enjoy reading books in my free time. => 40,2103,3760,3642,294,452,1737,220,3766,13
+She plays the piano beautifully. => 9526,5749,220,3322,9211,16525,13
+The cat chased the mouse around the room. => 2278,3857,33091,220,3322,9719,926,220,3322,1808,13
+I love eating pizza with extra cheese. => 40,959,3936,8298,365,2857,5399,13
+He always wears a hat wherever he goes. => 5205,1009,20877,257,2385,8660,415,1709,13
+The flowers in the garden are blooming. => 2278,8085,294,220,3322,7431,366,45294,13
+She danced gracefully on the stage. => 9526,32909,10042,2277,322,220,3322,3233,13
+The dog barked loudly in the park. => 2278,3000,16202,292,22958,294,220,3322,3884,13
+We went swimming in the ocean yesterday. => 4360,1437,11989,294,220,3322,7810,5186,13
+He speaks fluent French and Spanish. => 5205,10789,40799,5522,293,8058,13
+The train arrived at the station on time. => 2278,220,83,7146,6678,412,220,3322,5214,322,220,3766,13
+She cooked a delicious meal for her family. => 9526,9267,257,4809,6791,337,720,1605,13

ggml/examples/python/README.md

@@ -0,0 +1,115 @@
+# Simple autogenerated Python bindings for ggml
+
+This folder contains:
+
+- Scripts to generate full Python bindings from ggml headers (+ stubs for autocompletion in IDEs)
+- Some barebones utils (see [ggml/utils.py](./ggml/utils.py)):
+  - `ggml.utils.init` builds a context that's freed automatically when the pointer gets GC'd
+  - `ggml.utils.copy` **copies between same-shaped tensors (numpy or ggml), w/ automatic (de/re)quantization**
+  - `ggml.utils.numpy` returns a numpy view over a ggml tensor; if it's quantized, it returns a copy (requires `allow_copy=True`)
+- Very basic examples (anyone wants to port [llama2.c](https://github.com/karpathy/llama2.c)?)
+
+Provided you set `GGML_LIBRARY=.../path/to/libggml_shared.so` (see instructions below), it's trivial to do some operations on quantized tensors:
+
+```python
+# Make sure libllama.so is in your [DY]LD_LIBRARY_PATH, or set GGML_LIBRARY=.../libggml_shared.so
+
+from ggml import lib, ffi
+from ggml.utils import init, copy, numpy
+import numpy as np
+
+ctx = init(mem_size=12*1024*1024)
+n = 256
+n_threads = 4
+
+a = lib.ggml_new_tensor_1d(ctx, lib.GGML_TYPE_Q5_K, n)
+b = lib.ggml_new_tensor_1d(ctx, lib.GGML_TYPE_F32, n) # Can't both be quantized
+sum = lib.ggml_add(ctx, a, b) # all zeroes for now. Will be quantized too!
+
+gf = ffi.new('struct ggml_cgraph*')
+lib.ggml_build_forward_expand(gf, sum)
+
+copy(np.array([i for i in range(n)], np.float32), a)
+copy(np.array([i*100 for i in range(n)], np.float32), b)
+
+lib.ggml_graph_compute_with_ctx(ctx, gf, n_threads)
+
+print(numpy(a, allow_copy=True))
+#  0.    1.0439453   2.0878906   3.131836    4.1757812   5.2197266. ...
+print(numpy(b))
+#  0.  100.        200.        300.        400.        500.         ...
+print(numpy(sum, allow_copy=True))
+#  0.  105.4375    210.875     316.3125    421.75      527.1875     ...
+```
+
+### Prerequisites
+
+You'll need a shared library of ggml to use the bindings.
+
+#### Build libggml_shared.so or libllama.so
+
+As of this writing the best is to use [ggerganov/llama.cpp](https://github.com/ggerganov/llama.cpp)'s generated `libggml_shared.so` or `libllama.so`, which you can build as follows:
+
+```bash
+git clone https://github.com/ggerganov/llama.cpp
+# On a CUDA-enabled system add -DLLAMA_CUBLAS=1
+# On a Mac add -DLLAMA_METAL=1
+cmake llama.cpp \
+  -B llama_build \
+  -DCMAKE_C_FLAGS=-Ofast \
+  -DLLAMA_NATIVE=1 \
+  -DLLAMA_LTO=1 \
+  -DBUILD_SHARED_LIBS=1 \
+  -DLLAMA_MPI=1 \
+  -DLLAMA_BUILD_TESTS=0 \
+  -DLLAMA_BUILD_EXAMPLES=0
+( cd llama_build && make -j )
+
+# On Mac, this will be libggml_shared.dylib instead
+export GGML_LIBRARY=$PWD/llama_build/libggml_shared.so
+# Alternatively, you can just copy it to your system's lib dir, e.g /usr/local/lib
+```
+
+#### (Optional) Regenerate the bindings and stubs
+
+If you added or changed any signatures of the C API, you'll want to regenerate the bindings ([ggml/cffi.py](./ggml/cffi.py)) and stubs ([ggml/__init__.pyi](./ggml/__init__.pyi)).
+
+Luckily it's a one-liner using [regenerate.py](./regenerate.py):
+
+```bash
+pip install -q cffi
+
+python regenerate.py
+```
+
+By default it assumes `llama.cpp` was cloned in ../../../llama.cpp (alongside the ggml folder). You can override this with:
+
+```bash
+C_INCLUDE_DIR=$LLAMA_CPP_DIR python regenerate.py
+```
+
+You can also edit [api.h](./api.h) to control which files should be included in the generated bindings (defaults to `llama.cpp/ggml*.h`)
+
+In fact, if you wanted to only generate bindings for the current version of the `ggml` repo itself (instead of `llama.cpp`; you'd loose support for k-quants), you could run:
+
+```bash
+API=../../include/ggml/ggml.h python regenerate.py
+```
+
+## Develop
+
+Run tests:
+
+```bash
+pytest
+```
+
+### Alternatives
+
+This example's goal is to showcase [cffi](https://cffi.readthedocs.io/)-generated bindings that are trivial to use and update, but there are already alternatives in the wild:
+
+- https://github.com/abetlen/ggml-python: these bindings seem to be hand-written and use [ctypes](https://docs.python.org/3/library/ctypes.html). It has [high-quality API reference docs](https://ggml-python.readthedocs.io/en/latest/api-reference/#ggml.ggml) that can be used with these bindings too, but it doesn't expose Metal, CUDA, MPI or OpenCL calls, doesn't support transparent (de/re)quantization like this example does (see [ggml.utils](./ggml/utils.py) module), and won't pick up your local changes.
+  
+- https://github.com/abetlen/llama-cpp-python: these expose the C++ `llama.cpp` interface, which this example cannot easily be extended to support (`cffi` only generates bindings of C libraries)
+
+- [pybind11](https://github.com/pybind/pybind11) and [nanobind](https://github.com/wjakob/nanobind) are two alternatives to cffi that support binding C++ libraries, but it doesn't seem either of them have an automatic generator (writing bindings is rather time-consuming).

ggml/examples/python/api.h

@@ -0,0 +1,14 @@
+/*
+  List here all the headers you want to expose in the Python bindings,
+  then run `python regenerate.py` (see details in README.md)
+*/
+
+#include "ggml.h"
+#include "ggml-metal.h"
+#include "ggml-opencl.h"
+
+// Headers below are currently only present in the llama.cpp repository, comment them out if you don't have them.
+#include "k_quants.h"
+#include "ggml-alloc.h"
+#include "ggml-cuda.h"
+#include "ggml-mpi.h"

ggml/examples/python/example_add_quant.py

@@ -0,0 +1,25 @@
+from ggml import lib, ffi
+from ggml.utils import init, copy, numpy
+import numpy as np
+
+ctx = init(mem_size=12*1024*1024) # automatically freed when pointer is GC'd
+n = 256
+n_threads = 4
+
+a = lib.ggml_new_tensor_1d(ctx, lib.GGML_TYPE_Q5_K, n)
+b = lib.ggml_new_tensor_1d(ctx, lib.GGML_TYPE_F32, n) # can't both be quantized
+sum = lib.ggml_add(ctx, a, b) # all zeroes for now. Will be quantized too!
+
+# See cffi's doc on how to allocate native memory: it's very simple!
+# https://cffi.readthedocs.io/en/latest/ref.html#ffi-interface
+gf = ffi.new('struct ggml_cgraph*')
+lib.ggml_build_forward_expand(gf, sum)
+
+copy(np.array([i for i in range(n)], np.float32), a)
+copy(np.array([i*100 for i in range(n)], np.float32), b)
+
+lib.ggml_graph_compute_with_ctx(ctx, gf, n_threads)
+
+print(numpy(a, allow_copy=True))
+print(numpy(b))
+print(numpy(sum, allow_copy=True))

ggml/examples/python/example_test_all_quants.py

@@ -0,0 +1,68 @@
+from ggml import ffi, lib
+from ggml.utils import init, numpy, copy
+import numpy as np
+from math import pi, cos, sin, ceil
+
+import matplotlib.pyplot as plt
+
+ctx = init(mem_size=100*1024*1024) # Will be auto-GC'd
+n = 256
+
+orig = np.array([
+    [
+        cos(j * 2 * pi / n) * (sin(i * 2 * pi / n))
+        for j in range(n)
+    ]
+    for i in range(n)
+], np.float32)
+orig_tensor = lib.ggml_new_tensor_2d(ctx, lib.GGML_TYPE_F32, n, n)
+copy(orig, orig_tensor)
+
+quants = [
+    type for type in range(lib.GGML_TYPE_COUNT)
+    if lib.ggml_is_quantized(type) and
+       type not in [lib.GGML_TYPE_Q8_1, lib.GGML_TYPE_Q8_K] # Apparently not supported
+]
+# quants = [lib.GGML_TYPE_Q2_K] # Test a single one
+
+def get_name(type):
+    name = lib.ggml_type_name(type)
+    return ffi.string(name).decode('utf-8') if name else '?'
+
+quants.sort(key=get_name)
+quants.insert(0, None)
+print(quants)
+
+ncols=4
+nrows = ceil(len(quants) / ncols)
+
+plt.figure(figsize=(ncols * 5, nrows * 5), layout='tight')
+
+for i, type in enumerate(quants):
+    plt.subplot(nrows, ncols, i + 1)
+    try:
+        if type == None:
+            plt.title('Original')
+            plt.imshow(orig)
+        else:
+            quantized_tensor = lib.ggml_new_tensor_2d(ctx, type, n, n)
+            copy(orig_tensor, quantized_tensor)
+            quantized = numpy(quantized_tensor, allow_copy=True)
+            d = quantized - orig
+            results = {
+                "l2": np.linalg.norm(d, 2),
+                "linf": np.linalg.norm(d, np.inf),
+                "compression":
+                    round(lib.ggml_nbytes(orig_tensor) /
+                          lib.ggml_nbytes(quantized_tensor), 1)
+            }
+            name = get_name(type)
+            print(f'{name}: {results}')
+
+            plt.title(f'{name} ({results["compression"]}x smaller)')
+            plt.imshow(quantized, interpolation='nearest')
+        
+    except Exception as e:
+        print(f'Error: {e}')
+
+plt.show()

ggml/examples/python/ggml/__init__.py

@@ -0,0 +1,58 @@
+"""
+  Python bindings for the ggml library.
+
+  Usage example:
+
+      from ggml import lib, ffi
+      from ggml.utils import init, copy, numpy
+      import numpy as np
+
+      ctx = init(mem_size=10*1024*1024)
+      n = 1024
+      n_threads = 4
+
+      a = lib.ggml_new_tensor_1d(ctx, lib.GGML_TYPE_Q5_K, n)
+      b = lib.ggml_new_tensor_1d(ctx, lib.GGML_TYPE_F32, n)
+      sum = lib.ggml_add(ctx, a, b)
+
+      gf = ffi.new('struct ggml_cgraph*')
+      lib.ggml_build_forward_expand(gf, sum)
+
+      copy(np.array([i for i in range(n)], np.float32), a)
+      copy(np.array([i*100 for i in range(n)], np.float32), b)
+      lib.ggml_graph_compute_with_ctx(ctx, gf, n_threads)
+
+      print(numpy(sum, allow_copy=True))
+
+  See https://cffi.readthedocs.io/en/latest/cdef.html for more on cffi.
+"""
+
+try:
+    from ggml.cffi import ffi as ffi
+except ImportError as e:
+    raise ImportError(f"Couldn't find ggml bindings ({e}). Run `python regenerate.py` or check your PYTHONPATH.")
+
+import os, platform
+
+__exact_library = os.environ.get("GGML_LIBRARY")
+if __exact_library:
+    __candidates = [__exact_library]
+elif platform.system() == "Windows":
+    __candidates = ["ggml_shared.dll", "llama.dll"]
+else:
+    __candidates = ["libggml_shared.so", "libllama.so"]
+    if platform.system() == "Darwin":
+        __candidates += ["libggml_shared.dylib", "libllama.dylib"]
+
+for i, name in enumerate(__candidates):
+    try:
+        # This is where all the functions, enums and constants are defined
+        lib = ffi.dlopen(name)
+    except OSError:
+        if i < len(__candidates) - 1:
+            continue
+        raise OSError(f"Couldn't find ggml's shared library (tried names: {__candidates}). Add its directory to DYLD_LIBRARY_PATH (on Mac) or LD_LIBRARY_PATH, or define GGML_LIBRARY.")
+
+# This contains the cffi helpers such as new, cast, string, etc.
+# https://cffi.readthedocs.io/en/latest/ref.html#ffi-interface
+ffi = ffi

ggml/examples/python/ggml/__init__.pyi

@@ -0,0 +1,2431 @@
+# auto-generated file
+import ggml.ffi as ffi
+import numpy as np
+class lib:
+  @property
+  def GGML_BACKEND_CPU(self) -> int: ...
+  @property
+  def GGML_BACKEND_GPU(self) -> int: ...
+  @property
+  def GGML_BACKEND_GPU_SPLIT(self) -> int: ...
+  @property
+  def GGML_FTYPE_ALL_F32(self) -> int: ...
+  @property
+  def GGML_FTYPE_MOSTLY_F16(self) -> int: ...
+  @property
+  def GGML_FTYPE_MOSTLY_Q2_K(self) -> int: ...
+  @property
+  def GGML_FTYPE_MOSTLY_Q3_K(self) -> int: ...
+  @property
+  def GGML_FTYPE_MOSTLY_Q4_0(self) -> int: ...
+  @property
+  def GGML_FTYPE_MOSTLY_Q4_1(self) -> int: ...
+  @property
+  def GGML_FTYPE_MOSTLY_Q4_1_SOME_F16(self) -> int: ...
+  @property
+  def GGML_FTYPE_MOSTLY_Q4_K(self) -> int: ...
+  @property
+  def GGML_FTYPE_MOSTLY_Q5_0(self) -> int: ...
+  @property
+  def GGML_FTYPE_MOSTLY_Q5_1(self) -> int: ...
+  @property
+  def GGML_FTYPE_MOSTLY_Q5_K(self) -> int: ...
+  @property
+  def GGML_FTYPE_MOSTLY_Q6_K(self) -> int: ...
+  @property
+  def GGML_FTYPE_MOSTLY_Q8_0(self) -> int: ...
+  @property
+  def GGML_FTYPE_UNKNOWN(self) -> int: ...
+  @property
+  def GGML_LINESEARCH_BACKTRACKING_ARMIJO(self) -> int: ...
+  @property
+  def GGML_LINESEARCH_BACKTRACKING_STRONG_WOLFE(self) -> int: ...
+  @property
+  def GGML_LINESEARCH_BACKTRACKING_WOLFE(self) -> int: ...
+  @property
+  def GGML_LINESEARCH_DEFAULT(self) -> int: ...
+  @property
+  def GGML_LINESEARCH_FAIL(self) -> int: ...
+  @property
+  def GGML_LINESEARCH_INVALID_PARAMETERS(self) -> int: ...
+  @property
+  def GGML_LINESEARCH_MAXIMUM_ITERATIONS(self) -> int: ...
+  @property
+  def GGML_LINESEARCH_MAXIMUM_STEP(self) -> int: ...
+  @property
+  def GGML_LINESEARCH_MINIMUM_STEP(self) -> int: ...
+  @property
+  def GGML_OBJECT_GRAPH(self) -> int: ...
+  @property
+  def GGML_OBJECT_TENSOR(self) -> int: ...
+  @property
+  def GGML_OBJECT_WORK_BUFFER(self) -> int: ...
+  @property
+  def GGML_OPT_ADAM(self) -> int: ...
+  @property
+  def GGML_OPT_DID_NOT_CONVERGE(self) -> int: ...
+  @property
+  def GGML_OPT_FAIL(self) -> int: ...
+  @property
+  def GGML_OPT_INVALID_WOLFE(self) -> int: ...
+  @property
+  def GGML_OPT_LBFGS(self) -> int: ...
+  @property
+  def GGML_OPT_NO_CONTEXT(self) -> int: ...
+  @property
+  def GGML_OPT_OK(self) -> int: ...
+  @property
+  def GGML_OP_ACC(self) -> int: ...
+  @property
+  def GGML_OP_ADD(self) -> int: ...
+  @property
+  def GGML_OP_ADD1(self) -> int: ...
+  @property
+  def GGML_OP_ALIBI(self) -> int: ...
+  @property
+  def GGML_OP_ARGMAX(self) -> int: ...
+  @property
+  def GGML_OP_CLAMP(self) -> int: ...
+  @property
+  def GGML_OP_CONT(self) -> int: ...
+  @property
+  def GGML_OP_CONV_1D(self) -> int: ...
+  @property
+  def GGML_OP_CONV_2D(self) -> int: ...
+  @property
+  def GGML_OP_COUNT(self) -> int: ...
+  @property
+  def GGML_OP_CPY(self) -> int: ...
+  @property
+  def GGML_OP_CROSS_ENTROPY_LOSS(self) -> int: ...
+  @property
+  def GGML_OP_CROSS_ENTROPY_LOSS_BACK(self) -> int: ...
+  @property
+  def GGML_OP_DIAG(self) -> int: ...
+  @property
+  def GGML_OP_DIAG_MASK_INF(self) -> int: ...
+  @property
+  def GGML_OP_DIAG_MASK_ZERO(self) -> int: ...
+  @property
+  def GGML_OP_DIV(self) -> int: ...
+  @property
+  def GGML_OP_DUP(self) -> int: ...
+  @property
+  def GGML_OP_FLASH_ATTN(self) -> int: ...
+  @property
+  def GGML_OP_FLASH_ATTN_BACK(self) -> int: ...
+  @property
+  def GGML_OP_FLASH_FF(self) -> int: ...
+  @property
+  def GGML_OP_GET_ROWS(self) -> int: ...
+  @property
+  def GGML_OP_GET_ROWS_BACK(self) -> int: ...
+  @property
+  def GGML_OP_LOG(self) -> int: ...
+  @property
+  def GGML_OP_MAP_BINARY(self) -> int: ...
+  @property
+  def GGML_OP_MAP_CUSTOM1(self) -> int: ...
+  @property
+  def GGML_OP_MAP_CUSTOM1_F32(self) -> int: ...
+  @property
+  def GGML_OP_MAP_CUSTOM2(self) -> int: ...
+  @property
+  def GGML_OP_MAP_CUSTOM2_F32(self) -> int: ...
+  @property
+  def GGML_OP_MAP_CUSTOM3(self) -> int: ...
+  @property
+  def GGML_OP_MAP_CUSTOM3_F32(self) -> int: ...
+  @property
+  def GGML_OP_MAP_UNARY(self) -> int: ...
+  @property
+  def GGML_OP_MEAN(self) -> int: ...
+  @property
+  def GGML_OP_MUL(self) -> int: ...
+  @property
+  def GGML_OP_MUL_MAT(self) -> int: ...
+  @property
+  def GGML_OP_NONE(self) -> int: ...
+  @property
+  def GGML_OP_NORM(self) -> int: ...
+  @property
+  def GGML_OP_OUT_PROD(self) -> int: ...
+  @property
+  def GGML_OP_PERMUTE(self) -> int: ...
+  @property
+  def GGML_OP_POOL_1D(self) -> int: ...
+  @property
+  def GGML_OP_POOL_2D(self) -> int: ...
+  @property
+  def GGML_OP_POOL_AVG(self) -> int: ...
+  @property
+  def GGML_OP_POOL_COUNT(self) -> int: ...
+  @property
+  def GGML_OP_POOL_MAX(self) -> int: ...
+  @property
+  def GGML_OP_REPEAT(self) -> int: ...
+  @property
+  def GGML_OP_REPEAT_BACK(self) -> int: ...
+  @property
+  def GGML_OP_RESHAPE(self) -> int: ...
+  @property
+  def GGML_OP_RMS_NORM(self) -> int: ...
+  @property
+  def GGML_OP_RMS_NORM_BACK(self) -> int: ...
+  @property
+  def GGML_OP_ROPE(self) -> int: ...
+  @property
+  def GGML_OP_ROPE_BACK(self) -> int: ...
+  @property
+  def GGML_OP_SCALE(self) -> int: ...
+  @property
+  def GGML_OP_SET(self) -> int: ...
+  @property
+  def GGML_OP_SILU_BACK(self) -> int: ...
+  @property
+  def GGML_OP_SOFT_MAX(self) -> int: ...
+  @property
+  def GGML_OP_SOFT_MAX_BACK(self) -> int: ...
+  @property
+  def GGML_OP_SQR(self) -> int: ...
+  @property
+  def GGML_OP_SQRT(self) -> int: ...
+  @property
+  def GGML_OP_SUB(self) -> int: ...
+  @property
+  def GGML_OP_SUM(self) -> int: ...
+  @property
+  def GGML_OP_SUM_ROWS(self) -> int: ...
+  @property
+  def GGML_OP_TRANSPOSE(self) -> int: ...
+  @property
+  def GGML_OP_UNARY(self) -> int: ...
+  @property
+  def GGML_OP_VIEW(self) -> int: ...
+  @property
+  def GGML_OP_WIN_PART(self) -> int: ...
+  @property
+  def GGML_OP_WIN_UNPART(self) -> int: ...
+  @property
+  def GGML_TASK_COMPUTE(self) -> int: ...
+  @property
+  def GGML_TASK_FINALIZE(self) -> int: ...
+  @property
+  def GGML_TASK_INIT(self) -> int: ...
+  @property
+  def GGML_TYPE_COUNT(self) -> int: ...
+  @property
+  def GGML_TYPE_F16(self) -> int: ...
+  @property
+  def GGML_TYPE_F32(self) -> int: ...
+  @property
+  def GGML_TYPE_I16(self) -> int: ...
+  @property
+  def GGML_TYPE_I32(self) -> int: ...
+  @property
+  def GGML_TYPE_I8(self) -> int: ...
+  @property
+  def GGML_TYPE_Q2_K(self) -> int: ...
+  @property
+  def GGML_TYPE_Q3_K(self) -> int: ...
+  @property
+  def GGML_TYPE_Q4_0(self) -> int: ...
+  @property
+  def GGML_TYPE_Q4_1(self) -> int: ...
+  @property
+  def GGML_TYPE_Q4_K(self) -> int: ...
+  @property
+  def GGML_TYPE_Q5_0(self) -> int: ...
+  @property
+  def GGML_TYPE_Q5_1(self) -> int: ...
+  @property
+  def GGML_TYPE_Q5_K(self) -> int: ...
+  @property
+  def GGML_TYPE_Q6_K(self) -> int: ...
+  @property
+  def GGML_TYPE_Q8_0(self) -> int: ...
+  @property
+  def GGML_TYPE_Q8_1(self) -> int: ...
+  @property
+  def GGML_TYPE_Q8_K(self) -> int: ...
+  @property
+  def GGML_UNARY_OP_ABS(self) -> int: ...
+  @property
+  def GGML_UNARY_OP_ELU(self) -> int: ...
+  @property
+  def GGML_UNARY_OP_GELU(self) -> int: ...
+  @property
+  def GGML_UNARY_OP_GELU_QUICK(self) -> int: ...
+  @property
+  def GGML_UNARY_OP_NEG(self) -> int: ...
+  @property
+  def GGML_UNARY_OP_RELU(self) -> int: ...
+  @property
+  def GGML_UNARY_OP_SGN(self) -> int: ...
+  @property
+  def GGML_UNARY_OP_SILU(self) -> int: ...
+  @property
+  def GGML_UNARY_OP_STEP(self) -> int: ...
+  @property
+  def GGML_UNARY_OP_TANH(self) -> int: ...
+  @property
+  def GGUF_TYPE_ARRAY(self) -> int: ...
+  @property
+  def GGUF_TYPE_BOOL(self) -> int: ...
+  @property
+  def GGUF_TYPE_COUNT(self) -> int: ...
+  @property
+  def GGUF_TYPE_FLOAT32(self) -> int: ...
+  @property
+  def GGUF_TYPE_INT16(self) -> int: ...
+  @property
+  def GGUF_TYPE_INT32(self) -> int: ...
+  @property
+  def GGUF_TYPE_INT8(self) -> int: ...
+  @property
+  def GGUF_TYPE_STRING(self) -> int: ...
+  @property
+  def GGUF_TYPE_UINT16(self) -> int: ...
+  @property
+  def GGUF_TYPE_UINT32(self) -> int: ...
+  @property
+  def GGUF_TYPE_UINT8(self) -> int: ...
+  def abort_callback(data: ffi.CData) -> bool:
+    """
+    abort ggml_graph_compute when true
+
+            bool (*abort_callback)(void * data);
+    """
+    ...
+  def dequantize_row_q2_K(x: ffi.CData, y: ffi.CData, k: int) -> None:
+    """
+    Dequantization
+
+    void dequantize_row_q2_K(const block_q2_K * restrict x, float * restrict y, int k);
+    """
+    ...
+  def dequantize_row_q3_K(x: ffi.CData, y: ffi.CData, k: int) -> None:
+    """void dequantize_row_q3_K(const block_q3_K * restrict x, float * restrict y, int k);"""
+    ...
+  def dequantize_row_q4_K(x: ffi.CData, y: ffi.CData, k: int) -> None:
+    """void dequantize_row_q4_K(const block_q4_K * restrict x, float * restrict y, int k);"""
+    ...
+  def dequantize_row_q5_K(x: ffi.CData, y: ffi.CData, k: int) -> None:
+    """void dequantize_row_q5_K(const block_q5_K * restrict x, float * restrict y, int k);"""
+    ...
+  def dequantize_row_q6_K(x: ffi.CData, y: ffi.CData, k: int) -> None:
+    """void dequantize_row_q6_K(const block_q6_K * restrict x, float * restrict y, int k);"""
+    ...
+  def dequantize_row_q8_K(x: ffi.CData, y: ffi.CData, k: int) -> None:
+    """void dequantize_row_q8_K(const block_q8_K * restrict x, float * restrict y, int k);"""
+    ...
+  def ggml_abs(ctx: ffi.CData, a: ffi.CData) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_abs(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a);
+    """
+    ...
+  def ggml_abs_inplace(ctx: ffi.CData, a: ffi.CData) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_abs_inplace(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a);
+    """
+    ...
+  def ggml_acc(ctx: ffi.CData, a: ffi.CData, b: ffi.CData, nb1: int, nb2: int, nb3: int, offset: int) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_acc(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a,
+                struct ggml_tensor  * b,
+                size_t                nb1,
+                size_t                nb2,
+                size_t                nb3,
+                size_t                offset);
+    """
+    ...
+  def ggml_acc_inplace(ctx: ffi.CData, a: ffi.CData, b: ffi.CData, nb1: int, nb2: int, nb3: int, offset: int) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_acc_inplace(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a,
+                struct ggml_tensor  * b,
+                size_t                nb1,
+                size_t                nb2,
+                size_t                nb3,
+                size_t                offset);
+    """
+    ...
+  def ggml_add(ctx: ffi.CData, a: ffi.CData, b: ffi.CData) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_add(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a,
+                struct ggml_tensor  * b);
+    """
+    ...
+  def ggml_add1(ctx: ffi.CData, a: ffi.CData, b: ffi.CData) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_add1(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a,
+                struct ggml_tensor  * b);
+    """
+    ...
+  def ggml_add1_inplace(ctx: ffi.CData, a: ffi.CData, b: ffi.CData) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_add1_inplace(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a,
+                struct ggml_tensor  * b);
+    """
+    ...
+  def ggml_add_inplace(ctx: ffi.CData, a: ffi.CData, b: ffi.CData) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_add_inplace(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a,
+                struct ggml_tensor  * b);
+    """
+    ...
+  def ggml_alibi(ctx: ffi.CData, a: ffi.CData, n_past: int, n_head: int, bias_max: float) -> ffi.CData:
+    """
+    alibi position embedding
+    in-place, returns view(a)
+
+        struct ggml_tensor * ggml_alibi(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a,
+                int                   n_past,
+                int                   n_head,
+                float                 bias_max);
+    """
+    ...
+  def ggml_allocr_alloc(alloc: ffi.CData, tensor: ffi.CData) -> None:
+    """GGML_API void   ggml_allocr_alloc(struct ggml_allocr * alloc, struct ggml_tensor * tensor);"""
+    ...
+  def ggml_allocr_alloc_graph(alloc: ffi.CData, graph: ffi.CData) -> int:
+    """GGML_API size_t ggml_allocr_alloc_graph(struct ggml_allocr * alloc, struct ggml_cgraph * graph);"""
+    ...
+  def ggml_allocr_free(alloc: ffi.CData) -> None:
+    """GGML_API void   ggml_allocr_free(struct ggml_allocr * alloc);"""
+    ...
+  def ggml_allocr_is_measure(alloc: ffi.CData) -> bool:
+    """GGML_API bool   ggml_allocr_is_measure(struct ggml_allocr * alloc);"""
+    ...
+  def ggml_allocr_new(data: ffi.CData, size: int, alignment: int) -> ffi.CData:
+    """GGML_API struct ggml_allocr * ggml_allocr_new(void * data, size_t size, size_t alignment);"""
+    ...
+  def ggml_allocr_new_measure(alignment: int) -> ffi.CData:
+    """GGML_API struct ggml_allocr * ggml_allocr_new_measure(size_t alignment);"""
+    ...
+  def ggml_allocr_reset(alloc: ffi.CData) -> None:
+    """GGML_API void   ggml_allocr_reset(struct ggml_allocr * alloc);"""
+    ...
+  def ggml_allocr_set_parse_seq(alloc: ffi.CData, list: ffi.CData, n: int) -> None:
+    """
+    tell the allocator to parse nodes following the order described in the list
+    you should call this if your graph are optimized to execute out-of-order
+
+    GGML_API void   ggml_allocr_set_parse_seq(struct ggml_allocr * alloc, int * list, int n);
+    """
+    ...
+  def ggml_are_same_shape(t0: ffi.CData, t1: ffi.CData) -> bool:
+    """    GGML_API bool ggml_are_same_shape(const struct ggml_tensor * t0, const struct ggml_tensor * t1);"""
+    ...
+  def ggml_argmax(ctx: ffi.CData, a: ffi.CData) -> ffi.CData:
+    """
+    argmax along rows
+
+        GGML_API struct ggml_tensor * ggml_argmax(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a);
+    """
+    ...
+  def ggml_blck_size(type: int) -> int:
+    """    GGML_API int     ggml_blck_size (enum ggml_type type);"""
+    ...
+  def ggml_build_backward(ctx: ffi.CData, gf: ffi.CData, keep: bool) -> ffi.CData:
+    """    GGML_API struct ggml_cgraph ggml_build_backward(struct ggml_context * ctx, struct ggml_cgraph * gf, bool keep);"""
+    ...
+  def ggml_build_forward(tensor: ffi.CData) -> ffi.CData:
+    """    GGML_API struct ggml_cgraph ggml_build_forward (struct ggml_tensor * tensor);"""
+    ...
+  def ggml_build_forward_ctx(ctx: ffi.CData, tensor: ffi.CData) -> ffi.CData:
+    """    GGML_API struct ggml_cgraph * ggml_build_forward_ctx(struct ggml_context * ctx, struct ggml_tensor * tensor);"""
+    ...
+  def ggml_build_forward_expand(cgraph: ffi.CData, tensor: ffi.CData) -> None:
+    """    GGML_API void ggml_build_forward_expand(struct ggml_cgraph * cgraph, struct ggml_tensor * tensor);"""
+    ...
+  def ggml_cl_can_mul_mat(src0: ffi.CData, src1: ffi.CData, dst: ffi.CData) -> bool:
+    """bool   ggml_cl_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);"""
+    ...
+  def ggml_cl_free_data(tensor: ffi.CData) -> None:
+    """void ggml_cl_free_data(const struct ggml_tensor* tensor);"""
+    ...
+  def ggml_cl_host_free(ptr: ffi.CData) -> None:
+    """void   ggml_cl_host_free(void * ptr);"""
+    ...
+  def ggml_cl_host_malloc(size: int) -> ffi.CData:
+    """void * ggml_cl_host_malloc(size_t size);"""
+    ...
+  def ggml_cl_init() -> None:
+    """void ggml_cl_init(void);"""
+    ...
+  def ggml_cl_mul(src0: ffi.CData, src1: ffi.CData, dst: ffi.CData) -> None:
+    """void   ggml_cl_mul(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);"""
+    ...
+  def ggml_cl_mul_mat(src0: ffi.CData, src1: ffi.CData, dst: ffi.CData, wdata: ffi.CData, wsize: int) -> None:
+    """void   ggml_cl_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst, void * wdata, size_t wsize);"""
+    ...
+  def ggml_cl_mul_mat_get_wsize(src0: ffi.CData, src1: ffi.CData, dst: ffi.CData) -> int:
+    """size_t ggml_cl_mul_mat_get_wsize(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);"""
+    ...
+  def ggml_cl_transform_tensor(data: ffi.CData, tensor: ffi.CData) -> None:
+    """void ggml_cl_transform_tensor(void * data, struct ggml_tensor * tensor);"""
+    ...
+  def ggml_clamp(ctx: ffi.CData, a: ffi.CData, min: float, max: float) -> ffi.CData:
+    """
+    clamp
+    in-place, returns view(a)
+
+        struct ggml_tensor * ggml_clamp(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a,
+                float                 min,
+                float                 max);
+    """
+    ...
+  def ggml_cont(ctx: ffi.CData, a: ffi.CData) -> ffi.CData:
+    """
+    make contiguous
+
+        GGML_API struct ggml_tensor * ggml_cont(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a);
+    """
+    ...
+  def ggml_cont_inplace(ctx: ffi.CData, a: ffi.CData) -> ffi.CData:
+    """
+    make contiguous, in-place
+
+        GGML_API struct ggml_tensor * ggml_cont_inplace(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a);
+    """
+    ...
+  def ggml_conv_1d(ctx: ffi.CData, a: ffi.CData, b: ffi.CData, s0: int, p0: int, d0: int) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_conv_1d(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a,
+                struct ggml_tensor  * b,
+                int                   s0,  // stride
+                int                   p0,  // padding
+                int                   d0); // dilation
+    """
+    ...
+  def ggml_conv_1d_ph(ctx: ffi.CData, a: ffi.CData, b: ffi.CData, s: int, d: int) -> ffi.CData:
+    """
+    conv_1d with padding = half
+    alias for ggml_conv_1d(a, b, s, a->ne[0]/2, d)
+
+        GGML_API struct ggml_tensor * ggml_conv_1d_ph(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a,
+                struct ggml_tensor  * b,
+                int                   s,
+                int                   d);
+    """
+    ...
+  def ggml_conv_2d(ctx: ffi.CData, a: ffi.CData, b: ffi.CData, s0: int, s1: int, p0: int, p1: int, d0: int, d1: int) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_conv_2d(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a,
+                struct ggml_tensor  * b,
+                int                   s0,
+                int                   s1,
+                int                   p0,
+                int                   p1,
+                int                   d0,
+                int                   d1);
+    """
+    ...
+  def ggml_cpu_has_arm_fma() -> int:
+    """    GGML_API int ggml_cpu_has_arm_fma    (void);"""
+    ...
+  def ggml_cpu_has_avx() -> int:
+    """    GGML_API int ggml_cpu_has_avx        (void);"""
+    ...
+  def ggml_cpu_has_avx2() -> int:
+    """    GGML_API int ggml_cpu_has_avx2       (void);"""
+    ...
+  def ggml_cpu_has_avx512() -> int:
+    """    GGML_API int ggml_cpu_has_avx512     (void);"""
+    ...
+  def ggml_cpu_has_avx512_vbmi() -> int:
+    """    GGML_API int ggml_cpu_has_avx512_vbmi(void);"""
+    ...
+  def ggml_cpu_has_avx512_vnni() -> int:
+    """    GGML_API int ggml_cpu_has_avx512_vnni(void);"""
+    ...
+  def ggml_cpu_has_blas() -> int:
+    """    GGML_API int ggml_cpu_has_blas       (void);"""
+    ...
+  def ggml_cpu_has_clblast() -> int:
+    """    GGML_API int ggml_cpu_has_clblast    (void);"""
+    ...
+  def ggml_cpu_has_cublas() -> int:
+    """    GGML_API int ggml_cpu_has_cublas     (void);"""
+    ...
+  def ggml_cpu_has_f16c() -> int:
+    """    GGML_API int ggml_cpu_has_f16c       (void);"""
+    ...
+  def ggml_cpu_has_fma() -> int:
+    """    GGML_API int ggml_cpu_has_fma        (void);"""
+    ...
+  def ggml_cpu_has_fp16_va() -> int:
+    """    GGML_API int ggml_cpu_has_fp16_va    (void);"""
+    ...
+  def ggml_cpu_has_gpublas() -> int:
+    """    GGML_API int ggml_cpu_has_gpublas    (void);"""
+    ...
+  def ggml_cpu_has_neon() -> int:
+    """    GGML_API int ggml_cpu_has_neon       (void);"""
+    ...
+  def ggml_cpu_has_sse3() -> int:
+    """    GGML_API int ggml_cpu_has_sse3       (void);"""
+    ...
+  def ggml_cpu_has_vsx() -> int:
+    """    GGML_API int ggml_cpu_has_vsx        (void);"""
+    ...
+  def ggml_cpu_has_wasm_simd() -> int:
+    """    GGML_API int ggml_cpu_has_wasm_simd  (void);"""
+    ...
+  def ggml_cpy(ctx: ffi.CData, a: ffi.CData, b: ffi.CData) -> ffi.CData:
+    """
+    a -> b, return view(b)
+
+        GGML_API struct ggml_tensor * ggml_cpy(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a,
+                struct ggml_tensor  * b);
+    """
+    ...
+  def ggml_cpy_inplace(ctx: ffi.CData, a: ffi.CData, b: ffi.CData) -> ffi.CData:
+    """
+    a -> b, in-place, return view(b)
+
+        GGML_API struct ggml_tensor * ggml_cpy_inplace(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a,
+                struct ggml_tensor  * b);
+    """
+    ...
+  def ggml_cross_entropy_loss(ctx: ffi.CData, a: ffi.CData, b: ffi.CData) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_cross_entropy_loss(
+                struct ggml_context         * ctx,
+                struct ggml_tensor          * a,
+                struct ggml_tensor          * b);
+    """
+    ...
+  def ggml_cross_entropy_loss_back(ctx: ffi.CData, a: ffi.CData, b: ffi.CData, c: ffi.CData) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_cross_entropy_loss_back(
+                struct ggml_context         * ctx,
+                struct ggml_tensor          * a,
+                struct ggml_tensor          * b,
+                struct ggml_tensor          * c);
+    """
+    ...
+  def ggml_cuda_assign_buffers(tensor: ffi.CData) -> None:
+    """GGML_API void   ggml_cuda_assign_buffers(struct ggml_tensor * tensor);"""
+    ...
+  def ggml_cuda_assign_buffers_force_inplace(tensor: ffi.CData) -> None:
+    """GGML_API void   ggml_cuda_assign_buffers_force_inplace(struct ggml_tensor * tensor);"""
+    ...
+  def ggml_cuda_assign_buffers_no_scratch(tensor: ffi.CData) -> None:
+    """GGML_API void   ggml_cuda_assign_buffers_no_scratch(struct ggml_tensor * tensor);"""
+    ...
+  def ggml_cuda_can_mul_mat(src0: ffi.CData, src1: ffi.CData, dst: ffi.CData) -> bool:
+    """GGML_API bool   ggml_cuda_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);"""
+    ...
+  def ggml_cuda_compute_forward(params: ffi.CData, tensor: ffi.CData) -> bool:
+    """GGML_API bool   ggml_cuda_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * tensor);"""
+    ...
+  def ggml_cuda_free_data(tensor: ffi.CData) -> None:
+    """GGML_API void   ggml_cuda_free_data(struct ggml_tensor * tensor);"""
+    ...
+  def ggml_cuda_free_scratch() -> None:
+    """GGML_API void   ggml_cuda_free_scratch(void);"""
+    ...
+  def ggml_cuda_get_device_count() -> int:
+    """GGML_API int    ggml_cuda_get_device_count(void);"""
+    ...
+  def ggml_cuda_get_device_description(device: int, description: ffi.CData, description_size: int) -> None:
+    """GGML_API void   ggml_cuda_get_device_description(int device, char * description, size_t description_size);"""
+    ...
+  def ggml_cuda_host_free(ptr: ffi.CData) -> None:
+    """GGML_API void   ggml_cuda_host_free(void * ptr);"""
+    ...
+  def ggml_cuda_host_malloc(size: int) -> ffi.CData:
+    """GGML_API void * ggml_cuda_host_malloc(size_t size);"""
+    ...
+  def ggml_cuda_set_main_device(main_device: int) -> None:
+    """GGML_API void   ggml_cuda_set_main_device(int main_device);"""
+    ...
+  def ggml_cuda_set_mul_mat_q(mul_mat_q: bool) -> None:
+    """GGML_API void   ggml_cuda_set_mul_mat_q(bool mul_mat_q);"""
+    ...
+  def ggml_cuda_set_scratch_size(scratch_size: int) -> None:
+    """GGML_API void   ggml_cuda_set_scratch_size(size_t scratch_size);"""
+    ...
+  def ggml_cuda_set_tensor_split(tensor_split: ffi.CData) -> None:
+    """GGML_API void   ggml_cuda_set_tensor_split(const float * tensor_split);"""
+    ...
+  def ggml_cuda_transform_tensor(data: ffi.CData, tensor: ffi.CData) -> None:
+    """GGML_API void   ggml_cuda_transform_tensor(void * data, struct ggml_tensor * tensor);"""
+    ...
+  def ggml_cycles() -> int:
+    """    GGML_API int64_t ggml_cycles(void);"""
+    ...
+  def ggml_cycles_per_ms() -> int:
+    """    GGML_API int64_t ggml_cycles_per_ms(void);"""
+    ...
+  def ggml_diag(ctx: ffi.CData, a: ffi.CData) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_diag(
+            struct ggml_context     * ctx,
+            struct ggml_tensor      * a);
+    """
+    ...
+  def ggml_diag_mask_inf(ctx: ffi.CData, a: ffi.CData, n_past: int) -> ffi.CData:
+    """
+    set elements above the diagonal to -INF
+
+        GGML_API struct ggml_tensor * ggml_diag_mask_inf(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a,
+                int                   n_past);
+    """
+    ...
+  def ggml_diag_mask_inf_inplace(ctx: ffi.CData, a: ffi.CData, n_past: int) -> ffi.CData:
+    """
+    in-place, returns view(a)
+
+        GGML_API struct ggml_tensor * ggml_diag_mask_inf_inplace(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a,
+                int                   n_past);
+    """
+    ...
+  def ggml_diag_mask_zero(ctx: ffi.CData, a: ffi.CData, n_past: int) -> ffi.CData:
+    """
+    set elements above the diagonal to 0
+
+        GGML_API struct ggml_tensor * ggml_diag_mask_zero(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a,
+                int                   n_past);
+    """
+    ...
+  def ggml_diag_mask_zero_inplace(ctx: ffi.CData, a: ffi.CData, n_past: int) -> ffi.CData:
+    """
+    in-place, returns view(a)
+
+        GGML_API struct ggml_tensor * ggml_diag_mask_zero_inplace(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a,
+                int                   n_past);
+    """
+    ...
+  def ggml_div(ctx: ffi.CData, a: ffi.CData, b: ffi.CData) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_div(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a,
+                struct ggml_tensor  * b);
+    """
+    ...
+  def ggml_div_inplace(ctx: ffi.CData, a: ffi.CData, b: ffi.CData) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_div_inplace(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a,
+                struct ggml_tensor  * b);
+    """
+    ...
+  def ggml_dup(ctx: ffi.CData, a: ffi.CData) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_dup(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a);
+    """
+    ...
+  def ggml_dup_inplace(ctx: ffi.CData, a: ffi.CData) -> ffi.CData:
+    """
+    in-place, returns view(a)
+
+        GGML_API struct ggml_tensor * ggml_dup_inplace(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a);
+    """
+    ...
+  def ggml_dup_tensor(ctx: ffi.CData, src: ffi.CData) -> ffi.CData:
+    """    GGML_API struct ggml_tensor * ggml_dup_tensor (struct ggml_context * ctx, const struct ggml_tensor * src);"""
+    ...
+  def ggml_element_size(tensor: ffi.CData) -> int:
+    """    GGML_API size_t  ggml_element_size(const struct ggml_tensor * tensor);"""
+    ...
+  def ggml_elu(ctx: ffi.CData, a: ffi.CData) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_elu(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a);
+    """
+    ...
+  def ggml_elu_inplace(ctx: ffi.CData, a: ffi.CData) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_elu_inplace(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a);
+    """
+    ...
+  def ggml_flash_attn(ctx: ffi.CData, q: ffi.CData, k: ffi.CData, v: ffi.CData, masked: bool) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_flash_attn(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * q,
+                struct ggml_tensor  * k,
+                struct ggml_tensor  * v,
+                bool                  masked);
+    """
+    ...
+  def ggml_flash_attn_back(ctx: ffi.CData, q: ffi.CData, k: ffi.CData, v: ffi.CData, d: ffi.CData, masked: bool) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_flash_attn_back(
+               struct ggml_context * ctx,
+               struct ggml_tensor  * q,
+               struct ggml_tensor  * k,
+               struct ggml_tensor  * v,
+               struct ggml_tensor  * d,
+               bool                  masked);
+    """
+    ...
+  def ggml_flash_ff(ctx: ffi.CData, a: ffi.CData, b0: ffi.CData, b1: ffi.CData, c0: ffi.CData, c1: ffi.CData) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_flash_ff(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a,
+                struct ggml_tensor  * b0,
+                struct ggml_tensor  * b1,
+                struct ggml_tensor  * c0,
+                struct ggml_tensor  * c1);
+    """
+    ...
+  def ggml_format_name(tensor: ffi.CData, fmt: ffi.CData, *args2) -> ffi.CData:
+    """    GGML_API struct ggml_tensor * ggml_format_name(      struct ggml_tensor * tensor, const char * fmt, ...);"""
+    ...
+  def ggml_fp16_to_fp32(x: np.float16) -> float:
+    """
+    convert FP16 <-> FP32
+
+        GGML_API float       ggml_fp16_to_fp32(ggml_fp16_t x);
+    """
+    ...
+  def ggml_fp16_to_fp32_row(x: ffi.CData, y: ffi.CData, n: int) -> None:
+    """    GGML_API void ggml_fp16_to_fp32_row(const ggml_fp16_t * x, float * y, int n);"""
+    ...
+  def ggml_fp32_to_fp16(x: float) -> np.float16:
+    """    GGML_API ggml_fp16_t ggml_fp32_to_fp16(float x);"""
+    ...
+  def ggml_fp32_to_fp16_row(x: ffi.CData, y: ffi.CData, n: int) -> None:
+    """    GGML_API void ggml_fp32_to_fp16_row(const float * x, ggml_fp16_t * y, int n);"""
+    ...
+  def ggml_free(ctx: ffi.CData) -> None:
+    """    GGML_API void                  ggml_free(struct ggml_context * ctx);"""
+    ...
+  def ggml_ftype_to_ggml_type(ftype: int) -> int:
+    """
+    TODO: temporary until model loading of ggml examples is refactored
+
+        GGML_API enum ggml_type ggml_ftype_to_ggml_type(enum ggml_ftype ftype);
+    """
+    ...
+  def ggml_gelu(ctx: ffi.CData, a: ffi.CData) -> ffi.CData:
+    """
+    TODO: double-check this computation is correct
+
+        GGML_API struct ggml_tensor * ggml_gelu(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a);
+    """
+    ...
+  def ggml_gelu_inplace(ctx: ffi.CData, a: ffi.CData) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_gelu_inplace(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a);
+    """
+    ...
+  def ggml_gelu_quick(ctx: ffi.CData, a: ffi.CData) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_gelu_quick(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a);
+    """
+    ...
+  def ggml_gelu_quick_inplace(ctx: ffi.CData, a: ffi.CData) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_gelu_quick_inplace(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a);
+    """
+    ...
+  def ggml_get_data(tensor: ffi.CData) -> ffi.CData:
+    """    GGML_API void *  ggml_get_data    (const struct ggml_tensor * tensor);"""
+    ...
+  def ggml_get_data_f32(tensor: ffi.CData) -> ffi.CData:
+    """    GGML_API float * ggml_get_data_f32(const struct ggml_tensor * tensor);"""
+    ...
+  def ggml_get_f32_1d(tensor: ffi.CData, i: int) -> float:
+    """    GGML_API float   ggml_get_f32_1d(const struct ggml_tensor * tensor, int i);"""
+    ...
+  def ggml_get_i32_1d(tensor: ffi.CData, i: int) -> int:
+    """    GGML_API int32_t ggml_get_i32_1d(const struct ggml_tensor * tensor, int i);"""
+    ...
+  def ggml_get_max_tensor_size(ctx: ffi.CData) -> int:
+    """    GGML_API size_t  ggml_get_max_tensor_size(const struct ggml_context * ctx);"""
+    ...
+  def ggml_get_mem_buffer(ctx: ffi.CData) -> ffi.CData:
+    """    GGML_API void *  ggml_get_mem_buffer     (const struct ggml_context * ctx);"""
+    ...
+  def ggml_get_mem_size(ctx: ffi.CData) -> int:
+    """    GGML_API size_t  ggml_get_mem_size       (const struct ggml_context * ctx);"""
+    ...
+  def ggml_get_name(tensor: ffi.CData) -> ffi.CData:
+    """    GGML_API const char *         ggml_get_name   (const struct ggml_tensor * tensor);"""
+    ...
+  def ggml_get_no_alloc(ctx: ffi.CData) -> bool:
+    """    GGML_API bool    ggml_get_no_alloc(struct ggml_context * ctx);"""
+    ...
+  def ggml_get_rows(ctx: ffi.CData, a: ffi.CData, b: ffi.CData) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_get_rows(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a,
+                struct ggml_tensor  * b);
+    """
+    ...
+  def ggml_get_rows_back(ctx: ffi.CData, a: ffi.CData, b: ffi.CData, c: ffi.CData) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_get_rows_back(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a,
+                struct ggml_tensor  * b,
+                struct ggml_tensor  * c);
+    """
+    ...
+  def ggml_get_tensor(ctx: ffi.CData, name: ffi.CData) -> ffi.CData:
+    """    GGML_API struct ggml_tensor * ggml_get_tensor(struct ggml_context * ctx, const char * name);"""
+    ...
+  def ggml_get_unary_op(tensor: ffi.CData) -> int:
+    """    GGML_API enum ggml_unary_op ggml_get_unary_op(const struct ggml_tensor * tensor);"""
+    ...
+  def ggml_graph_compute(cgraph: ffi.CData, cplan: ffi.CData) -> int:
+    """    GGML_API               int ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cplan * cplan);"""
+    ...
+  def ggml_graph_compute_with_ctx(ctx: ffi.CData, cgraph: ffi.CData, n_threads: int) -> None:
+    """
+    same as ggml_graph_compute() but the work data is allocated as a part of the context
+    note: the drawback of this API is that you must have ensured that the context has enough memory for the work data
+
+        GGML_API void ggml_graph_compute_with_ctx(struct ggml_context * ctx, struct ggml_cgraph * cgraph, int n_threads);
+    """
+    ...
+  def ggml_graph_dump_dot(gb: ffi.CData, gf: ffi.CData, filename: ffi.CData) -> None:
+    """
+    dump the graph into a file using the dot format
+
+        GGML_API void ggml_graph_dump_dot(const struct ggml_cgraph * gb, const struct ggml_cgraph * gf, const char * filename);
+    """
+    ...
+  def ggml_graph_export(cgraph: ffi.CData, fname: ffi.CData) -> None:
+    """    GGML_API void               ggml_graph_export(const struct ggml_cgraph * cgraph, const char * fname);"""
+    ...
+  def ggml_graph_get_tensor(cgraph: ffi.CData, name: ffi.CData) -> ffi.CData:
+    """    GGML_API struct ggml_tensor * ggml_graph_get_tensor(struct ggml_cgraph * cgraph, const char * name);"""
+    ...
+  def ggml_graph_import(fname: ffi.CData, ctx_data: ffi.CData, ctx_eval: ffi.CData) -> ffi.CData:
+    """    GGML_API struct ggml_cgraph ggml_graph_import(const char * fname, struct ggml_context ** ctx_data, struct ggml_context ** ctx_eval);"""
+    ...
+  def ggml_graph_overhead() -> int:
+    """    GGML_API size_t ggml_graph_overhead(void);"""
+    ...
+  def ggml_graph_plan(cgraph: ffi.CData, n_threads: int) -> ffi.CData:
+    """
+    ggml_graph_plan() has to be called before ggml_graph_compute()
+    when plan.work_size > 0, caller must allocate memory for plan.work_data
+
+        GGML_API struct ggml_cplan ggml_graph_plan   (struct ggml_cgraph * cgraph, int n_threads /*= GGML_DEFAULT_N_THREADS*/);
+    """
+    ...
+  def ggml_graph_print(cgraph: ffi.CData) -> None:
+    """
+    print info and performance information for the graph
+
+        GGML_API void ggml_graph_print(const struct ggml_cgraph * cgraph);
+    """
+    ...
+  def ggml_graph_reset(cgraph: ffi.CData) -> None:
+    """    GGML_API              void ggml_graph_reset  (struct ggml_cgraph * cgraph);"""
+    ...
+  def ggml_init(params: ffi.CData) -> ffi.CData:
+    """    GGML_API struct ggml_context * ggml_init(struct ggml_init_params params);"""
+    ...
+  def ggml_init_cublas() -> None:
+    """GGML_API void   ggml_init_cublas(void);"""
+    ...
+  def ggml_internal_get_type_traits(type: int) -> ffi.CData:
+    """    ggml_type_traits_t ggml_internal_get_type_traits(enum ggml_type type);"""
+    ...
+  def ggml_is_contiguous(tensor: ffi.CData) -> bool:
+    """    GGML_API bool ggml_is_contiguous(const struct ggml_tensor * tensor);"""
+    ...
+  def ggml_is_numa() -> bool:
+    """    GGML_API bool    ggml_is_numa(void); // true if init detected that system has >1 NUMA node"""
+    ...
+  def ggml_is_permuted(tensor: ffi.CData) -> bool:
+    """    GGML_API bool ggml_is_permuted  (const struct ggml_tensor * tensor);"""
+    ...
+  def ggml_is_quantized(type: int) -> bool:
+    """    GGML_API bool    ggml_is_quantized(enum ggml_type type);"""
+    ...
+  def ggml_is_transposed(tensor: ffi.CData) -> bool:
+    """    GGML_API bool ggml_is_transposed(const struct ggml_tensor * tensor);"""
+    ...
+  def ggml_log(ctx: ffi.CData, a: ffi.CData) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_log(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a);
+    """
+    ...
+  def ggml_log_inplace(ctx: ffi.CData, a: ffi.CData) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_log_inplace(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a);
+    """
+    ...
+  def ggml_map_binary_f32(ctx: ffi.CData, a: ffi.CData, b: ffi.CData, fun: ffi.CData) -> ffi.CData:
+    """
+        GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_binary_f32(
+                struct ggml_context         * ctx,
+                struct ggml_tensor          * a,
+                struct ggml_tensor          * b,
+                       ggml_binary_op_f32_t   fun),
+            "use ggml_map_custom2 instead");
+    """
+    ...
+  def ggml_map_binary_inplace_f32(ctx: ffi.CData, a: ffi.CData, b: ffi.CData, fun: ffi.CData) -> ffi.CData:
+    """
+        GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_binary_inplace_f32(
+                struct ggml_context         * ctx,
+                struct ggml_tensor          * a,
+                struct ggml_tensor          * b,
+                       ggml_binary_op_f32_t   fun),
+            "use ggml_map_custom2_inplace instead");
+    """
+    ...
+  def ggml_map_custom1(ctx: ffi.CData, a: ffi.CData, fun: ffi.CData, n_tasks: int, userdata: ffi.CData) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_map_custom1(
+                struct ggml_context   * ctx,
+                struct ggml_tensor    * a,
+                ggml_custom1_op_t       fun,
+                int                     n_tasks,
+                void                  * userdata);
+    """
+    ...
+  def ggml_map_custom1_f32(ctx: ffi.CData, a: ffi.CData, fun: ffi.CData) -> ffi.CData:
+    """
+        GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom1_f32(
+                struct ggml_context          * ctx,
+                struct ggml_tensor           * a,
+                       ggml_custom1_op_f32_t   fun),
+            "use ggml_map_custom1 instead");
+    """
+    ...
+  def ggml_map_custom1_inplace(ctx: ffi.CData, a: ffi.CData, fun: ffi.CData, n_tasks: int, userdata: ffi.CData) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_map_custom1_inplace(
+                struct ggml_context   * ctx,
+                struct ggml_tensor    * a,
+                ggml_custom1_op_t       fun,
+                int                     n_tasks,
+                void                  * userdata);
+    """
+    ...
+  def ggml_map_custom1_inplace_f32(ctx: ffi.CData, a: ffi.CData, fun: ffi.CData) -> ffi.CData:
+    """
+        GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom1_inplace_f32(
+                struct ggml_context          * ctx,
+                struct ggml_tensor           * a,
+                       ggml_custom1_op_f32_t   fun),
+            "use ggml_map_custom1_inplace instead");
+    """
+    ...
+  def ggml_map_custom2(ctx: ffi.CData, a: ffi.CData, b: ffi.CData, fun: ffi.CData, n_tasks: int, userdata: ffi.CData) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_map_custom2(
+                struct ggml_context   * ctx,
+                struct ggml_tensor    * a,
+                struct ggml_tensor    * b,
+                ggml_custom2_op_t       fun,
+                int                     n_tasks,
+                void                  * userdata);
+    """
+    ...
+  def ggml_map_custom2_f32(ctx: ffi.CData, a: ffi.CData, b: ffi.CData, fun: ffi.CData) -> ffi.CData:
+    """
+        GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom2_f32(
+                struct ggml_context          * ctx,
+                struct ggml_tensor           * a,
+                struct ggml_tensor           * b,
+                       ggml_custom2_op_f32_t   fun),
+            "use ggml_map_custom2 instead");
+    """
+    ...
+  def ggml_map_custom2_inplace(ctx: ffi.CData, a: ffi.CData, b: ffi.CData, fun: ffi.CData, n_tasks: int, userdata: ffi.CData) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_map_custom2_inplace(
+                struct ggml_context   * ctx,
+                struct ggml_tensor    * a,
+                struct ggml_tensor    * b,
+                ggml_custom2_op_t       fun,
+                int                     n_tasks,
+                void                  * userdata);
+    """
+    ...
+  def ggml_map_custom2_inplace_f32(ctx: ffi.CData, a: ffi.CData, b: ffi.CData, fun: ffi.CData) -> ffi.CData:
+    """
+        GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom2_inplace_f32(
+                struct ggml_context          * ctx,
+                struct ggml_tensor           * a,
+                struct ggml_tensor           * b,
+                       ggml_custom2_op_f32_t   fun),
+            "use ggml_map_custom2_inplace instead");
+    """
+    ...
+  def ggml_map_custom3(ctx: ffi.CData, a: ffi.CData, b: ffi.CData, c: ffi.CData, fun: ffi.CData, n_tasks: int, userdata: ffi.CData) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_map_custom3(
+                struct ggml_context   * ctx,
+                struct ggml_tensor    * a,
+                struct ggml_tensor    * b,
+                struct ggml_tensor    * c,
+                ggml_custom3_op_t       fun,
+                int                     n_tasks,
+                void                  * userdata);
+    """
+    ...
+  def ggml_map_custom3_f32(ctx: ffi.CData, a: ffi.CData, b: ffi.CData, c: ffi.CData, fun: ffi.CData) -> ffi.CData:
+    """
+        GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom3_f32(
+                struct ggml_context          * ctx,
+                struct ggml_tensor           * a,
+                struct ggml_tensor           * b,
+                struct ggml_tensor           * c,
+                       ggml_custom3_op_f32_t   fun),
+            "use ggml_map_custom3 instead");
+    """
+    ...
+  def ggml_map_custom3_inplace(ctx: ffi.CData, a: ffi.CData, b: ffi.CData, c: ffi.CData, fun: ffi.CData, n_tasks: int, userdata: ffi.CData) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_map_custom3_inplace(
+                struct ggml_context   * ctx,
+                struct ggml_tensor    * a,
+                struct ggml_tensor    * b,
+                struct ggml_tensor    * c,
+                ggml_custom3_op_t       fun,
+                int                     n_tasks,
+                void                  * userdata);
+    """
+    ...
+  def ggml_map_custom3_inplace_f32(ctx: ffi.CData, a: ffi.CData, b: ffi.CData, c: ffi.CData, fun: ffi.CData) -> ffi.CData:
+    """
+        GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_custom3_inplace_f32(
+                struct ggml_context          * ctx,
+                struct ggml_tensor           * a,
+                struct ggml_tensor           * b,
+                struct ggml_tensor           * c,
+                       ggml_custom3_op_f32_t   fun),
+            "use ggml_map_custom3_inplace instead");
+    """
+    ...
+  def ggml_map_unary_f32(ctx: ffi.CData, a: ffi.CData, fun: ffi.CData) -> ffi.CData:
+    """
+        GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_unary_f32(
+                struct ggml_context        * ctx,
+                struct ggml_tensor         * a,
+                       ggml_unary_op_f32_t   fun),
+            "use ggml_map_custom1 instead");
+    """
+    ...
+  def ggml_map_unary_inplace_f32(ctx: ffi.CData, a: ffi.CData, fun: ffi.CData) -> ffi.CData:
+    """
+        GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_map_unary_inplace_f32(
+                struct ggml_context        * ctx,
+                struct ggml_tensor         * a,
+                       ggml_unary_op_f32_t   fun),
+            "use ggml_map_custom1_inplace instead");
+    """
+    ...
+  def ggml_mean(ctx: ffi.CData, a: ffi.CData) -> ffi.CData:
+    """
+    mean along rows
+
+        GGML_API struct ggml_tensor * ggml_mean(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a);
+    """
+    ...
+  def ggml_metal_add_buffer(ctx: ffi.CData, name: ffi.CData, data: ffi.CData, size: int, max_size: int) -> bool:
+    """
+    creates a mapping between a host memory buffer and a device memory buffer
+    - make sure to map all buffers used in the graph before calling ggml_metal_graph_compute
+    - the mapping is used during computation to determine the arguments of the compute kernels
+    - you don't need to keep the host memory buffer allocated as it is never accessed by Metal
+    - max_size specifies the maximum size of a tensor and is used to create shared views such
+    that it is guaranteed that the tensor will fit in at least one of the views
+    
+
+    bool ggml_metal_add_buffer(
+            struct ggml_metal_context * ctx,
+                           const char * name,
+                                 void * data,
+                               size_t   size,
+                               size_t   max_size);
+    """
+    ...
+  def ggml_metal_free(ctx: ffi.CData) -> None:
+    """void ggml_metal_free(struct ggml_metal_context * ctx);"""
+    ...
+  def ggml_metal_get_concur_list(ctx: ffi.CData) -> ffi.CData:
+    """
+    output the concur_list for ggml_alloc
+
+    int * ggml_metal_get_concur_list(struct ggml_metal_context * ctx);
+    """
+    ...
+  def ggml_metal_get_tensor(ctx: ffi.CData, t: ffi.CData) -> None:
+    """
+    get data from the device into host memory
+
+    void ggml_metal_get_tensor(struct ggml_metal_context * ctx, struct ggml_tensor * t);
+    """
+    ...
+  def ggml_metal_graph_compute(ctx: ffi.CData, gf: ffi.CData) -> None:
+    """
+    same as ggml_graph_compute but uses Metal
+    creates gf->n_threads command buffers in parallel
+
+    void ggml_metal_graph_compute(struct ggml_metal_context * ctx, struct ggml_cgraph * gf);
+    """
+    ...
+  def ggml_metal_graph_find_concurrency(ctx: ffi.CData, gf: ffi.CData, check_mem: bool) -> None:
+    """
+    try to find operations that can be run concurrently in the graph
+    you should run it again if the topology of your graph changes
+
+    void ggml_metal_graph_find_concurrency(struct ggml_metal_context * ctx, struct ggml_cgraph * gf, bool check_mem);
+    """
+    ...
+  def ggml_metal_host_free(data: ffi.CData) -> None:
+    """void   ggml_metal_host_free  (void * data);"""
+    ...
+  def ggml_metal_host_malloc(n: int) -> ffi.CData:
+    """void * ggml_metal_host_malloc(size_t n);"""
+    ...
+  def ggml_metal_if_optimized(ctx: ffi.CData) -> int:
+    """
+    if the graph has been optimized for concurrently dispatch, return length of the concur_list if optimized
+
+    int ggml_metal_if_optimized(struct ggml_metal_context * ctx);
+    """
+    ...
+  def ggml_metal_init(n_cb: int) -> ffi.CData:
+    """
+    number of command buffers to use
+
+    struct ggml_metal_context * ggml_metal_init(int n_cb);
+    """
+    ...
+  def ggml_metal_set_n_cb(ctx: ffi.CData, n_cb: int) -> None:
+    """
+    set the number of command buffers to use
+
+    void ggml_metal_set_n_cb(struct ggml_metal_context * ctx, int n_cb);
+    """
+    ...
+  def ggml_metal_set_tensor(ctx: ffi.CData, t: ffi.CData) -> None:
+    """
+    set data from host memory into the device
+
+    void ggml_metal_set_tensor(struct ggml_metal_context * ctx, struct ggml_tensor * t);
+    """
+    ...
+  def ggml_mpi_backend_free() -> None:
+    """void ggml_mpi_backend_free(void);"""
+    ...
+  def ggml_mpi_backend_init() -> None:
+    """void ggml_mpi_backend_init(void);"""
+    ...
+  def ggml_mpi_eval_init(ctx_mpi: ffi.CData, n_tokens: ffi.CData, n_past: ffi.CData, n_threads: ffi.CData) -> None:
+    """
+    void ggml_mpi_eval_init(
+            struct ggml_mpi_context * ctx_mpi,
+                                int * n_tokens,
+                                int * n_past,
+                                int * n_threads);
+    """
+    ...
+  def ggml_mpi_free(ctx: ffi.CData) -> None:
+    """void ggml_mpi_free(struct ggml_mpi_context * ctx);"""
+    ...
+  def ggml_mpi_graph_compute_post(ctx_mpi: ffi.CData, gf: ffi.CData, n_layers: int) -> None:
+    """
+    void ggml_mpi_graph_compute_post(
+            struct ggml_mpi_context * ctx_mpi,
+                 struct ggml_cgraph * gf,
+                                int   n_layers);
+    """
+    ...
+  def ggml_mpi_graph_compute_pre(ctx_mpi: ffi.CData, gf: ffi.CData, n_layers: int) -> None:
+    """
+    void ggml_mpi_graph_compute_pre(
+            struct ggml_mpi_context * ctx_mpi,
+                 struct ggml_cgraph * gf,
+                                int   n_layers);
+    """
+    ...
+  def ggml_mpi_init() -> ffi.CData:
+    """struct ggml_mpi_context * ggml_mpi_init(void);"""
+    ...
+  def ggml_mpi_rank(ctx: ffi.CData) -> int:
+    """int ggml_mpi_rank(struct ggml_mpi_context * ctx);"""
+    ...
+  def ggml_mul(ctx: ffi.CData, a: ffi.CData, b: ffi.CData) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_mul(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a,
+                struct ggml_tensor  * b);
+    """
+    ...
+  def ggml_mul_inplace(ctx: ffi.CData, a: ffi.CData, b: ffi.CData) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_mul_inplace(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a,
+                struct ggml_tensor  * b);
+    """
+    ...
+  def ggml_mul_mat(ctx: ffi.CData, a: ffi.CData, b: ffi.CData) -> ffi.CData:
+    """
+    A: n columns, m rows
+    B: n columns, p rows  (i.e. we transpose it internally)
+    result is m columns, p rows
+
+        GGML_API struct ggml_tensor * ggml_mul_mat(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a,
+                struct ggml_tensor  * b);
+    """
+    ...
+  def ggml_nbytes(tensor: ffi.CData) -> int:
+    """    GGML_API size_t  ggml_nbytes      (const struct ggml_tensor * tensor);"""
+    ...
+  def ggml_nbytes_pad(tensor: ffi.CData) -> int:
+    """    GGML_API size_t  ggml_nbytes_pad  (const struct ggml_tensor * tensor); // same as ggml_nbytes() but padded to GGML_MEM_ALIGN"""
+    ...
+  def ggml_nbytes_split(tensor: ffi.CData, nrows_split: int) -> int:
+    """    GGML_API size_t  ggml_nbytes_split(const struct ggml_tensor * tensor, int nrows_split);"""
+    ...
+  def ggml_neg(ctx: ffi.CData, a: ffi.CData) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_neg(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a);
+    """
+    ...
+  def ggml_neg_inplace(ctx: ffi.CData, a: ffi.CData) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_neg_inplace(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a);
+    """
+    ...
+  def ggml_nelements(tensor: ffi.CData) -> int:
+    """    GGML_API int64_t ggml_nelements   (const struct ggml_tensor * tensor);"""
+    ...
+  def ggml_new_f32(ctx: ffi.CData, value: float) -> ffi.CData:
+    """    GGML_API struct ggml_tensor * ggml_new_f32(struct ggml_context * ctx, float value);"""
+    ...
+  def ggml_new_graph(ctx: ffi.CData) -> ffi.CData:
+    """
+    graph allocation in a context
+
+        GGML_API struct ggml_cgraph * ggml_new_graph        (struct ggml_context * ctx);
+    """
+    ...
+  def ggml_new_i32(ctx: ffi.CData, value: int) -> ffi.CData:
+    """    GGML_API struct ggml_tensor * ggml_new_i32(struct ggml_context * ctx, int32_t value);"""
+    ...
+  def ggml_new_tensor(ctx: ffi.CData, type: int, n_dims: int, ne: ffi.CData) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_new_tensor(
+                struct ggml_context * ctx,
+                enum   ggml_type type,
+                int    n_dims,
+                const int64_t *ne);
+    """
+    ...
+  def ggml_new_tensor_1d(ctx: ffi.CData, type: int, ne0: int) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_new_tensor_1d(
+                struct ggml_context * ctx,
+                enum   ggml_type type,
+                int64_t ne0);
+    """
+    ...
+  def ggml_new_tensor_2d(ctx: ffi.CData, type: int, ne0: int, ne1: int) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_new_tensor_2d(
+                struct ggml_context * ctx,
+                enum   ggml_type type,
+                int64_t ne0,
+                int64_t ne1);
+    """
+    ...
+  def ggml_new_tensor_3d(ctx: ffi.CData, type: int, ne0: int, ne1: int, ne2: int) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_new_tensor_3d(
+                struct ggml_context * ctx,
+                enum   ggml_type type,
+                int64_t ne0,
+                int64_t ne1,
+                int64_t ne2);
+    """
+    ...
+  def ggml_new_tensor_4d(ctx: ffi.CData, type: int, ne0: int, ne1: int, ne2: int, ne3: int) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_new_tensor_4d(
+                struct ggml_context * ctx,
+                enum   ggml_type type,
+                int64_t ne0,
+                int64_t ne1,
+                int64_t ne2,
+                int64_t ne3);
+    """
+    ...
+  def ggml_norm(ctx: ffi.CData, a: ffi.CData) -> ffi.CData:
+    """
+    normalize along rows
+    TODO: eps is hardcoded to 1e-5 for now
+
+        GGML_API struct ggml_tensor * ggml_norm(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a);
+    """
+    ...
+  def ggml_norm_inplace(ctx: ffi.CData, a: ffi.CData) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_norm_inplace(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a);
+    """
+    ...
+  def ggml_nrows(tensor: ffi.CData) -> int:
+    """    GGML_API int64_t ggml_nrows       (const struct ggml_tensor * tensor);"""
+    ...
+  def ggml_numa_init() -> None:
+    """    GGML_API void    ggml_numa_init(void); // call once for better performance on NUMA systems"""
+    ...
+  def ggml_op_name(op: int) -> ffi.CData:
+    """    GGML_API const char * ggml_op_name  (enum ggml_op   op);"""
+    ...
+  def ggml_op_symbol(op: int) -> ffi.CData:
+    """    GGML_API const char * ggml_op_symbol(enum ggml_op   op);"""
+    ...
+  def ggml_opt(ctx: ffi.CData, params: ffi.CData, f: ffi.CData) -> int:
+    """
+    optimize the function defined by the tensor f
+
+        GGML_API enum ggml_opt_result ggml_opt(
+                struct ggml_context * ctx,
+                struct ggml_opt_params params,
+                struct ggml_tensor * f);
+    """
+    ...
+  def ggml_opt_default_params(type: int) -> ffi.CData:
+    """    GGML_API struct ggml_opt_params ggml_opt_default_params(enum ggml_opt_type type);"""
+    ...
+  def ggml_opt_init(ctx: ffi.CData, opt: ffi.CData, params: ffi.CData, nx: int) -> None:
+    """
+    initialize optimizer context
+
+        GGML_API void ggml_opt_init(
+                struct ggml_context * ctx,
+                struct ggml_opt_context * opt,
+                struct ggml_opt_params params,
+                int64_t nx);
+    """
+    ...
+  def ggml_opt_resume(ctx: ffi.CData, opt: ffi.CData, f: ffi.CData) -> int:
+    """
+    continue optimizing the function defined by the tensor f
+
+        GGML_API enum ggml_opt_result ggml_opt_resume(
+                struct ggml_context * ctx,
+                struct ggml_opt_context * opt,
+                struct ggml_tensor * f);
+    """
+    ...
+  def ggml_opt_resume_g(ctx: ffi.CData, opt: ffi.CData, f: ffi.CData, gf: ffi.CData, gb: ffi.CData) -> int:
+    """
+    continue optimizing the function defined by the tensor f
+
+        GGML_API enum ggml_opt_result ggml_opt_resume_g(
+                struct ggml_context * ctx,
+                struct ggml_opt_context * opt,
+                struct ggml_tensor * f,
+                struct ggml_cgraph * gf,
+                struct ggml_cgraph * gb);
+    """
+    ...
+  def ggml_out_prod(ctx: ffi.CData, a: ffi.CData, b: ffi.CData) -> ffi.CData:
+    """
+    A: m columns, n rows,
+    B: p columns, n rows,
+    result is m columns, p rows
+
+        GGML_API struct ggml_tensor * ggml_out_prod(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a,
+                struct ggml_tensor  * b);
+    """
+    ...
+  def ggml_permute(ctx: ffi.CData, a: ffi.CData, axis0: int, axis1: int, axis2: int, axis3: int) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_permute(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a,
+                int                   axis0,
+                int                   axis1,
+                int                   axis2,
+                int                   axis3);
+    """
+    ...
+  def ggml_pool_1d(ctx: ffi.CData, a: ffi.CData, op: int, k0: int, s0: int, p0: int) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_pool_1d(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a,
+                enum ggml_op_pool     op,
+                int                   k0, // kernel size
+                int                   s0, // stride
+                int                   p0); // padding
+    """
+    ...
+  def ggml_pool_2d(ctx: ffi.CData, a: ffi.CData, op: int, k0: int, k1: int, s0: int, s1: int, p0: int, p1: int) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_pool_2d(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a,
+                enum ggml_op_pool     op,
+                int                   k0,
+                int                   k1,
+                int                   s0,
+                int                   s1,
+                int                   p0,
+                int                   p1);
+    """
+    ...
+  def ggml_print_object(obj: ffi.CData) -> None:
+    """    GGML_API void    ggml_print_object (const struct ggml_object * obj);"""
+    ...
+  def ggml_print_objects(ctx: ffi.CData) -> None:
+    """    GGML_API void    ggml_print_objects(const struct ggml_context * ctx);"""
+    ...
+  def ggml_quantize_chunk(type: int, src: ffi.CData, dst: ffi.CData, start: int, n: int, hist: ffi.CData) -> int:
+    """    GGML_API size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst, int start, int n, int64_t * hist);"""
+    ...
+  def ggml_quantize_q2_K(src: ffi.CData, dst: ffi.CData, n: int, k: int, hist: ffi.CData) -> int:
+    """
+    Quantization with histogram collection
+
+    size_t ggml_quantize_q2_K(const float * src, void * dst, int n, int k, int64_t * hist);
+    """
+    ...
+  def ggml_quantize_q3_K(src: ffi.CData, dst: ffi.CData, n: int, k: int, hist: ffi.CData) -> int:
+    """size_t ggml_quantize_q3_K(const float * src, void * dst, int n, int k, int64_t * hist);"""
+    ...
+  def ggml_quantize_q4_0(src: ffi.CData, dst: ffi.CData, n: int, k: int, hist: ffi.CData) -> int:
+    """    GGML_API size_t ggml_quantize_q4_0(const float * src, void * dst, int n, int k, int64_t * hist);"""
+    ...
+  def ggml_quantize_q4_1(src: ffi.CData, dst: ffi.CData, n: int, k: int, hist: ffi.CData) -> int:
+    """    GGML_API size_t ggml_quantize_q4_1(const float * src, void * dst, int n, int k, int64_t * hist);"""
+    ...
+  def ggml_quantize_q4_K(src: ffi.CData, dst: ffi.CData, n: int, k: int, hist: ffi.CData) -> int:
+    """size_t ggml_quantize_q4_K(const float * src, void * dst, int n, int k, int64_t * hist);"""
+    ...
+  def ggml_quantize_q5_0(src: ffi.CData, dst: ffi.CData, n: int, k: int, hist: ffi.CData) -> int:
+    """    GGML_API size_t ggml_quantize_q5_0(const float * src, void * dst, int n, int k, int64_t * hist);"""
+    ...
+  def ggml_quantize_q5_1(src: ffi.CData, dst: ffi.CData, n: int, k: int, hist: ffi.CData) -> int:
+    """    GGML_API size_t ggml_quantize_q5_1(const float * src, void * dst, int n, int k, int64_t * hist);"""
+    ...
+  def ggml_quantize_q5_K(src: ffi.CData, dst: ffi.CData, n: int, k: int, hist: ffi.CData) -> int:
+    """size_t ggml_quantize_q5_K(const float * src, void * dst, int n, int k, int64_t * hist);"""
+    ...
+  def ggml_quantize_q6_K(src: ffi.CData, dst: ffi.CData, n: int, k: int, hist: ffi.CData) -> int:
+    """size_t ggml_quantize_q6_K(const float * src, void * dst, int n, int k, int64_t * hist);"""
+    ...
+  def ggml_quantize_q8_0(src: ffi.CData, dst: ffi.CData, n: int, k: int, hist: ffi.CData) -> int:
+    """    GGML_API size_t ggml_quantize_q8_0(const float * src, void * dst, int n, int k, int64_t * hist);"""
+    ...
+  def ggml_relu(ctx: ffi.CData, a: ffi.CData) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_relu(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a);
+    """
+    ...
+  def ggml_relu_inplace(ctx: ffi.CData, a: ffi.CData) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_relu_inplace(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a);
+    """
+    ...
+  def ggml_repeat(ctx: ffi.CData, a: ffi.CData, b: ffi.CData) -> ffi.CData:
+    """
+    if a is the same shape as b, and a is not parameter, return a
+    otherwise, return a new tensor: repeat(a) to fit in b
+
+        GGML_API struct ggml_tensor * ggml_repeat(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a,
+                struct ggml_tensor  * b);
+    """
+    ...
+  def ggml_repeat_back(ctx: ffi.CData, a: ffi.CData, b: ffi.CData) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_repeat_back(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a,
+                struct ggml_tensor  * b);
+    """
+    ...
+  def ggml_reshape(ctx: ffi.CData, a: ffi.CData, b: ffi.CData) -> ffi.CData:
+    """
+    return view(a), b specifies the new shape
+    TODO: when we start computing gradient, make a copy instead of view
+
+        GGML_API struct ggml_tensor * ggml_reshape(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a,
+                struct ggml_tensor  * b);
+    """
+    ...
+  def ggml_reshape_1d(ctx: ffi.CData, a: ffi.CData, ne0: int) -> ffi.CData:
+    """
+    return view(a)
+    TODO: when we start computing gradient, make a copy instead of view
+
+        GGML_API struct ggml_tensor * ggml_reshape_1d(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a,
+                int64_t               ne0);
+    """
+    ...
+  def ggml_reshape_2d(ctx: ffi.CData, a: ffi.CData, ne0: int, ne1: int) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_reshape_2d(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a,
+                int64_t               ne0,
+                int64_t               ne1);
+    """
+    ...
+  def ggml_reshape_3d(ctx: ffi.CData, a: ffi.CData, ne0: int, ne1: int, ne2: int) -> ffi.CData:
+    """
+    return view(a)
+    TODO: when we start computing gradient, make a copy instead of view
+
+        GGML_API struct ggml_tensor * ggml_reshape_3d(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a,
+                int64_t               ne0,
+                int64_t               ne1,
+                int64_t               ne2);
+    """
+    ...
+  def ggml_reshape_4d(ctx: ffi.CData, a: ffi.CData, ne0: int, ne1: int, ne2: int, ne3: int) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_reshape_4d(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a,
+                int64_t               ne0,
+                int64_t               ne1,
+                int64_t               ne2,
+                int64_t               ne3);
+    """
+    ...
+  def ggml_rms_norm(ctx: ffi.CData, a: ffi.CData, eps: float) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_rms_norm(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a,
+                float                 eps);
+    """
+    ...
+  def ggml_rms_norm_back(ctx: ffi.CData, a: ffi.CData, b: ffi.CData) -> ffi.CData:
+    """
+    a - x
+    b - dy
+    TODO: update with configurable eps
+
+        GGML_API struct ggml_tensor * ggml_rms_norm_back(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a,
+                struct ggml_tensor  * b);
+    """
+    ...
+  def ggml_rms_norm_inplace(ctx: ffi.CData, a: ffi.CData, eps: float) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_rms_norm_inplace(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a,
+                float                 eps);
+    """
+    ...
+  def ggml_rope(ctx: ffi.CData, a: ffi.CData, n_past: int, n_dims: int, mode: int, n_ctx: int) -> ffi.CData:
+    """
+    rotary position embedding
+    if mode & 1 == 1, skip n_past elements
+    if mode & 2 == 1, GPT-NeoX style
+    if mode & 4 == 1, ChatGLM style
+    TODO: avoid creating a new tensor every time
+
+        GGML_API struct ggml_tensor * ggml_rope(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a,
+                int                   n_past,
+                int                   n_dims,
+                int                   mode,
+                int                   n_ctx);
+    """
+    ...
+  def ggml_rope_back(ctx: ffi.CData, a: ffi.CData, n_past: int, n_dims: int, mode: int, n_ctx: int) -> ffi.CData:
+    """
+    rotary position embedding backward, i.e compute dx from dy
+    a - dy
+
+        GGML_API struct ggml_tensor * ggml_rope_back(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a,
+                int                   n_past,
+                int                   n_dims,
+                int                   mode,
+                int                   n_ctx);
+    """
+    ...
+  def ggml_rope_custom(ctx: ffi.CData, a: ffi.CData, n_past: int, n_dims: int, mode: int, n_ctx: int, freq_base: float, freq_scale: float) -> ffi.CData:
+    """
+    custom RoPE
+
+        GGML_API struct ggml_tensor * ggml_rope_custom(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a,
+                int                   n_past,
+                int                   n_dims,
+                int                   mode,
+                int                   n_ctx,
+                float                 freq_base,
+                float                 freq_scale);
+    """
+    ...
+  def ggml_rope_custom_inplace(ctx: ffi.CData, a: ffi.CData, n_past: int, n_dims: int, mode: int, n_ctx: int, freq_base: float, freq_scale: float) -> ffi.CData:
+    """
+    in-place, returns view(a)
+
+        GGML_API struct ggml_tensor * ggml_rope_custom_inplace(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a,
+                int                   n_past,
+                int                   n_dims,
+                int                   mode,
+                int                   n_ctx,
+                float                 freq_base,
+                float                 freq_scale);
+    """
+    ...
+  def ggml_rope_inplace(ctx: ffi.CData, a: ffi.CData, n_past: int, n_dims: int, mode: int, n_ctx: int) -> ffi.CData:
+    """
+    in-place, returns view(a)
+
+        GGML_API struct ggml_tensor * ggml_rope_inplace(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a,
+                int                   n_past,
+                int                   n_dims,
+                int                   mode,
+                int                   n_ctx);
+    """
+    ...
+  def ggml_scale(ctx: ffi.CData, a: ffi.CData, b: ffi.CData) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_scale(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a,
+                struct ggml_tensor  * b);
+    """
+    ...
+  def ggml_scale_inplace(ctx: ffi.CData, a: ffi.CData, b: ffi.CData) -> ffi.CData:
+    """
+    in-place, returns view(a)
+
+        GGML_API struct ggml_tensor * ggml_scale_inplace(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a,
+                struct ggml_tensor  * b);
+    """
+    ...
+  def ggml_set(ctx: ffi.CData, a: ffi.CData, b: ffi.CData, nb1: int, nb2: int, nb3: int, offset: int) -> ffi.CData:
+    """
+    b -> view(a,offset,nb1,nb2,3), return modified a
+
+        GGML_API struct ggml_tensor * ggml_set(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a,
+                struct ggml_tensor  * b,
+                size_t                nb1,
+                size_t                nb2,
+                size_t                nb3,
+                size_t                offset);
+    """
+    ...
+  def ggml_set_1d(ctx: ffi.CData, a: ffi.CData, b: ffi.CData, offset: int) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_set_1d(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a,
+                struct ggml_tensor  * b,
+                size_t                offset);
+    """
+    ...
+  def ggml_set_1d_inplace(ctx: ffi.CData, a: ffi.CData, b: ffi.CData, offset: int) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_set_1d_inplace(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a,
+                struct ggml_tensor  * b,
+                size_t                offset);
+    """
+    ...
+  def ggml_set_2d(ctx: ffi.CData, a: ffi.CData, b: ffi.CData, nb1: int, offset: int) -> ffi.CData:
+    """
+    b -> view(a,offset,nb1,nb2,3), return modified a
+
+        GGML_API struct ggml_tensor * ggml_set_2d(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a,
+                struct ggml_tensor  * b,
+                size_t                nb1,
+                size_t                offset);
+    """
+    ...
+  def ggml_set_2d_inplace(ctx: ffi.CData, a: ffi.CData, b: ffi.CData, nb1: int, offset: int) -> ffi.CData:
+    """
+    b -> view(a,offset,nb1,nb2,3), return view(a)
+
+        GGML_API struct ggml_tensor * ggml_set_2d_inplace(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a,
+                struct ggml_tensor  * b,
+                size_t                nb1,
+                size_t                offset);
+    """
+    ...
+  def ggml_set_f32(tensor: ffi.CData, value: float) -> ffi.CData:
+    """    GGML_API struct ggml_tensor * ggml_set_f32 (struct ggml_tensor * tensor, float value);"""
+    ...
+  def ggml_set_f32_1d(tensor: ffi.CData, i: int, value: float) -> None:
+    """    GGML_API void    ggml_set_f32_1d(const struct ggml_tensor * tensor, int i, float value);"""
+    ...
+  def ggml_set_i32(tensor: ffi.CData, value: int) -> ffi.CData:
+    """    GGML_API struct ggml_tensor * ggml_set_i32 (struct ggml_tensor * tensor, int32_t value);"""
+    ...
+  def ggml_set_i32_1d(tensor: ffi.CData, i: int, value: int) -> None:
+    """    GGML_API void    ggml_set_i32_1d(const struct ggml_tensor * tensor, int i, int32_t value);"""
+    ...
+  def ggml_set_inplace(ctx: ffi.CData, a: ffi.CData, b: ffi.CData, nb1: int, nb2: int, nb3: int, offset: int) -> ffi.CData:
+    """
+    b -> view(a,offset,nb1,nb2,3), return view(a)
+
+        GGML_API struct ggml_tensor * ggml_set_inplace(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a,
+                struct ggml_tensor  * b,
+                size_t                nb1,
+                size_t                nb2,
+                size_t                nb3,
+                size_t                offset);
+    """
+    ...
+  def ggml_set_name(tensor: ffi.CData, name: ffi.CData) -> ffi.CData:
+    """    GGML_API struct ggml_tensor * ggml_set_name   (      struct ggml_tensor * tensor, const char * name);"""
+    ...
+  def ggml_set_no_alloc(ctx: ffi.CData, no_alloc: bool) -> None:
+    """    GGML_API void    ggml_set_no_alloc(struct ggml_context * ctx, bool no_alloc);"""
+    ...
+  def ggml_set_param(ctx: ffi.CData, tensor: ffi.CData) -> None:
+    """
+        GGML_API void ggml_set_param(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * tensor);
+    """
+    ...
+  def ggml_set_scratch(ctx: ffi.CData, scratch: ffi.CData) -> int:
+    """    GGML_API size_t  ggml_set_scratch (struct ggml_context * ctx, struct ggml_scratch scratch);"""
+    ...
+  def ggml_set_zero(tensor: ffi.CData) -> ffi.CData:
+    """    GGML_API struct ggml_tensor * ggml_set_zero(struct ggml_tensor * tensor);"""
+    ...
+  def ggml_sgn(ctx: ffi.CData, a: ffi.CData) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_sgn(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a);
+    """
+    ...
+  def ggml_sgn_inplace(ctx: ffi.CData, a: ffi.CData) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_sgn_inplace(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a);
+    """
+    ...
+  def ggml_silu(ctx: ffi.CData, a: ffi.CData) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_silu(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a);
+    """
+    ...
+  def ggml_silu_back(ctx: ffi.CData, a: ffi.CData, b: ffi.CData) -> ffi.CData:
+    """
+    a - x
+    b - dy
+
+        GGML_API struct ggml_tensor * ggml_silu_back(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a,
+                struct ggml_tensor  * b);
+    """
+    ...
+  def ggml_silu_inplace(ctx: ffi.CData, a: ffi.CData) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_silu_inplace(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a);
+    """
+    ...
+  def ggml_soft_max(ctx: ffi.CData, a: ffi.CData) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_soft_max(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a);
+    """
+    ...
+  def ggml_soft_max_back(ctx: ffi.CData, a: ffi.CData, b: ffi.CData) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_soft_max_back(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a,
+                struct ggml_tensor  * b);
+    """
+    ...
+  def ggml_soft_max_back_inplace(ctx: ffi.CData, a: ffi.CData, b: ffi.CData) -> ffi.CData:
+    """
+    in-place, returns view(a)
+
+        GGML_API struct ggml_tensor * ggml_soft_max_back_inplace(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a,
+                struct ggml_tensor  * b);
+    """
+    ...
+  def ggml_soft_max_inplace(ctx: ffi.CData, a: ffi.CData) -> ffi.CData:
+    """
+    in-place, returns view(a)
+
+        GGML_API struct ggml_tensor * ggml_soft_max_inplace(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a);
+    """
+    ...
+  def ggml_sqr(ctx: ffi.CData, a: ffi.CData) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_sqr(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a);
+    """
+    ...
+  def ggml_sqr_inplace(ctx: ffi.CData, a: ffi.CData) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_sqr_inplace(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a);
+    """
+    ...
+  def ggml_sqrt(ctx: ffi.CData, a: ffi.CData) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_sqrt(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a);
+    """
+    ...
+  def ggml_sqrt_inplace(ctx: ffi.CData, a: ffi.CData) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_sqrt_inplace(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a);
+    """
+    ...
+  def ggml_step(ctx: ffi.CData, a: ffi.CData) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_step(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a);
+    """
+    ...
+  def ggml_step_inplace(ctx: ffi.CData, a: ffi.CData) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_step_inplace(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a);
+    """
+    ...
+  def ggml_sub(ctx: ffi.CData, a: ffi.CData, b: ffi.CData) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_sub(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a,
+                struct ggml_tensor  * b);
+    """
+    ...
+  def ggml_sub_inplace(ctx: ffi.CData, a: ffi.CData, b: ffi.CData) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_sub_inplace(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a,
+                struct ggml_tensor  * b);
+    """
+    ...
+  def ggml_sum(ctx: ffi.CData, a: ffi.CData) -> ffi.CData:
+    """
+    return scalar
+
+        GGML_API struct ggml_tensor * ggml_sum(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a);
+    """
+    ...
+  def ggml_sum_rows(ctx: ffi.CData, a: ffi.CData) -> ffi.CData:
+    """
+    sums along rows, with input shape [a,b,c,d] return shape [1,b,c,d]
+
+        GGML_API struct ggml_tensor * ggml_sum_rows(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a);
+    """
+    ...
+  def ggml_tanh(ctx: ffi.CData, a: ffi.CData) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_tanh(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a);
+    """
+    ...
+  def ggml_tanh_inplace(ctx: ffi.CData, a: ffi.CData) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_tanh_inplace(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a);
+    """
+    ...
+  def ggml_tensor_overhead() -> int:
+    """
+    use this to compute the memory overhead of a tensor
+
+        GGML_API size_t ggml_tensor_overhead(void);
+    """
+    ...
+  def ggml_time_init() -> None:
+    """    GGML_API void    ggml_time_init(void); // call this once at the beginning of the program"""
+    ...
+  def ggml_time_ms() -> int:
+    """    GGML_API int64_t ggml_time_ms(void);"""
+    ...
+  def ggml_time_us() -> int:
+    """    GGML_API int64_t ggml_time_us(void);"""
+    ...
+  def ggml_transpose(ctx: ffi.CData, a: ffi.CData) -> ffi.CData:
+    """
+    alias for ggml_permute(ctx, a, 1, 0, 2, 3)
+
+        GGML_API struct ggml_tensor * ggml_transpose(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a);
+    """
+    ...
+  def ggml_type_name(type: int) -> ffi.CData:
+    """    GGML_API const char * ggml_type_name(enum ggml_type type);"""
+    ...
+  def ggml_type_size(type: int) -> int:
+    """    GGML_API size_t  ggml_type_size (enum ggml_type type); // size in bytes for all elements in a block"""
+    ...
+  def ggml_type_sizef(type: int) -> float:
+    """    GGML_API float   ggml_type_sizef(enum ggml_type type); // ggml_type_size()/ggml_blck_size() as float"""
+    ...
+  def ggml_unary(ctx: ffi.CData, a: ffi.CData, op: int) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_unary(
+                struct ggml_context * ctx,
+                 struct ggml_tensor * a,
+                 enum ggml_unary_op op);
+    """
+    ...
+  def ggml_unary_inplace(ctx: ffi.CData, a: ffi.CData, op: int) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_unary_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            enum ggml_unary_op op);
+    """
+    ...
+  def ggml_used_mem(ctx: ffi.CData) -> int:
+    """    GGML_API size_t  ggml_used_mem(const struct ggml_context * ctx);"""
+    ...
+  def ggml_vec_dot_q2_K_q8_K(n: int, s: ffi.CData, vx: ffi.CData, vy: ffi.CData) -> None:
+    """
+    Dot product
+
+    void ggml_vec_dot_q2_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+    """
+    ...
+  def ggml_vec_dot_q3_K_q8_K(n: int, s: ffi.CData, vx: ffi.CData, vy: ffi.CData) -> None:
+    """void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);"""
+    ...
+  def ggml_vec_dot_q4_K_q8_K(n: int, s: ffi.CData, vx: ffi.CData, vy: ffi.CData) -> None:
+    """void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);"""
+    ...
+  def ggml_vec_dot_q5_K_q8_K(n: int, s: ffi.CData, vx: ffi.CData, vy: ffi.CData) -> None:
+    """void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);"""
+    ...
+  def ggml_vec_dot_q6_K_q8_K(n: int, s: ffi.CData, vx: ffi.CData, vy: ffi.CData) -> None:
+    """void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);"""
+    ...
+  def ggml_view_1d(ctx: ffi.CData, a: ffi.CData, ne0: int, offset: int) -> ffi.CData:
+    """
+    offset in bytes
+
+        GGML_API struct ggml_tensor * ggml_view_1d(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a,
+                int64_t               ne0,
+                size_t                offset);
+    """
+    ...
+  def ggml_view_2d(ctx: ffi.CData, a: ffi.CData, ne0: int, ne1: int, nb1: int, offset: int) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_view_2d(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a,
+                int64_t               ne0,
+                int64_t               ne1,
+                size_t                nb1, // row stride in bytes
+                size_t                offset);
+    """
+    ...
+  def ggml_view_3d(ctx: ffi.CData, a: ffi.CData, ne0: int, ne1: int, ne2: int, nb1: int, nb2: int, offset: int) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_view_3d(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a,
+                int64_t               ne0,
+                int64_t               ne1,
+                int64_t               ne2,
+                size_t                nb1, // row   stride in bytes
+                size_t                nb2, // slice stride in bytes
+                size_t                offset);
+    """
+    ...
+  def ggml_view_4d(ctx: ffi.CData, a: ffi.CData, ne0: int, ne1: int, ne2: int, ne3: int, nb1: int, nb2: int, nb3: int, offset: int) -> ffi.CData:
+    """
+        GGML_API struct ggml_tensor * ggml_view_4d(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a,
+                int64_t               ne0,
+                int64_t               ne1,
+                int64_t               ne2,
+                int64_t               ne3,
+                size_t                nb1, // row   stride in bytes
+                size_t                nb2, // slice stride in bytes
+                size_t                nb3,
+                size_t                offset);
+    """
+    ...
+  def ggml_view_tensor(ctx: ffi.CData, src: ffi.CData) -> ffi.CData:
+    """    GGML_API struct ggml_tensor * ggml_view_tensor(struct ggml_context * ctx, const struct ggml_tensor * src);"""
+    ...
+  def ggml_win_part(ctx: ffi.CData, a: ffi.CData, w: int) -> ffi.CData:
+    """
+    partition into non-overlapping windows with padding if needed
+    example:
+    a:   768   64   64    1
+    w:    14
+    res: 768   14   14    25
+    used in sam
+
+        GGML_API struct ggml_tensor * ggml_win_part(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a,
+                int                   w);
+    """
+    ...
+  def ggml_win_unpart(ctx: ffi.CData, a: ffi.CData, w0: int, h0: int, w: int) -> ffi.CData:
+    """
+    reverse of ggml_win_part
+    used in sam
+
+        GGML_API struct ggml_tensor * ggml_win_unpart(
+                struct ggml_context * ctx,
+                struct ggml_tensor  * a,
+                int                   w0,
+                int                   h0,
+                int                   w);
+    """
+    ...
+  def gguf_add_tensor(ctx: ffi.CData, tensor: ffi.CData) -> None:
+    """
+    manage tensor info
+
+        GGML_API void gguf_add_tensor(struct gguf_context * ctx, const struct ggml_tensor * tensor);
+    """
+    ...
+  def gguf_find_key(ctx: ffi.CData, key: ffi.CData) -> int:
+    """    GGML_API int          gguf_find_key(struct gguf_context * ctx, const char * key);"""
+    ...
+  def gguf_find_tensor(ctx: ffi.CData, name: ffi.CData) -> int:
+    """    GGML_API int    gguf_find_tensor      (struct gguf_context * ctx, const char * name);"""
+    ...
+  def gguf_free(ctx: ffi.CData) -> None:
+    """    GGML_API void gguf_free(struct gguf_context * ctx);"""
+    ...
+  def gguf_get_alignment(ctx: ffi.CData) -> int:
+    """    GGML_API size_t gguf_get_alignment  (struct gguf_context * ctx);"""
+    ...
+  def gguf_get_arr_data(ctx: ffi.CData, i: int) -> ffi.CData:
+    """    GGML_API const void * gguf_get_arr_data(struct gguf_context * ctx, int i);"""
+    ...
+  def gguf_get_arr_n(ctx: ffi.CData, i: int) -> int:
+    """    GGML_API int          gguf_get_arr_n   (struct gguf_context * ctx, int i);"""
+    ...
+  def gguf_get_arr_str(ctx: ffi.CData, key_id: int, i: int) -> ffi.CData:
+    """    GGML_API const char * gguf_get_arr_str (struct gguf_context * ctx, int key_id, int i);"""
+    ...
+  def gguf_get_arr_type(ctx: ffi.CData, i: int) -> int:
+    """    GGML_API enum gguf_type gguf_get_arr_type(struct gguf_context * ctx, int i);"""
+    ...
+  def gguf_get_data(ctx: ffi.CData) -> ffi.CData:
+    """    GGML_API void * gguf_get_data       (struct gguf_context * ctx);"""
+    ...
+  def gguf_get_data_offset(ctx: ffi.CData) -> int:
+    """    GGML_API size_t gguf_get_data_offset(struct gguf_context * ctx);"""
+    ...
+  def gguf_get_key(ctx: ffi.CData, i: int) -> ffi.CData:
+    """    GGML_API const char * gguf_get_key (struct gguf_context * ctx, int i);"""
+    ...
+  def gguf_get_kv_type(ctx: ffi.CData, i: int) -> int:
+    """    GGML_API enum gguf_type gguf_get_kv_type (struct gguf_context * ctx, int i);"""
+    ...
+  def gguf_get_meta_data(ctx: ffi.CData, data: ffi.CData) -> None:
+    """    GGML_API void   gguf_get_meta_data(struct gguf_context * ctx, void * data);"""
+    ...
+  def gguf_get_meta_size(ctx: ffi.CData) -> int:
+    """
+    get the size in bytes of the meta data (header, kv pairs, tensor info) including padding
+
+        GGML_API size_t gguf_get_meta_size(struct gguf_context * ctx);
+    """
+    ...
+  def gguf_get_n_kv(ctx: ffi.CData) -> int:
+    """    GGML_API int          gguf_get_n_kv(struct gguf_context * ctx);"""
+    ...
+  def gguf_get_n_tensors(ctx: ffi.CData) -> int:
+    """    GGML_API int    gguf_get_n_tensors    (struct gguf_context * ctx);"""
+    ...
+  def gguf_get_tensor_name(ctx: ffi.CData, i: int) -> ffi.CData:
+    """    GGML_API char * gguf_get_tensor_name  (struct gguf_context * ctx, int i);"""
+    ...
+  def gguf_get_tensor_offset(ctx: ffi.CData, i: int) -> int:
+    """    GGML_API size_t gguf_get_tensor_offset(struct gguf_context * ctx, int i);"""
+    ...
+  def gguf_get_val_bool(ctx: ffi.CData, i: int) -> bool:
+    """    GGML_API bool         gguf_get_val_bool(struct gguf_context * ctx, int i);"""
+    ...
+  def gguf_get_val_f32(ctx: ffi.CData, i: int) -> float:
+    """    GGML_API float        gguf_get_val_f32 (struct gguf_context * ctx, int i);"""
+    ...
+  def gguf_get_val_i16(ctx: ffi.CData, i: int) -> int:
+    """    GGML_API int16_t      gguf_get_val_i16 (struct gguf_context * ctx, int i);"""
+    ...
+  def gguf_get_val_i32(ctx: ffi.CData, i: int) -> int:
+    """    GGML_API int32_t      gguf_get_val_i32 (struct gguf_context * ctx, int i);"""
+    ...
+  def gguf_get_val_i8(ctx: ffi.CData, i: int) -> int:
+    """    GGML_API int8_t       gguf_get_val_i8  (struct gguf_context * ctx, int i);"""
+    ...
+  def gguf_get_val_str(ctx: ffi.CData, i: int) -> ffi.CData:
+    """    GGML_API const char * gguf_get_val_str (struct gguf_context * ctx, int i);"""
+    ...
+  def gguf_get_val_u16(ctx: ffi.CData, i: int) -> int:
+    """    GGML_API uint16_t     gguf_get_val_u16 (struct gguf_context * ctx, int i);"""
+    ...
+  def gguf_get_val_u32(ctx: ffi.CData, i: int) -> int:
+    """    GGML_API uint32_t     gguf_get_val_u32 (struct gguf_context * ctx, int i);"""
+    ...
+  def gguf_get_val_u8(ctx: ffi.CData, i: int) -> int:
+    """
+    results are undefined if the wrong type is used for the key
+
+        GGML_API uint8_t      gguf_get_val_u8  (struct gguf_context * ctx, int i);
+    """
+    ...
+  def gguf_get_version(ctx: ffi.CData) -> int:
+    """    GGML_API int    gguf_get_version    (struct gguf_context * ctx);"""
+    ...
+  def gguf_init_empty() -> ffi.CData:
+    """    GGML_API struct gguf_context * gguf_init_empty(void);"""
+    ...
+  def gguf_init_from_file(fname: ffi.CData, params: ffi.CData) -> ffi.CData:
+    """    GGML_API struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_params params);"""
+    ...
+  def gguf_set_arr_data(ctx: ffi.CData, key: ffi.CData, type: int, data: ffi.CData, n: int) -> None:
+    """    GGML_API void gguf_set_arr_data(struct gguf_context * ctx, const char * key, enum gguf_type type, const void * data, int n);"""
+    ...
+  def gguf_set_arr_str(ctx: ffi.CData, key: ffi.CData, data: ffi.CData, n: int) -> None:
+    """    GGML_API void gguf_set_arr_str (struct gguf_context * ctx, const char * key, const char ** data, int n);"""
+    ...
+  def gguf_set_kv(ctx: ffi.CData, src: ffi.CData) -> None:
+    """
+    set or add KV pairs from another context
+
+        GGML_API void gguf_set_kv(struct gguf_context * ctx, struct gguf_context * src);
+    """
+    ...
+  def gguf_set_tensor_data(ctx: ffi.CData, name: ffi.CData, data: ffi.CData, size: int) -> None:
+    """    GGML_API void gguf_set_tensor_data(struct gguf_context * ctx, const char * name, const void * data, size_t size);"""
+    ...
+  def gguf_set_tensor_type(ctx: ffi.CData, name: ffi.CData, type: int) -> None:
+    """    GGML_API void gguf_set_tensor_type(struct gguf_context * ctx, const char * name, enum ggml_type type);"""
+    ...
+  def gguf_set_val_bool(ctx: ffi.CData, key: ffi.CData, val: bool) -> None:
+    """    GGML_API void gguf_set_val_bool(struct gguf_context * ctx, const char * key, bool     val);"""
+    ...
+  def gguf_set_val_f32(ctx: ffi.CData, key: ffi.CData, val: float) -> None:
+    """    GGML_API void gguf_set_val_f32 (struct gguf_context * ctx, const char * key, float    val);"""
+    ...
+  def gguf_set_val_i16(ctx: ffi.CData, key: ffi.CData, val: int) -> None:
+    """    GGML_API void gguf_set_val_i16 (struct gguf_context * ctx, const char * key, int16_t  val);"""
+    ...
+  def gguf_set_val_i32(ctx: ffi.CData, key: ffi.CData, val: int) -> None:
+    """    GGML_API void gguf_set_val_i32 (struct gguf_context * ctx, const char * key, int32_t  val);"""
+    ...
+  def gguf_set_val_i8(ctx: ffi.CData, key: ffi.CData, val: int) -> None:
+    """    GGML_API void gguf_set_val_i8  (struct gguf_context * ctx, const char * key, int8_t   val);"""
+    ...
+  def gguf_set_val_str(ctx: ffi.CData, key: ffi.CData, val: ffi.CData) -> None:
+    """    GGML_API void gguf_set_val_str (struct gguf_context * ctx, const char * key, const char * val);"""
+    ...
+  def gguf_set_val_u16(ctx: ffi.CData, key: ffi.CData, val: int) -> None:
+    """    GGML_API void gguf_set_val_u16 (struct gguf_context * ctx, const char * key, uint16_t val);"""
+    ...
+  def gguf_set_val_u32(ctx: ffi.CData, key: ffi.CData, val: int) -> None:
+    """    GGML_API void gguf_set_val_u32 (struct gguf_context * ctx, const char * key, uint32_t val);"""
+    ...
+  def gguf_set_val_u8(ctx: ffi.CData, key: ffi.CData, val: int) -> None:
+    """
+    overrides existing values or adds a new one
+
+        GGML_API void gguf_set_val_u8  (struct gguf_context * ctx, const char * key, uint8_t  val);
+    """
+    ...
+  def gguf_type_name(type: int) -> ffi.CData:
+    """    GGML_API const char * gguf_type_name(enum gguf_type type);"""
+    ...
+  def gguf_write_to_file(ctx: ffi.CData, fname: ffi.CData, only_meta: bool) -> None:
+    """
+    write the entire context to a binary file
+
+        GGML_API void gguf_write_to_file(struct gguf_context * ctx, const char * fname, bool only_meta);
+    """
+    ...
+  def quantize_row_q2_K(x: ffi.CData, y: ffi.CData, k: int) -> None:
+    """void quantize_row_q2_K(const float * restrict x, void * restrict y, int k);"""
+    ...
+  def quantize_row_q2_K_reference(x: ffi.CData, y: ffi.CData, k: int) -> None:
+    """
+    Quantization
+
+    void quantize_row_q2_K_reference(const float * restrict x, block_q2_K * restrict y, int k);
+    """
+    ...
+  def quantize_row_q3_K(x: ffi.CData, y: ffi.CData, k: int) -> None:
+    """void quantize_row_q3_K(const float * restrict x, void * restrict y, int k);"""
+    ...
+  def quantize_row_q3_K_reference(x: ffi.CData, y: ffi.CData, k: int) -> None:
+    """void quantize_row_q3_K_reference(const float * restrict x, block_q3_K * restrict y, int k);"""
+    ...
+  def quantize_row_q4_K(x: ffi.CData, y: ffi.CData, k: int) -> None:
+    """void quantize_row_q4_K(const float * restrict x, void * restrict y, int k);"""
+    ...
+  def quantize_row_q4_K_reference(x: ffi.CData, y: ffi.CData, k: int) -> None:
+    """void quantize_row_q4_K_reference(const float * restrict x, block_q4_K * restrict y, int k);"""
+    ...
+  def quantize_row_q5_K(x: ffi.CData, y: ffi.CData, k: int) -> None:
+    """void quantize_row_q5_K(const float * restrict x, void * restrict y, int k);"""
+    ...
+  def quantize_row_q5_K_reference(x: ffi.CData, y: ffi.CData, k: int) -> None:
+    """void quantize_row_q5_K_reference(const float * restrict x, block_q5_K * restrict y, int k);"""
+    ...
+  def quantize_row_q6_K(x: ffi.CData, y: ffi.CData, k: int) -> None:
+    """void quantize_row_q6_K(const float * restrict x, void * restrict y, int k);"""
+    ...
+  def quantize_row_q6_K_reference(x: ffi.CData, y: ffi.CData, k: int) -> None:
+    """void quantize_row_q6_K_reference(const float * restrict x, block_q6_K * restrict y, int k);"""
+    ...
+  def quantize_row_q8_K(x: ffi.CData, y: ffi.CData, k: int) -> None:
+    """void quantize_row_q8_K(const float * restrict x, void * restrict y, int k);"""
+    ...
+  def quantize_row_q8_K_reference(x: ffi.CData, y: ffi.CData, k: int) -> None:
+    """void quantize_row_q8_K_reference(const float * restrict x, block_q8_K * restrict y, int k);"""
+    ...

ggml/examples/python/ggml/ffi/__init__.pyi

@@ -0,0 +1,7 @@
+# Phony stubs.
+
+class CData:
+    pass
+
+class CType:
+    pass

ggml/examples/python/ggml/utils.py

@@ -0,0 +1,182 @@
+"""
+  Common helpers for working with ggml + numpy
+"""
+from ggml import ffi, lib
+from typing import Union, Optional
+import numpy as np
+
+def init(mem_size: int, mem_buffer: ffi.CData = ffi.NULL, no_alloc: bool = False) -> ffi.CData:
+    """
+      Initialize a ggml context, which will be freed automatically when the pointer is garbage collected.
+    """
+    params = ffi.new('struct ggml_init_params*')
+    params.mem_size = mem_size
+    params.mem_buffer = mem_buffer
+    params.no_alloc = no_alloc
+    return ffi.gc(lib.ggml_init(params[0]), lib.ggml_free)
+ 
+TensorLike = Union[ffi.CData, np.ndarray]
+
+def copy(from_tensor: TensorLike, to_tensor: TensorLike, allow_requantize: bool = True):
+    """
+      Copy the contents of one tensor to another, doing any necessary (de/re)quantization transparently.
+      Works across numpy & ggml tensors, but they must have the same shape (and be contiguous).
+
+      Parameters
+      ----------
+      from_tensor : TensorLike
+          The tensor to copy from (a numpy array or possibly-quantized ggml tensor)
+      to_tensor : TensorLike
+          The tensor to copy to (a numpy array or possibly-quantized ggml tensor)
+      allow_requantize : bool
+          If False, will throw an error if requantization is required (i.e. both from_tensor
+          and to_tensor are quantized with different quantization types)
+    """
+    if id(from_tensor) == id(to_tensor):
+        return
+ 
+    __expect_same_layout("source", from_tensor, "destination", to_tensor)
+    __check_shape_consistent_with_type(from_tensor)
+    __check_shape_consistent_with_type(to_tensor)
+
+    from_type = __get_type(from_tensor)
+    to_type = __get_type(to_tensor)
+
+    if from_type == to_type:
+        ffi.memmove(__get_data(to_tensor), __get_data(from_tensor), __get_nbytes(from_tensor))
+    else:
+        assert allow_requantize or not lib.ggml_is_quantized(from_type) or not lib.ggml_is_quantized(to_type), \
+            f"Requantizing from {__type_name(from_type)} to {__type_name(to_type)} is disabled. Force with allow_requantize=True"
+ 
+        __set_floats(to_tensor, __get_floats(from_tensor))
+
+def numpy(tensor: ffi.CData, allow_copy: Union[bool, np.ndarray] = False, allow_requantize=False) -> np.ndarray:
+    """
+      Convert a ggml tensor to a numpy array.
+      If the tensor isn't quantized, the returned numpy array will be a view over its data.
+ 
+      If it is quantized (and allow_copy is True), the copy will involve dequantization and the returned array will
+      be a copy of the original tensor (any changes to the numpy array won't then be reflected back to the tensor).
+
+      Parameters
+      ----------
+      tensor : ffi.CData
+          The tensor to convert to a numpy array
+      allow_copy : bool or np.ndarray
+          If False, will throw an error if the tensor is quantized (since dequantization requires extra memory).
+          If True, will dequantize the tensor and return a copy of the data in a new float32 numpy array.
+          If an np.ndarray, will copy the data into the given array (which must be the same shape as the tensor) when dequantization is needed
+      allow_requantize : bool
+          If allow_copy is a tensor with a different quantization type than the source tensor, will throw an error unless allow_requantize is True.
+    """
+    shape = __get_shape(tensor)
+
+    if lib.ggml_is_quantized(tensor.type):
+        if allow_copy == False:
+            raise ValueError(f"{__describe(tensor)} is quantized, conversion to numpy requires a copy (pass allow_copy=True; changes to the numpy array won't affect the original).")
+        elif isinstance(allow_copy, np.ndarray):
+            __expect_same_layout("source tensor", tensor, "dequantization output tensor", allow_copy)
+            destination = allow_copy
+        else:
+            destination = np.empty(shape, dtype=np.float32)
+
+        copy(tensor, destination, allow_requantize=allow_requantize)
+        return destination
+    else:
+        dtype = __type_to_dtype(tensor.type)
+        if not dtype:
+            raise NotImplementedError(f'Cannot convert {__describe(tensor)} to numpy')
+
+        assert __is_contiguous(tensor), f"Cannot convert {__describe(tensor)} to numpy (support contiguous tensors only)"
+        nbytes = lib.ggml_nelements(tensor) * lib.ggml_type_size(tensor.type)
+        array = np.frombuffer(ffi.buffer(lib.ggml_get_data(tensor), nbytes), dtype=dtype)
+        array.shape = shape
+        return array
+
+def __type_name(type: int) -> str:
+    name = lib.ggml_type_name(type)
+    return ffi.string(name).decode('utf-8') if name else None
+
+__k_quant_types = set([
+  lib.GGML_TYPE_Q2_K,
+  lib.GGML_TYPE_Q3_K,
+  lib.GGML_TYPE_Q4_K,
+  lib.GGML_TYPE_Q5_K,
+  lib.GGML_TYPE_Q6_K,
+  lib.GGML_TYPE_Q8_K,
+])
+
+__type_to_dtype_dict = {
+  lib.GGML_TYPE_I8: np.int8,
+  lib.GGML_TYPE_I16: np.int16,
+  lib.GGML_TYPE_I32: np.int32,
+  lib.GGML_TYPE_F16: np.float16,
+  lib.GGML_TYPE_F32: np.float32,
+}
+
+def __type_to_dtype(type: int) -> Optional[np.dtype]: return __type_to_dtype_dict.get(type)
+def __dtype_to_type(dtype: np.dtype):
+    if dtype == np.float32: return lib.GGML_TYPE_F32
+    elif dtype == np.float16: return lib.GGML_TYPE_F16
+    elif dtype == np.int32: return lib.GGML_TYPE_I32
+    elif dtype == np.int16: return lib.GGML_TYPE_I16
+    elif dtype == np.int8: return lib.GGML_TYPE_I8
+    else: raise ValueError(f"Unsupported dtype: {dtype}")
+
+def __describe(tensor: ffi.CType): return f'Tensor[{__type_name(__get_type(tensor))}, {__get_shape(tensor)}]'
+def __get_type(tensor: TensorLike): return __dtype_to_type(tensor.dtype) if isinstance(tensor, np.ndarray) else tensor.type
+def __get_shape(x: TensorLike): return x.shape if isinstance(x, np.ndarray) else tuple([x.ne[i] for i in range(x.n_dims)])
+def __get_strides(x: TensorLike): return x.strides if isinstance(x, np.ndarray) else tuple([x.nb[i] for i in range(x.n_dims)])
+def __get_data(x: TensorLike) -> ffi.CData: return ffi.from_buffer(x) if isinstance(x, np.ndarray) else lib.ggml_get_data(x)
+def __get_nbytes(tensor: TensorLike): return tensor.nbytes if isinstance(tensor, np.ndarray) else lib.ggml_nbytes(tensor)
+def __get_nelements(tensor: TensorLike): return tensor.size if isinstance(tensor, np.ndarray) else lib.ggml_nelements(tensor)
+def __is_contiguous(tensor: TensorLike): return tensor.flags['C_CONTIGUOUS'] if isinstance(tensor, np.ndarray) else lib.ggml_is_contiguous(tensor)
+
+def __get_floats(tensor: TensorLike) -> ffi.CData:
+    data, type = __get_data(tensor), __get_type(tensor)
+    if type == lib.GGML_TYPE_F32:
+        return ffi.cast('float*', data)
+    else:
+      nelements = __get_nelements(tensor)
+      floats = ffi.new('float[]', nelements)
+      if type == lib.GGML_TYPE_F16:
+          lib.ggml_fp16_to_fp32_row(ffi.cast('uint16_t*', data), floats, nelements)
+      elif lib.ggml_is_quantized(type):
+          qtype = lib.ggml_internal_get_type_traits(type)
+          assert qtype.to_float, f"Type {__type_name(type)} is not supported by ggml"
+          qtype.to_float(data, floats, nelements)
+      else:
+          raise NotImplementedError(f'Cannot read floats from {__describe(tensor)}')
+      return floats
+
+def __set_floats(tensor: TensorLike, f32_data: ffi.CData) -> None:
+    data, type, nbytes = __get_data(tensor), __get_type(tensor), __get_nbytes(tensor)
+    if type == lib.GGML_TYPE_F32:
+        ffi.memmove(data, f32_data, nbytes)
+    else:
+      nelements = __get_nelements(tensor)
+      if type == lib.GGML_TYPE_F16:
+          lib.ggml_fp32_to_fp16_row(f32_data, ffi.cast('uint16_t*', data), nelements)
+      elif lib.ggml_is_quantized(type):
+          qtype = lib.ggml_internal_get_type_traits(type)
+          assert qtype.from_float, f"Type {__type_name(type)} is not supported by ggml"
+          qtype.from_float(f32_data, data, nelements)
+      else:
+          raise NotImplementedError(f'Cannot write floats to {__describe(tensor)}')
+
+def __expect_same_layout(name1: str, tensor1: TensorLike, name2: str, tensor2: TensorLike):
+    shape1, shape2 = __get_shape(tensor1), __get_shape(tensor2)
+    assert shape1 == shape2, f"Shape mismatch: {name1} has {shape1} but {name2} has {shape2}"
+    assert __is_contiguous(tensor1) and __is_contiguous(tensor2), f"Only contiguous tensors are supported (got {name1} with strides {__get_strides(tensor1)} and {name2} with strides {__get_strides(tensor2)})"
+
+def __check_shape_consistent_with_type(tensor: TensorLike):
+    type = __get_type(tensor)
+    if not lib.ggml_is_quantized(type):
+        return
+    shape = __get_shape(tensor)
+
+    block_size = lib.ggml_blck_size(type)
+    assert not (block_size == 0 and type in __k_quant_types), f"Can't quantize, native library was not compiled with USE_K_QUANTS!"
+    assert block_size > 0, f"Invalid block size {block_size} for type {__type_name(type)}"
+    for i, d in enumerate(shape):
+        assert d % block_size == 0, f"Dimension {i} of {__describe(tensor)} is not divisible by {block_size}, required for quantization."

ggml/examples/python/regenerate.py

@@ -0,0 +1,42 @@
+# Generates bindings for the ggml library.
+#
+# cffi requires prior C preprocessing of the headers, and it uses pycparser which chokes on a couple of things
+# so we help it a bit (e.g. replace sizeof expressions with their value, remove exotic syntax found in Darwin headers).
+import os, sys, re, subprocess
+import cffi
+from stubs import generate_stubs
+
+API = os.environ.get('API', 'api.h')
+CC = os.environ.get('CC') or 'gcc'
+C_INCLUDE_DIR = os.environ.get('C_INCLUDE_DIR', '../../../llama.cpp')
+CPPFLAGS = [
+    "-I", C_INCLUDE_DIR,
+    '-D__fp16=uint16_t',  # pycparser doesn't support __fp16
+    '-D__attribute__(x)=',
+    '-D_Static_assert(x, m)=',
+] + [x for x in os.environ.get('CPPFLAGS', '').split(' ') if x != '']
+
+try: header = subprocess.run([CC, "-E", *CPPFLAGS, API], capture_output=True, text=True, check=True).stdout
+except subprocess.CalledProcessError as e: print(f'{e.stderr}\n{e}', file=sys.stderr); raise
+
+header = '\n'.join([l for l in header.split('\n') if '__darwin_va_list' not in l]) # pycparser hates this
+
+# Replace constant size expressions w/ their value (compile & run a mini exe for each, because why not).
+# First, extract anyting *inside* square brackets and anything that looks like a sizeof call.
+for expr in set(re.findall(f'(?<=\\[)[^\\]]+(?=])|sizeof\\s*\\([^()]+\\)', header)):
+    if re.match(r'^(\d+|\s*)$', expr): continue # skip constants and empty bracket contents
+    subprocess.run([CC, "-o", "eval_size_expr", *CPPFLAGS, "-x", "c", "-"], text=True, check=True,
+                   input=f'''#include <stdio.h>
+                             #include "{API}"
+                             int main() {{ printf("%lu", (size_t)({expr})); }}''')
+    size = subprocess.run(["./eval_size_expr"], capture_output=True, text=True, check=True).stdout
+    print(f'Computed constexpr {expr} = {size}')
+    header = header.replace(expr, size)
+
+ffibuilder = cffi.FFI()
+ffibuilder.cdef(header)
+ffibuilder.set_source(f'ggml.cffi', None) # we're not compiling a native extension, as this quickly gets hairy
+ffibuilder.compile(verbose=True)
+
+with open("ggml/__init__.pyi", "wt") as f:
+    f.write(generate_stubs(header))

ggml/examples/python/stubs.py

@@ -0,0 +1,128 @@
+"""
+  This generates .pyi stubs for the cffi Python bindings generated by regenerate.py
+"""
+import sys, re, itertools
+sys.path.extend(['.', '..']) # for pycparser
+
+from pycparser import c_ast, parse_file, CParser
+import pycparser.plyparser
+from pycparser.c_ast import PtrDecl, TypeDecl, FuncDecl, EllipsisParam, IdentifierType, Struct, Enum, Typedef
+from typing import Tuple
+
+__c_type_to_python_type = {
+    'void': 'None', '_Bool': 'bool',
+    'char': 'int', 'short': 'int', 'int': 'int', 'long': 'int',
+    'ptrdiff_t': 'int', 'size_t': 'int',
+    'int8_t': 'int', 'uint8_t': 'int',
+    'int16_t': 'int', 'uint16_t': 'int',
+    'int32_t': 'int', 'uint32_t': 'int',
+    'int64_t': 'int', 'uint64_t': 'int',
+    'float': 'float', 'double': 'float',
+    'ggml_fp16_t': 'np.float16',
+}
+
+def format_type(t: TypeDecl):
+    if isinstance(t, PtrDecl) or isinstance(t, Struct):
+        return 'ffi.CData'
+    if isinstance(t, Enum):
+        return 'int'
+    if isinstance(t, TypeDecl):
+        return format_type(t.type)
+    if isinstance(t, IdentifierType):
+        assert len(t.names) == 1, f'Expected a single name, got {t.names}'
+        return __c_type_to_python_type.get(t.names[0]) or 'ffi.CData'
+    return t.name
+
+class PythonStubFuncDeclVisitor(c_ast.NodeVisitor):
+    def __init__(self):
+        self.sigs = {}
+        self.sources = {}
+
+    def get_source_snippet_lines(self, coord: pycparser.plyparser.Coord) -> Tuple[list[str], list[str]]:
+        if coord.file not in self.sources:
+            with open(coord.file, 'rt') as f:
+                self.sources[coord.file] = f.readlines()
+        source_lines = self.sources[coord.file]
+        ncomment_lines = len(list(itertools.takewhile(lambda i: re.search(r'^\s*(//|/\*)', source_lines[i]), range(coord.line - 2, -1, -1))))
+        comment_lines = [l.strip() for l in source_lines[coord.line - 1 - ncomment_lines:coord.line - 1]]
+        decl_lines = []
+        for line in source_lines[coord.line - 1:]:
+            decl_lines.append(line.rstrip())
+            if (';' in line) or ('{' in line): break
+        return (comment_lines, decl_lines)
+
+    def visit_Enum(self, node: Enum):
+        if node.values is not None:
+          for e in node.values.enumerators:
+              self.sigs[e.name] = f'  @property\n  def {e.name}(self) -> int: ...'
+
+    def visit_Typedef(self, node: Typedef):
+        pass
+
+    def visit_FuncDecl(self, node: FuncDecl):
+        ret_type = node.type
+        is_ptr = False
+        while isinstance(ret_type, PtrDecl):
+            ret_type = ret_type.type
+            is_ptr = True
+
+        fun_name = ret_type.declname
+        if fun_name.startswith('__'):
+            return
+
+        args = []
+        argnames = []
+        def gen_name(stem):
+            i = 1
+            while True:
+                new_name = stem if i == 1 else f'{stem}{i}'
+                if new_name not in argnames: return new_name
+                i += 1
+
+        for a in node.args.params:
+            if isinstance(a, EllipsisParam):
+                arg_name = gen_name('args')
+                argnames.append(arg_name)
+                args.append('*' + gen_name('args'))
+            elif format_type(a.type) == 'None':
+                continue
+            else:
+                arg_name = a.name or gen_name('arg')
+                argnames.append(arg_name)
+                args.append(f'{arg_name}: {format_type(a.type)}')
+
+        ret = format_type(ret_type if not is_ptr else node.type)
+
+        comment_lines, decl_lines = self.get_source_snippet_lines(node.coord)
+
+        lines = [f'  def {fun_name}({", ".join(args)}) -> {ret}:']
+        if len(comment_lines) == 0 and len(decl_lines) == 1:
+            lines += [f'    """{decl_lines[0]}"""']
+        else:
+            lines += ['    """']
+            lines += [f'    {c.lstrip("/* ")}' for c in comment_lines]
+            if len(comment_lines) > 0:
+                lines += ['']
+            lines += [f'    {d}' for d in decl_lines]
+            lines += ['    """']
+        lines += ['    ...']
+        self.sigs[fun_name] = '\n'.join(lines)
+
+def generate_stubs(header: str):
+    """
+      Generates a .pyi Python stub file for the GGML API using C header files.
+    """
+
+    v = PythonStubFuncDeclVisitor()
+    v.visit(CParser().parse(header, "<input>"))
+
+    keys = list(v.sigs.keys())
+    keys.sort()
+
+    return '\n'.join([
+        '# auto-generated file',
+        'import ggml.ffi as ffi',
+        'import numpy as np',
+        'class lib:',
+        *[v.sigs[k] for k in keys]
+    ])

ggml/examples/python/test_tensor.py

@@ -0,0 +1,258 @@
+import pytest
+from pytest import raises
+
+from ggml import lib, ffi
+from ggml.utils import init, copy, numpy
+import numpy as np
+import numpy.testing as npt
+
+@pytest.fixture()
+def ctx():
+    print("setup")
+    yield init(mem_size=10*1024*1024)
+    print("teardown")
+
+class TestNumPy:
+    
+    # Single element
+
+    def test_set_get_single_i32(self, ctx):
+        i = lib.ggml_new_i32(ctx, 42)
+        assert lib.ggml_get_i32_1d(i, 0) == 42
+        assert numpy(i) == np.array([42], dtype=np.int32)
+
+    def test_set_get_single_f32(self, ctx):
+        i = lib.ggml_new_f32(ctx, 4.2)
+        
+        epsilon = 0.000001 # Not sure why so large a difference??
+        pytest.approx(lib.ggml_get_f32_1d(i, 0), 4.2, epsilon)
+        pytest.approx(numpy(i), np.array([4.2], dtype=np.float32), epsilon)
+
+    def _test_copy_np_to_ggml(self, a: np.ndarray, t: ffi.CData):
+        a2 = a.copy() # Clone original
+        copy(a, t)
+        npt.assert_array_equal(numpy(t), a2)
+
+    # I32
+
+    def test_copy_np_to_ggml_1d_i32(self, ctx):
+        t = lib.ggml_new_tensor_1d(ctx, lib.GGML_TYPE_I32, 10)
+        a = np.arange(10, dtype=np.int32)
+        self._test_copy_np_to_ggml(a, t)
+
+    def test_copy_np_to_ggml_2d_i32(self, ctx):
+        t = lib.ggml_new_tensor_2d(ctx, lib.GGML_TYPE_I32, 2, 3)
+        a = np.arange(2 * 3, dtype=np.int32).reshape((2, 3))
+        self._test_copy_np_to_ggml(a, t)
+
+    def test_copy_np_to_ggml_3d_i32(self, ctx):
+        t = lib.ggml_new_tensor_3d(ctx, lib.GGML_TYPE_I32, 2, 3, 4)
+        a = np.arange(2 * 3 * 4, dtype=np.int32).reshape((2, 3, 4))
+        self._test_copy_np_to_ggml(a, t)
+
+    def test_copy_np_to_ggml_4d_i32(self, ctx):
+        t = lib.ggml_new_tensor_4d(ctx, lib.GGML_TYPE_I32, 2, 3, 4, 5)
+        a = np.arange(2 * 3 * 4 * 5, dtype=np.int32).reshape((2, 3, 4, 5))
+        self._test_copy_np_to_ggml(a, t)
+
+    def test_copy_np_to_ggml_4d_n_i32(self, ctx):
+        dims = [2, 3, 4, 5] # GGML_MAX_DIMS is 4, going beyond would crash
+        pdims = ffi.new('int64_t[]', len(dims))
+        for i, d in enumerate(dims): pdims[i] = d
+        t = lib.ggml_new_tensor(ctx, lib.GGML_TYPE_I32, len(dims), pdims)
+        a = np.arange(np.prod(dims), dtype=np.int32).reshape(tuple(pdims))
+        self._test_copy_np_to_ggml(a, t)
+
+    # F32
+
+    def test_copy_np_to_ggml_1d_f32(self, ctx):
+        t = lib.ggml_new_tensor_1d(ctx, lib.GGML_TYPE_F32, 10)
+        a = np.arange(10, dtype=np.float32)
+        self._test_copy_np_to_ggml(a, t)
+
+    def test_copy_np_to_ggml_2d_f32(self, ctx):
+        t = lib.ggml_new_tensor_2d(ctx, lib.GGML_TYPE_F32, 2, 3)
+        a = np.arange(2 * 3, dtype=np.float32).reshape((2, 3))
+        self._test_copy_np_to_ggml(a, t)
+
+    def test_copy_np_to_ggml_3d_f32(self, ctx):
+        t = lib.ggml_new_tensor_3d(ctx, lib.GGML_TYPE_F32, 2, 3, 4)
+        a = np.arange(2 * 3 * 4, dtype=np.float32).reshape((2, 3, 4))
+        self._test_copy_np_to_ggml(a, t)
+
+    def test_copy_np_to_ggml_4d_f32(self, ctx):
+        t = lib.ggml_new_tensor_4d(ctx, lib.GGML_TYPE_F32, 2, 3, 4, 5)
+        a = np.arange(2 * 3 * 4 * 5, dtype=np.float32).reshape((2, 3, 4, 5))
+        self._test_copy_np_to_ggml(a, t)
+
+    def test_copy_np_to_ggml_4d_n_f32(self, ctx):
+        dims = [2, 3, 4, 5] # GGML_MAX_DIMS is 4, going beyond would crash
+        pdims = ffi.new('int64_t[]', len(dims))
+        for i, d in enumerate(dims): pdims[i] = d
+        t = lib.ggml_new_tensor(ctx, lib.GGML_TYPE_F32, len(dims), pdims)
+        a = np.arange(np.prod(dims), dtype=np.float32).reshape(tuple(pdims))
+        self._test_copy_np_to_ggml(a, t)
+
+    # F16
+
+    def test_copy_np_to_ggml_1d_f16(self, ctx):
+        t = lib.ggml_new_tensor_1d(ctx, lib.GGML_TYPE_F16, 10)
+        a = np.arange(10, dtype=np.float16)
+        self._test_copy_np_to_ggml(a, t)
+
+    def test_copy_np_to_ggml_2d_f16(self, ctx):
+        t = lib.ggml_new_tensor_2d(ctx, lib.GGML_TYPE_F16, 2, 3)
+        a = np.arange(2 * 3, dtype=np.float16).reshape((2, 3))
+        self._test_copy_np_to_ggml(a, t)
+
+    def test_copy_np_to_ggml_3d_f16(self, ctx):
+        t = lib.ggml_new_tensor_3d(ctx, lib.GGML_TYPE_F16, 2, 3, 4)
+        a = np.arange(2 * 3 * 4, dtype=np.float16).reshape((2, 3, 4))
+        self._test_copy_np_to_ggml(a, t)
+
+    def test_copy_np_to_ggml_4d_f16(self, ctx):
+        t = lib.ggml_new_tensor_4d(ctx, lib.GGML_TYPE_F16, 2, 3, 4, 5)
+        a = np.arange(2 * 3 * 4 * 5, dtype=np.float16).reshape((2, 3, 4, 5))
+        self._test_copy_np_to_ggml(a, t)
+
+    def test_copy_np_to_ggml_4d_n_f16(self, ctx):
+        dims = [2, 3, 4, 5] # GGML_MAX_DIMS is 4, going beyond would crash
+        pdims = ffi.new('int64_t[]', len(dims))
+        for i, d in enumerate(dims): pdims[i] = d
+        t = lib.ggml_new_tensor(ctx, lib.GGML_TYPE_F16, len(dims), pdims)
+        a = np.arange(np.prod(dims), dtype=np.float16).reshape(tuple(pdims))
+        self._test_copy_np_to_ggml(a, t)
+
+    # Mismatching shapes
+
+    def test_copy_mismatching_shapes_1d(self, ctx):
+        t = lib.ggml_new_tensor_1d(ctx, lib.GGML_TYPE_F32, 10)
+        a = np.arange(10, dtype=np.float32)
+        copy(a, t) # OK
+        
+        a = a.reshape((5, 2))
+        with raises(AssertionError): copy(a, t)
+        with raises(AssertionError): copy(t, a)
+            
+    def test_copy_mismatching_shapes_2d(self, ctx):
+        t = lib.ggml_new_tensor_2d(ctx, lib.GGML_TYPE_F32, 2, 3)
+        a = np.arange(6, dtype=np.float32)
+        copy(a.reshape((2, 3)), t) # OK
+        
+        a = a.reshape((3, 2))
+        with raises(AssertionError): copy(a, t)
+        with raises(AssertionError): copy(t, a)
+
+    def test_copy_mismatching_shapes_3d(self, ctx):
+        t = lib.ggml_new_tensor_3d(ctx, lib.GGML_TYPE_F32, 2, 3, 4)
+        a = np.arange(24, dtype=np.float32)
+        copy(a.reshape((2, 3, 4)), t) # OK
+        
+        a = a.reshape((2, 4, 3))
+        with raises(AssertionError): copy(a, t)
+        with raises(AssertionError): copy(t, a)
+
+    def test_copy_mismatching_shapes_4d(self, ctx):
+        t = lib.ggml_new_tensor_4d(ctx, lib.GGML_TYPE_F32, 2, 3, 4, 5)
+        a = np.arange(24*5, dtype=np.float32)
+        copy(a.reshape((2, 3, 4, 5)), t) # OK
+        
+        a = a.reshape((2, 3, 5, 4))
+        with raises(AssertionError): copy(a, t)
+        with raises(AssertionError): copy(t, a)
+
+    def test_copy_f16_to_f32(self, ctx):
+        t = lib.ggml_new_tensor_1d(ctx, lib.GGML_TYPE_F32, 1)
+        a = np.array([123.45], dtype=np.float16)
+        copy(a, t)
+        np.testing.assert_allclose(lib.ggml_get_f32_1d(t, 0), 123.45, rtol=1e-3)
+
+    def test_copy_f32_to_f16(self, ctx):
+        t = lib.ggml_new_tensor_1d(ctx, lib.GGML_TYPE_F16, 1)
+        a = np.array([123.45], dtype=np.float32)
+        copy(a, t)
+        np.testing.assert_allclose(lib.ggml_get_f32_1d(t, 0), 123.45, rtol=1e-3)
+
+    def test_copy_f16_to_Q5_K(self, ctx):
+        n = 256
+        t = lib.ggml_new_tensor_1d(ctx, lib.GGML_TYPE_Q5_K, n)
+        a = np.arange(n, dtype=np.float16)
+        copy(a, t)
+        np.testing.assert_allclose(a, numpy(t, allow_copy=True), rtol=0.05)
+
+    def test_copy_Q5_K_to_f16(self, ctx):
+        n = 256
+        t = lib.ggml_new_tensor_1d(ctx, lib.GGML_TYPE_Q5_K, n)
+        copy(np.arange(n, dtype=np.float32), t)
+        a = np.arange(n, dtype=np.float16)
+        copy(t, a)
+        np.testing.assert_allclose(a, numpy(t, allow_copy=True), rtol=0.05)
+
+    def test_copy_i16_f32_mismatching_types(self, ctx):
+        t = lib.ggml_new_tensor_1d(ctx, lib.GGML_TYPE_F32, 1)
+        a = np.arange(1, dtype=np.int16)
+        with raises(NotImplementedError): copy(a, t)
+        with raises(NotImplementedError): copy(t, a)
+
+class TestTensorCopy:
+
+    def test_copy_self(self, ctx):
+        t = lib.ggml_new_i32(ctx, 42)
+        copy(t, t)
+        assert lib.ggml_get_i32_1d(t, 0) == 42
+
+    def test_copy_1d(self, ctx):
+        t1 = lib.ggml_new_tensor_1d(ctx, lib.GGML_TYPE_F32, 10)
+        t2 = lib.ggml_new_tensor_1d(ctx, lib.GGML_TYPE_F32, 10)
+        a = np.arange(10, dtype=np.float32)
+        copy(a, t1)
+        copy(t1, t2)
+        assert np.allclose(a, numpy(t2))
+        assert np.allclose(numpy(t1), numpy(t2))
+
+class TestGraph:
+
+    def test_add(self, ctx):
+        n = 256
+        ta = lib.ggml_new_tensor_1d(ctx, lib.GGML_TYPE_F32, n)
+        tb = lib.ggml_new_tensor_1d(ctx, lib.GGML_TYPE_F32, n)
+        tsum = lib.ggml_add(ctx, ta, tb)
+        assert tsum.type == lib.GGML_TYPE_F32
+
+        gf = ffi.new('struct ggml_cgraph*')
+        lib.ggml_build_forward_expand(gf, tsum)
+
+        a = np.arange(0, n, dtype=np.float32)
+        b = np.arange(n, 0, -1, dtype=np.float32)
+        copy(a, ta)
+        copy(b, tb)
+
+        lib.ggml_graph_compute_with_ctx(ctx, gf, 1)
+
+        assert np.allclose(numpy(tsum, allow_copy=True), a + b)
+
+class TestQuantization:
+
+    def test_quantized_add(self, ctx):
+        n = 256
+        ta = lib.ggml_new_tensor_1d(ctx, lib.GGML_TYPE_Q5_K, n)
+        tb = lib.ggml_new_tensor_1d(ctx, lib.GGML_TYPE_F32, n)
+        tsum = lib.ggml_add(ctx, ta, tb)
+        assert tsum.type == lib.GGML_TYPE_Q5_K
+
+        gf = ffi.new('struct ggml_cgraph*')
+        lib.ggml_build_forward_expand(gf, tsum)
+
+        a = np.arange(0, n, dtype=np.float32)
+        b = np.arange(n, 0, -1, dtype=np.float32)
+        copy(a, ta)
+        copy(b, tb)
+
+        lib.ggml_graph_compute_with_ctx(ctx, gf, 1)
+
+        unquantized_sum = a + b
+        sum = numpy(tsum, allow_copy=True)
+
+        diff = np.linalg.norm(unquantized_sum - sum, np.inf)
+        assert diff > 4
+        assert diff < 5

ggml/examples/replit/CMakeLists.txt

@@ -0,0 +1,13 @@
+#
+# replit
+
+set(TEST_TARGET replit)
+add_executable(${TEST_TARGET} main.cpp)
+target_link_libraries(${TEST_TARGET} PRIVATE ggml common common-ggml)
+
+#
+# replit-quantize
+
+set(TEST_TARGET replit-quantize)
+add_executable(${TEST_TARGET} quantize.cpp)
+target_link_libraries(${TEST_TARGET} PRIVATE ggml common common-ggml)

ggml/examples/replit/convert-h5-to-ggml.py

@@ -0,0 +1,117 @@
+from pathlib import Path
+import sys
+import struct
+import json
+import numpy as np
+from transformers import AutoModelForCausalLM, AutoTokenizer
+import sentencepiece.sentencepiece_model_pb2 as model
+
+if len(sys.argv) < 3:
+    print("Usage: convert-h5-to-ggml.py dir-model [use-f32]\n")
+    print("  ftype == 0 -> float32")
+    print("  ftype == 1 -> float16")
+    sys.exit(1)
+
+
+# output in the same directory as the model
+dir_model = sys.argv[1]
+fname_out = sys.argv[1] + "/ggml-model.bin"
+
+
+with open(dir_model + "/config.json", "r", encoding="utf-8") as f:
+    hparams = json.load(f)
+
+sp_proto = model.ModelProto()
+sp_proto.ParseFromString(open(Path(sys.argv[1]) / "spiece.model", "rb").read())
+
+
+# possible data types
+#   ftype == 0 -> float32
+#   ftype == 1 -> float16
+#
+# map from ftype to string
+ftype_str = ["f32", "f16"]
+
+ftype = 1
+if len(sys.argv) > 2:
+    ftype = int(sys.argv[2])
+    if ftype < 0 or ftype > 1:
+        print("Invalid ftype: " + str(ftype))
+        sys.exit(1)
+    fname_out = sys.argv[1] + "/ggml-model-" + ftype_str[ftype] + ".bin"
+
+
+tokenizer = AutoTokenizer.from_pretrained(dir_model, trust_remote_code=True)
+model = AutoModelForCausalLM.from_pretrained(
+    dir_model, low_cpu_mem_usage=True, trust_remote_code=True
+)
+# print (model)
+
+# print(tokenizer.encode('I believe the meaning of life is'))
+
+list_vars = model.state_dict()
+for name in list_vars.keys():
+    print(name, list_vars[name].shape, list_vars[name].dtype)
+
+fout = open(fname_out, "wb")
+
+print(hparams)
+
+fout.write(struct.pack("i", 0x67676D6C))  # magic: ggml in hex
+fout.write(struct.pack("i", hparams["d_model"]))
+fout.write(struct.pack("i", hparams["max_seq_len"]))
+fout.write(struct.pack("i", hparams["n_heads"]))
+fout.write(struct.pack("i", hparams["n_layers"]))
+fout.write(struct.pack("i", hparams["vocab_size"]))
+fout.write(struct.pack("i", ftype))
+
+
+# TODO: temporary hack to not deal with implementing the tokenizer
+for piece in sp_proto.pieces:
+    encoded_piece = piece.piece.encode("utf-8")
+    fout.write(struct.pack("i", len(encoded_piece)))
+    fout.write(encoded_piece)
+    fout.write(struct.pack("f", piece.score))
+
+if hparams["vocab_size"] > len(sp_proto.pieces):
+    for i in range(hparams["vocab_size"] - len(sp_proto.pieces)):
+        fout.write(struct.pack("i", 0))
+        fout.write(struct.pack("f", 0))
+
+for name in list_vars.keys():
+    data = list_vars[name].squeeze().numpy()
+    print("Processing variable: " + name + " with shape: ", data.shape)
+
+    n_dims = len(data.shape)
+
+    # ftype == 0 -> float32, ftype == 1 -> float16
+    ftype_cur = 0
+    if ftype != 0:
+        if name[-7:] == ".weight" and n_dims == 2:
+            print("  Converting to float16")
+            data = data.astype(np.float16)
+            ftype_cur = 1
+        else:
+            print("  Converting to float32")
+            data = data.astype(np.float32)
+            ftype_cur = 0
+    else:
+        if data.dtype != np.float32:
+            print("  Converting to float32")
+            data = data.astype(np.float32)
+            ftype_cur = 0
+
+    # header
+    str = name.encode("utf-8")
+    fout.write(struct.pack("iii", n_dims, len(str), ftype_cur))
+    for i in range(n_dims):
+        fout.write(struct.pack("i", data.shape[n_dims - 1 - i]))
+    fout.write(str)
+
+    # data
+    data.tofile(fout)
+
+fout.close()
+
+print("Done. Output file: " + fname_out)
+print("")

ggml/examples/replit/main.cpp

@@ -0,0 +1,795 @@
+#include "ggml/ggml.h"
+
+#include "common-ggml.h"
+#include "common.h"
+
+#include <cassert>
+#include <cmath>
+#include <cinttypes>
+#include <cstddef>
+#include <cstdio>
+#include <cstring>
+#include <fstream>
+#include <iostream>
+#include <map>
+#include <stdint.h>
+#include <string>
+#include <unordered_map>
+#include <utility>
+#include <vector>
+
+#if defined(_WIN32)
+#define NOMINMAX
+#include <Windows.h>
+bool is_stdin_terminal() {
+    auto in = GetStdHandle(STD_INPUT_HANDLE);
+    return GetFileType(in) == FILE_TYPE_CHAR;
+}
+#else
+#include <unistd.h>
+bool is_stdin_terminal() {
+    return isatty(STDIN_FILENO);
+}
+#endif
+
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
+using piece_t = std::pair<std::size_t, float>;
+using piece_map_t = std::unordered_map<std::string, piece_t>;
+
+struct replit_tokenizer {
+    gpt_vocab raw_vocab;
+    piece_map_t piece_map;
+    std::vector<std::string> vocab;
+};
+
+std::pair<std::vector<std::size_t>, float> encode_word(const std::string & word, const piece_map_t & model) {
+    std::vector<int> best_segmentations_starts(word.length() + 1, -1);
+    best_segmentations_starts[0] = 0;
+
+    std::vector<float> best_segmentations_scores(word.length() + 1, -std::numeric_limits<float>::infinity());
+    best_segmentations_scores[0] = 1.0;
+
+    for (size_t start_idx = 0; start_idx < word.length(); ++start_idx) {
+        float best_score_at_start = best_segmentations_scores[start_idx];
+        for (size_t end_idx = start_idx + 1; end_idx <= word.length(); ++end_idx) {
+            std::string token = word.substr(start_idx, end_idx - start_idx);
+            if (model.count(token) && best_score_at_start != -std::numeric_limits<float>::infinity()) {
+                float token_score = model.at(token).second;
+                float score = token_score + best_score_at_start;
+                if (best_segmentations_scores[end_idx] == -std::numeric_limits<float>::infinity() ||
+                    best_segmentations_scores[end_idx] > score) {
+                    best_segmentations_starts[end_idx] = start_idx;
+                    best_segmentations_scores[end_idx] = score;
+                }
+            }
+        }
+    }
+
+    if (best_segmentations_scores.back() == -std::numeric_limits<float>::infinity()) {
+        return std::make_pair(std::vector<std::size_t>{0}, 0.0f);
+    }
+
+    float score = best_segmentations_scores.back();
+    int start = best_segmentations_starts.back();
+    int end = word.length();
+    std::vector<std::size_t> tokens;
+    while (start != 0) {
+        const auto token_id = model.at(word.substr(start, end - start)).first;
+        tokens.insert(tokens.begin(), token_id);
+        int next_start = best_segmentations_starts[start];
+        end = start;
+        start = next_start;
+    }
+    const auto token_id = model.at(word.substr(start, end - start)).first;
+    tokens.insert(tokens.begin(), token_id);
+    return std::make_pair(tokens, score);
+}
+
+bool replit_tokenizer_load(replit_tokenizer & tokenizer, std::istream & fin, int max_vocab_size) {
+    std::string word;
+    std::vector<char> buf(128);
+
+    for (int i = 0; i < max_vocab_size; i++) {
+        uint32_t len;
+        fin.read((char *)&len, sizeof(len));
+
+        buf.resize(len);
+        fin.read((char *)buf.data(), len);
+        word.assign(buf.data(), len);
+
+        float score;
+        fin.read((char *)&score, sizeof(score));
+
+        tokenizer.piece_map[word] = std::make_pair(i, -score);
+        tokenizer.raw_vocab.id_to_token[i] = word;
+    }
+
+    return true;
+}
+
+std::string replace_all(const std::string & str,    // where to work
+                        const std::string & find,   // substitute 'find'
+                        const std::string & replace //      by 'replace'
+) {
+    using namespace std;
+    string result;
+    size_t find_len = find.size();
+    size_t pos, from = 0;
+    while (string::npos != (pos = str.find(find, from))) {
+        result.append(str, from, pos - from);
+        result.append(replace);
+        from = pos + find_len;
+    }
+    result.append(str, from, string::npos);
+    return result;
+}
+
+std::string ws_symbol = "\342\226\201";
+std::vector<std::size_t> replit_tokenizer_tokenize(replit_tokenizer & tokenizer, const std::string & text) {
+    std::vector<std::size_t> tokens;
+    auto normalized_text = replace_all(text, " ", ws_symbol);
+    auto tokenized = encode_word(normalized_text, tokenizer.piece_map);
+
+    return tokenized.first;
+}
+
+std::string replit_tokenizer_detokenize(replit_tokenizer & tokenizer, const std::vector<std::size_t> & tokens) {
+    std::string text;
+    for (auto token : tokens) {
+        text += tokenizer.raw_vocab.id_to_token[token];
+    }
+    auto denormalized_text = replace_all(text, ws_symbol, " ");
+    return denormalized_text;
+}
+
+// no defaults for now
+struct replit_hparams {
+    int32_t d_model = 0;
+    int32_t max_seq_len = 0;
+    int32_t n_heads = 0;
+    int32_t n_layers = 0;
+    int32_t n_vocab = 0;
+    int32_t ftype = 0;
+};
+
+struct replit_layer {
+    // pre normalization
+    struct ggml_tensor * norm_1_weight;
+
+    // attention
+    struct ggml_tensor * c_attn_wqkv_weight;
+    struct ggml_tensor * c_attn_out_proj_weight;
+
+    // post normalization
+    struct ggml_tensor * norm_2_weight;
+
+    // ff
+    struct ggml_tensor * ffn_up_proj;
+    struct ggml_tensor * ffn_down_proj;
+};
+
+struct replit_model {
+    replit_hparams hparams;
+
+    struct ggml_tensor * wte_weight;    // position embedding
+    struct ggml_tensor * norm_f_weight; // language model head
+
+    std::vector<replit_layer> layers;
+
+    // key + value memory
+    struct ggml_tensor * memory_k;
+    struct ggml_tensor * memory_v;
+
+    struct ggml_context * ctx;
+    std::map<std::string, struct ggml_tensor *> tensors;
+};
+
+// load the model's weights from a file
+bool replit_model_load(const std::string & fname, replit_model & model, replit_tokenizer & vocab) {
+    printf("%s: loading model from '%s' - please wait ...\n", __func__, fname.c_str());
+
+    auto fin = std::ifstream(fname, std::ios::binary);
+    if (!fin) {
+        fprintf(stderr, "%s: failed to open '%s'\n", __func__, fname.c_str());
+        return false;
+    }
+
+    // verify magic
+    {
+        uint32_t magic;
+        fin.read((char *)&magic, sizeof(magic));
+        if (magic != GGML_FILE_MAGIC) {
+            fprintf(stderr, "%s: invalid model file '%s' (bad magic)\n", __func__, fname.c_str());
+            return false;
+        }
+    }
+
+    // load hparams
+    {
+        auto & hparams = model.hparams;
+
+        fin.read((char *)&hparams.d_model, sizeof(hparams.d_model));
+        fin.read((char *)&hparams.max_seq_len, sizeof(hparams.max_seq_len));
+        fin.read((char *)&hparams.n_heads, sizeof(hparams.n_heads));
+        fin.read((char *)&hparams.n_layers, sizeof(hparams.n_layers));
+        fin.read((char *)&hparams.n_vocab, sizeof(hparams.n_vocab));
+        fin.read((char *)&hparams.ftype, sizeof(hparams.ftype));
+
+        const int32_t qntvr = hparams.ftype / GGML_QNT_VERSION_FACTOR;
+
+        printf("%s: d_model      = %d\n", __func__, hparams.d_model);
+        printf("%s: max_seq_len  = %d\n", __func__, hparams.max_seq_len);
+        printf("%s: n_heads      = %d\n", __func__, hparams.n_heads);
+        printf("%s: n_layers     = %d\n", __func__, hparams.n_layers);
+        printf("%s: n_vocab      = %d\n", __func__, hparams.n_vocab);
+        printf("%s: ftype        = %d\n", __func__, hparams.ftype);
+        printf("%s: qntvr        = %d\n", __func__, qntvr);
+
+        hparams.ftype %= GGML_QNT_VERSION_FACTOR;
+    }
+
+    // load vocab
+    replit_tokenizer_load(vocab, fin, model.hparams.n_vocab);
+
+    // for the big tensors, we have the option to store the data in 16-bit
+    // floats or quantized in order to save memory and also to speed up the
+    // computation
+    ggml_type wtype = ggml_ftype_to_ggml_type((ggml_ftype)(model.hparams.ftype));
+    if (wtype == GGML_TYPE_COUNT) {
+        fprintf(stderr, "%s: invalid model file '%s' (bad ftype value %d)\n", __func__, fname.c_str(),
+                model.hparams.ftype);
+        return false;
+    }
+
+    auto & ctx = model.ctx;
+
+    size_t ctx_size = 0;
+
+    {
+        const auto & hparams = model.hparams;
+
+        const int n_embd = hparams.d_model;
+        const int n_layer = hparams.n_layers;
+        const int n_ctx = hparams.max_seq_len;
+        const int n_vocab = hparams.n_vocab;
+
+        ctx_size += n_embd * n_vocab * ggml_type_sizef(wtype); // wte_weight
+        ctx_size += n_embd * ggml_type_sizef(GGML_TYPE_F32);   // ln_f_weight
+
+        ctx_size += n_layer * (n_embd * ggml_type_sizef(GGML_TYPE_F32));      // ln_1_weight
+        ctx_size += n_layer * (3 * n_embd * n_embd * ggml_type_sizef(wtype)); // attn_Wqkv_weight
+        ctx_size += n_layer * (n_embd * n_embd * ggml_type_sizef(wtype));     // attn_out_proj_weight
+        ctx_size += n_layer * (n_embd * ggml_type_sizef(GGML_TYPE_F32));      // ln_2_weight
+        ctx_size += n_layer * (4 * n_embd * n_embd * ggml_type_sizef(wtype)); // mlp_mlp_up_weight
+        ctx_size += n_layer * (n_embd * n_embd * 4 * ggml_type_sizef(wtype)); // mlp_mlp_down_weight
+
+        ctx_size += n_ctx * n_layer * n_embd * ggml_type_sizef(GGML_TYPE_F16); // memory_k
+        ctx_size += n_ctx * n_layer * n_embd * ggml_type_sizef(GGML_TYPE_F16); // memory_v
+
+        ctx_size += (1 + 6 * n_layer) * 512; // object overhead
+
+        printf("%s: ggml ctx size = %6.2f MB\n", __func__, ctx_size / (1024.0 * 1024.0));
+    }
+
+    // create the ggml context
+    {
+        struct ggml_init_params params = {
+            /*.mem_size   =*/ ctx_size,
+            /*.mem_buffer =*/ NULL,
+            /*.no_alloc   =*/ false,
+        };
+
+        model.ctx = ggml_init(params);
+        if (!model.ctx) {
+            fprintf(stderr, "%s: ggml_init() failed\n", __func__);
+            return false;
+        }
+    }
+
+    // prepare memory for the weights
+    {
+        const auto & hparams = model.hparams;
+
+        const size_t n_embd = hparams.d_model;
+        const size_t n_layer = hparams.n_layers;
+        const size_t n_vocab = hparams.n_vocab;
+
+        model.layers.resize(n_layer);
+
+        model.wte_weight = ggml_new_tensor_2d(ctx, wtype, n_embd, n_vocab);
+        model.norm_f_weight = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
+
+        // map by name
+        model.tensors["transformer.wte.weight"] = model.wte_weight;
+        model.tensors["transformer.norm_f.weight"] = model.norm_f_weight;
+
+        for (int i = 0; i < (int)n_layer; ++i) {
+            auto & layer = model.layers[i];
+
+            layer.norm_1_weight = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
+            layer.c_attn_wqkv_weight = ggml_new_tensor_2d(ctx, wtype, n_embd, 3 * n_embd);
+            layer.c_attn_out_proj_weight = ggml_new_tensor_2d(ctx, wtype, n_embd, n_embd);
+            layer.norm_2_weight = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
+            layer.ffn_up_proj = ggml_new_tensor_2d(ctx, wtype, n_embd, 4 * n_embd);
+            layer.ffn_down_proj = ggml_new_tensor_2d(ctx, wtype, 4 * n_embd, n_embd);
+
+            // map by name
+            model.tensors["transformer.blocks." + std::to_string(i) + ".norm_1.weight"] = layer.norm_1_weight;
+            model.tensors["transformer.blocks." + std::to_string(i) + ".attn.Wqkv.weight"] = layer.c_attn_wqkv_weight;
+            model.tensors["transformer.blocks." + std::to_string(i) + ".attn.out_proj.weight"] =
+                layer.c_attn_out_proj_weight;
+            model.tensors["transformer.blocks." + std::to_string(i) + ".norm_2.weight"] = layer.norm_2_weight;
+            model.tensors["transformer.blocks." + std::to_string(i) + ".ffn.up_proj.weight"] = layer.ffn_up_proj;
+            model.tensors["transformer.blocks." + std::to_string(i) + ".ffn.down_proj.weight"] = layer.ffn_down_proj;
+        }
+    }
+
+    // key + value memory
+    {
+        const auto & hparams = model.hparams;
+
+        const int n_embd = hparams.d_model;
+        const int n_layer = hparams.n_layers;
+        const int n_ctx = hparams.max_seq_len;
+
+        const int64_t n_mem = n_layer * n_ctx;
+        const int64_t n_elements = n_embd * n_mem;
+
+        model.memory_k = ggml_new_tensor_1d(ctx, GGML_TYPE_F16, n_elements);
+        model.memory_v = ggml_new_tensor_1d(ctx, GGML_TYPE_F16, n_elements);
+
+        const size_t memory_size = ggml_nbytes(model.memory_k) + ggml_nbytes(model.memory_v);
+
+        printf("%s: memory_size = %8.2f MB, n_mem = %" PRIu64 "\n", __func__, memory_size / 1024.0 / 1024.0, n_mem);
+    }
+
+    // load weights
+    {
+        int n_tensors = 0;
+        size_t total_size = 0;
+
+        printf("%s: ", __func__);
+
+        while (true) {
+            int32_t n_dims;
+            int32_t length;
+            int32_t ttype;
+
+            fin.read(reinterpret_cast<char *>(&n_dims), sizeof(n_dims));
+            fin.read(reinterpret_cast<char *>(&length), sizeof(length));
+            fin.read(reinterpret_cast<char *>(&ttype), sizeof(ttype));
+
+            if (fin.eof()) {
+                break;
+            }
+
+            int32_t nelements = 1;
+            int32_t ne[2] = {1, 1};
+            for (int i = 0; i < n_dims; ++i) {
+                fin.read(reinterpret_cast<char *>(&ne[i]), sizeof(ne[i]));
+                nelements *= ne[i];
+            }
+
+            std::string name(length, 0);
+            fin.read(&name[0], length);
+
+            if (model.tensors.find(name) == model.tensors.end()) {
+                fprintf(stderr, "%s: unknown tensor '%s' in model file\n", __func__, name.c_str());
+                return false;
+            }
+
+            auto tensor = model.tensors[name];
+            if (ggml_nelements(tensor) != nelements) {
+                fprintf(stderr, "%s: tensor '%s' has wrong size in model file\n", __func__, name.c_str());
+                return false;
+            }
+
+            if (tensor->ne[0] != ne[0] || tensor->ne[1] != ne[1]) {
+                fprintf(stderr,
+                        "%s: tensor '%s' has wrong shape in model file: got [%5d, "
+                        "%5d], expected [%5d, %5d]\n",
+                        __func__, name.c_str(), (int)tensor->ne[0], (int)tensor->ne[1], ne[0], ne[1]);
+                return false;
+            }
+
+            // for debugging
+            if (0) {
+                printf("%24s - [%5d, %5d], type = %6s, %6.2f MB, %9zu bytes\n", name.c_str(), ne[0], ne[1],
+                       ggml_type_name(ggml_type(ttype)), ggml_nbytes(tensor) / 1024.0 / 1024.0, ggml_nbytes(tensor));
+            }
+
+            const size_t bpe = ggml_type_size(ggml_type(ttype));
+
+            if ((nelements * bpe) / ggml_blck_size(tensor->type) != ggml_nbytes(tensor)) {
+                fprintf(stderr,
+                        "%s: tensor '%s' has wrong size in model file: got %zu, "
+                        "expected %zu\n",
+                        __func__, name.c_str(), ggml_nbytes(tensor), nelements * bpe);
+                return false;
+            }
+
+            fin.read(reinterpret_cast<char *>(tensor->data), ggml_nbytes(tensor));
+
+            total_size += ggml_nbytes(tensor);
+            if (++n_tensors % 8 == 0) {
+                printf(".");
+                fflush(stdout);
+            }
+        }
+
+        printf(" done\n");
+
+        printf("%s: model size = %8.2f MB / num tensors = %d\n", __func__, total_size / 1024.0 / 1024.0, n_tensors);
+    }
+
+    fin.close();
+
+    return true;
+}
+
+// evaluate the transformer
+//
+//   - model:     the model
+//   - n_threads: number of threads to use
+//   - n_past:    the context size so far
+//   - embd_inp:  the embeddings of the tokens in the context
+//   - embd_w:    the predicted logits for the next token
+//
+bool replit_eval(const replit_model & model, const int n_threads, const int n_past,
+                 const std::vector<gpt_vocab::id> & embd_inp, std::vector<float> & embd_w, bool logits_all,
+                 size_t & mem_per_token) {
+    const int N = embd_inp.size();
+
+    const auto & hparams = model.hparams;
+
+    const int n_embd = hparams.d_model;
+    const int n_layer = hparams.n_layers;
+    const int n_head = hparams.n_heads;
+    const int n_vocab = hparams.n_vocab;
+    const int n_ctx = hparams.max_seq_len;
+    const float eps = 1e-5f;
+
+    static size_t buf_size = 256u * 1024 * 1024;
+    static void * buf = malloc(buf_size);
+
+    if (mem_per_token > 0 && mem_per_token * N > buf_size) {
+        const size_t buf_size_new = 1.1 * (mem_per_token * N); // add 10% to account for ggml object overhead
+        // printf("\n%s: reallocating buffer from %zu to %zu bytes\n", __func__,
+        // buf_size, buf_size_new);
+
+        // reallocate
+        buf_size = buf_size_new;
+        buf = realloc(buf, buf_size);
+        if (buf == nullptr) {
+            fprintf(stderr, "%s: failed to allocate %zu bytes\n", __func__, buf_size);
+            return false;
+        }
+    }
+
+    struct ggml_init_params params = {
+        /*.mem_size   =*/ buf_size,
+        /*.mem_buffer =*/ buf,
+        /*.no_alloc   =*/ false,
+    };
+
+    struct ggml_context * ctx0 = ggml_init(params);
+    struct ggml_cgraph gf = {};
+
+    struct ggml_tensor * embd = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
+    memcpy(embd->data, embd_inp.data(), N * ggml_element_size(embd));
+
+    struct ggml_tensor * inpL = ggml_get_rows(ctx0, model.wte_weight, embd);
+
+    for (int il = 0; il < n_layer; ++il) {
+
+        struct ggml_tensor * cur;
+
+        // a = self.ln_1(x)
+        {
+            cur = ggml_norm(ctx0, inpL, eps);
+
+            cur = ggml_mul(ctx0, ggml_repeat(ctx0, model.layers[il].norm_1_weight, cur), cur);
+        }
+
+        // self-attention
+        //  b, _, past_key_value = self.attn(a, past_key_value=past_key_value,
+        //  attn_bias=attn_bias, attention_mask=attention_mask,
+        //  is_causal=is_causal)
+        {
+            // compute QKV
+            cur = ggml_mul_mat(ctx0, model.layers[il].c_attn_wqkv_weight, cur);
+
+            struct ggml_tensor * Qcur = ggml_view_2d(ctx0, cur, n_embd, N, cur->nb[1], 0 * sizeof(float) * n_embd);
+            struct ggml_tensor * Kcur = ggml_view_2d(ctx0, cur, n_embd, N, cur->nb[1], 1 * sizeof(float) * n_embd);
+            struct ggml_tensor * Vcur = ggml_view_2d(ctx0, cur, n_embd, N, cur->nb[1], 2 * sizeof(float) * n_embd);
+
+            // store key and value to memory
+            {
+                struct ggml_tensor * k =
+                    ggml_view_1d(ctx0, model.memory_k, N * n_embd,
+                                 (ggml_element_size(model.memory_k) * n_embd) * (il * n_ctx + n_past));
+                struct ggml_tensor * v =
+                    ggml_view_1d(ctx0, model.memory_v, N * n_embd,
+                                 (ggml_element_size(model.memory_v) * n_embd) * (il * n_ctx + n_past));
+
+                ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Kcur, k));
+                ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Vcur, v));
+            }
+
+            // Q = Qcur.contiguous().view(n_embd/n_head, n_head, N).permute(0,
+            // 2, 1, 3) [64, N, 12]
+            struct ggml_tensor * Q = ggml_permute(
+                ctx0, ggml_cpy(ctx0, Qcur, ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_embd / n_head, n_head, N)), 0, 2,
+                1, 3);
+
+            // K = Kmem.view(n_embd/n_head, n_head, n_past + N).permute(0, 2, 1,
+            // 3) [64, n_past + N, 12]
+            struct ggml_tensor * K =
+                ggml_permute(ctx0,
+                             ggml_reshape_3d(ctx0,
+                                             ggml_view_1d(ctx0, model.memory_k, (n_past + N) * n_embd,
+                                                          il * n_ctx * ggml_element_size(model.memory_k) * n_embd),
+                                             n_embd / n_head, n_head, n_past + N),
+                             0, 2, 1, 3);
+            // K * Q
+            struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);
+
+            // KQ_scaled = KQ / sqrt(n_embd/n_head)
+            struct ggml_tensor * KQ_scaled =
+                ggml_scale(ctx0, KQ, ggml_new_f32(ctx0, 1.0f / sqrt(float(n_embd) / n_head)));
+
+            struct ggml_tensor * KQ_scaled_alibi = ggml_alibi(ctx0, KQ_scaled, n_past, n_head, 8.0f);
+
+            // KQ_masked = mask_past(KQ_scaled)
+            struct ggml_tensor * KQ_masked = ggml_diag_mask_inf(ctx0, KQ_scaled_alibi, n_past);
+
+            // KQ = soft_max(KQ_masked)
+            struct ggml_tensor * KQ_soft_max = ggml_soft_max(ctx0, KQ_masked);
+
+            // V_trans = Vmem.view(n_embd/n_head, n_head, n_past + N).permute(1,
+            // 2, 0, 3).contiguous() [n_past + N, 64, 12]
+            struct ggml_tensor * V_trans = ggml_cpy(
+                ctx0,
+                ggml_permute(ctx0,
+                             ggml_reshape_3d(ctx0,
+                                             ggml_view_1d(ctx0, model.memory_v, (n_past + N) * n_embd,
+                                                          il * n_ctx * ggml_element_size(model.memory_v) * n_embd),
+                                             n_embd / n_head, n_head, n_past + N),
+                             1, 2, 0, 3),
+                ggml_new_tensor_3d(ctx0, model.memory_v->type, n_past + N, n_embd / n_head, n_head));
+
+            // KQV = transpose(V) * KQ_soft_max
+            struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V_trans, KQ_soft_max);
+
+            // KQV_merged = KQV.permute(0, 2, 1, 3)
+            struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
+
+            // cur = KQV_merged.contiguous().view(n_embd, N)
+            cur = ggml_cpy(ctx0, KQV_merged, ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, N));
+
+            // projection
+            { cur = ggml_mul_mat(ctx0, model.layers[il].c_attn_out_proj_weight, cur); }
+        }
+
+        inpL = ggml_add(ctx0, inpL, cur);
+
+        // m = self.ln_2(x)
+        {
+            cur = ggml_norm(ctx0, inpL, eps);
+
+            cur = ggml_mul(ctx0, ggml_repeat(ctx0, model.layers[il].norm_2_weight, cur), cur);
+        }
+
+        // n = self.mlp(m)
+        {
+
+            cur = ggml_mul_mat(ctx0, model.layers[il].ffn_up_proj, cur);
+
+            // GELU activation
+            cur = ggml_gelu(ctx0, cur);
+
+            // projection
+            // cur = proj_w*cur + proj_b
+            cur = ggml_mul_mat(ctx0, model.layers[il].ffn_down_proj, cur);
+        }
+
+        // x = x + n
+        inpL = ggml_add(ctx0, inpL, cur);
+    }
+
+    // norm
+    {
+        inpL = ggml_norm(ctx0, inpL, eps);
+        // inpL = ln_f_g*inpL
+        inpL = ggml_mul(ctx0, ggml_repeat(ctx0, model.norm_f_weight, inpL), inpL);
+    }
+
+    // output embedding weight tied to input embedding
+    inpL = ggml_mul_mat(ctx0, model.wte_weight, inpL);
+
+    // logits -> probs
+    // inpL = ggml_soft_max(ctx0, inpL);
+
+    // run the computation
+    ggml_build_forward_expand(&gf, inpL);
+    ggml_graph_compute_with_ctx(ctx0, &gf, n_threads);
+
+    // std::cout << "Qcur" << std::endl;
+    // print_tensor(Qcur);
+
+    // if (n_past%100 == 0) {
+    // ggml_graph_print(&gf);
+    // ggml_graph_dump_dot(&gf, NULL, "mpt-model.dot");
+    // }
+
+    if (logits_all) {
+        // return result for all tokens
+        embd_w.resize(n_vocab * N);
+        memcpy(embd_w.data(), (float *)ggml_get_data(inpL), sizeof(float) * n_vocab * N);
+    } else {
+        // return result for just the last token
+        embd_w.resize(n_vocab);
+        memcpy(embd_w.data(), (float *)ggml_get_data(inpL) + (n_vocab * (N - 1)), sizeof(float) * n_vocab);
+    }
+
+    if (mem_per_token == 0) {
+        mem_per_token = ggml_used_mem(ctx0) / N;
+    }
+    // printf("used_mem = %zu\n", ggml_used_mem(ctx0));
+
+    ggml_free(ctx0);
+
+    return true;
+}
+
+int main(int argc, char ** argv) {
+    const int64_t t_main_start_us = ggml_time_us();
+
+    gpt_params params;
+    params.model = "";
+
+    if (gpt_params_parse(argc, argv, params) == false) {
+        return 1;
+    }
+
+    if (params.seed < 0) {
+        params.seed = time(NULL);
+    }
+
+    printf("%s: seed = %d\n", __func__, params.seed);
+
+    std::mt19937 rng(params.seed);
+    if (params.prompt.empty()) {
+        if (!is_stdin_terminal()) {
+            std::string line;
+            while (std::getline(std::cin, line)) {
+                params.prompt = params.prompt + "\n" + line;
+            }
+        } else {
+            params.prompt = gpt_random_prompt(rng);
+        }
+    }
+
+    int64_t t_load_us = 0;
+
+    replit_tokenizer vocab;
+    replit_model model;
+
+    // load the model
+    {
+        const int64_t t_start_us = ggml_time_us();
+
+        if (!replit_model_load(params.model, model, vocab)) {
+            fprintf(stderr, "%s: failed to load model from '%s'\n", __func__, params.model.c_str());
+            return 1;
+        }
+
+        t_load_us = ggml_time_us() - t_start_us;
+    }
+
+    int n_past = 0;
+
+    int64_t t_sample_us = 0;
+    int64_t t_predict_us = 0;
+
+    std::vector<float> logits;
+
+    // tokenize the prompt
+    std::vector<std::size_t> embd_inp = replit_tokenizer_tokenize(vocab, params.prompt);
+
+    printf("%s: number of tokens in prompt = %zu\n", __func__, embd_inp.size());
+
+    for (size_t i = 0; i < embd_inp.size(); i++) {
+        printf("%s: token[%zu] = %6zu\n", __func__, i, embd_inp[i]);
+        // vocab.id_to_token.at(embd_inp[i]).c_str()
+    }
+    printf("\n");
+
+    params.n_predict = std::min(params.n_predict, model.hparams.max_seq_len - (int)embd_inp.size());
+
+    std::vector<gpt_vocab::id> embd;
+
+    // determine the required inference memory per token:
+    size_t mem_per_token = 0;
+    replit_eval(model, params.n_threads, 0, {0, 1, 2, 3}, logits, false, mem_per_token);
+
+    for (size_t i = embd.size(); i < embd_inp.size() + params.n_predict; i++) {
+        // predict
+        if (embd.size() > 0) {
+            const int64_t t_start_us = ggml_time_us();
+
+            if (!replit_eval(model, params.n_threads, n_past, embd, logits, false, mem_per_token)) {
+                printf("Failed to predict\n");
+                return 1;
+            }
+
+            t_predict_us += ggml_time_us() - t_start_us;
+        }
+
+        n_past += embd.size();
+        embd.clear();
+
+        if (i >= embd_inp.size()) {
+            // sample next token
+            const int top_k = params.top_k;
+            const float top_p = params.top_p;
+            const float temp = params.temp;
+
+            const int n_vocab = model.hparams.n_vocab;
+
+            gpt_vocab::id id = 0;
+
+            {
+                const int64_t t_start_sample_us = ggml_time_us();
+
+                id = gpt_sample_top_k_top_p(vocab.raw_vocab, logits.data() + (logits.size() - n_vocab), top_k, top_p,
+                                            temp, rng);
+
+                t_sample_us += ggml_time_us() - t_start_sample_us;
+            }
+
+            // add it to the context
+            embd.push_back(id);
+        } else {
+            // if here, it means we are still processing the input prompt
+            for (size_t k = i; k < embd_inp.size(); k++) {
+                embd.push_back(embd_inp[k]);
+                if (int32_t(embd.size()) > params.n_batch) {
+                    break;
+                }
+            }
+            i += embd.size() - 1;
+        }
+
+        // display text
+        for (auto id : embd) {
+            printf("%s", replit_tokenizer_detokenize(vocab, {static_cast<std::size_t>(id)}).c_str());
+        }
+        fflush(stdout);
+
+        // end of text token
+        if (embd.back() == 0) {
+            break;
+        }
+    }
+
+    // report timing
+    {
+        const int64_t t_main_end_us = ggml_time_us();
+
+        printf("\n\n");
+        printf("%s: mem per token = %8zu bytes\n", __func__, mem_per_token);
+        printf("%s:     load time = %8.2f ms\n", __func__, t_load_us / 1000.0f);
+        printf("%s:   sample time = %8.2f ms\n", __func__, t_sample_us / 1000.0f);
+        printf("%s:  predict time = %8.2f ms / %.2f ms per token\n", __func__, t_predict_us / 1000.0f,
+               t_predict_us / 1000.0f / n_past);
+        printf("%s:    total time = %8.2f ms\n", __func__, (t_main_end_us - t_main_start_us) / 1000.0f);
+    }
+
+    ggml_free(model.ctx);
+
+    return 0;
+}

ggml/examples/replit/quantize.cpp

@@ -0,0 +1,182 @@
+#include "ggml/ggml.h"
+
+#include "common-ggml.h"
+#include "common.h"
+
+#include <cassert>
+#include <cmath>
+#include <cstdio>
+#include <cstring>
+#include <fstream>
+#include <map>
+#include <regex>
+#include <string>
+#include <vector>
+
+struct mpt_hparams {
+    int32_t d_model     = 0;
+    int32_t max_seq_len = 0;
+    int32_t n_heads     = 0;
+    int32_t n_layers    = 0;
+    int32_t n_vocab     = 0;
+    int32_t ftype       = 0;
+};
+
+// quantize a model
+bool mpt_model_quantize(const std::string & fname_inp,
+                        const std::string & fname_out, ggml_ftype ftype) {
+
+    printf("%s: loading model from '%s'\n", __func__, fname_inp.c_str());
+
+    auto finp = std::ifstream(fname_inp, std::ios::binary);
+    if (!finp) {
+        fprintf(stderr, "%s: failed to open '%s' for reading\n", __func__,
+                fname_inp.c_str());
+        return false;
+    }
+
+    auto fout = std::ofstream(fname_out, std::ios::binary);
+    if (!fout) {
+        fprintf(stderr, "%s: failed to open '%s' for writing\n", __func__,
+                fname_out.c_str());
+        return false;
+    }
+
+    // verify magic
+    {
+        uint32_t magic;
+        finp.read((char *)&magic, sizeof(magic));
+        if (magic != GGML_FILE_MAGIC) {
+            fprintf(stderr, "%s: invalid model file '%s' (bad magic)\n",
+                    __func__, fname_inp.c_str());
+            return false;
+        }
+
+        fout.write((char *)&magic, sizeof(magic));
+    }
+
+    mpt_hparams hparams;
+
+    // load hparams
+    {
+        finp.read((char *) &hparams.d_model,     sizeof(hparams.d_model));
+        finp.read((char *) &hparams.max_seq_len, sizeof(hparams.max_seq_len));
+        finp.read((char *) &hparams.n_heads,     sizeof(hparams.n_heads));
+        finp.read((char *) &hparams.n_layers,    sizeof(hparams.n_layers));
+        finp.read((char *) &hparams.n_vocab,     sizeof(hparams.n_vocab));
+        finp.read((char *) &hparams.ftype,       sizeof(hparams.ftype));
+
+        const int32_t qntvr_src =    hparams.ftype / GGML_QNT_VERSION_FACTOR;
+        const int32_t ftype_dst = GGML_QNT_VERSION * GGML_QNT_VERSION_FACTOR + ftype;
+
+        printf("%s: d_model     = %d\n", __func__, hparams.d_model);
+        printf("%s: max_seq_len = %d\n", __func__, hparams.max_seq_len);
+        printf("%s: n_heads     = %d\n", __func__, hparams.n_heads);
+        printf("%s: n_layers    = %d\n", __func__, hparams.n_layers);
+        printf("%s: n_vocab     = %d\n", __func__, hparams.n_vocab);
+        printf("%s: ftype (src) = %d\n", __func__, hparams.ftype);
+        printf("%s: qntvr (src) = %d\n", __func__, qntvr_src);
+        printf("%s: ftype (dst) = %d\n", __func__, ftype_dst);
+        printf("%s: qntvr (dst) = %d\n", __func__, GGML_QNT_VERSION);
+
+        fout.write((char *) &hparams.d_model,     sizeof(hparams.d_model));
+        fout.write((char *) &hparams.max_seq_len, sizeof(hparams.max_seq_len));
+        fout.write((char *) &hparams.n_heads,     sizeof(hparams.n_heads));
+        fout.write((char *) &hparams.n_layers,    sizeof(hparams.n_layers));
+        fout.write((char *) &hparams.n_vocab,     sizeof(hparams.n_vocab));
+        fout.write((char *) &ftype_dst,           sizeof(ftype_dst));
+    }
+
+    // load vocab
+    {
+        const int32_t n_vocab = hparams.n_vocab;
+
+        std::string word;
+        for (int i = 0; i < n_vocab; i++) {
+            uint32_t len;
+            finp.read((char *)&len, sizeof(len));
+            fout.write((char *)&len, sizeof(len));
+
+            word.resize(len);
+            finp.read((char *)word.data(), len);
+            fout.write((char *)word.data(), len);
+
+            float prob;
+            finp.read((char *)&prob, sizeof(prob));
+            fout.write((char *)&prob, sizeof(prob));
+        }
+    }
+
+    printf("%s: quantizing tensors\n", __func__);
+
+    // regexes of tensor names to be quantized
+    const std::vector<std::string> to_quant = {
+        ".*weight",
+    };
+
+    if (!ggml_common_quantize_0(finp, fout, ftype, to_quant, {})) {
+        fprintf(stderr, "%s: failed to quantize model '%s'\n", __func__,
+                fname_inp.c_str());
+        return false;
+    }
+
+    finp.close();
+    fout.close();
+
+    return true;
+}
+
+// usage:
+//  ./replit-quantize models/replit/ggml-model.bin
+//  models/replit/ggml-model-quant.bin type
+//
+int main(int argc, char ** argv) {
+    if (argc != 4) {
+        fprintf(stderr, "usage: %s model-f32.bin model-quant.bin type\n",
+                argv[0]);
+        ggml_print_ftypes(stderr);
+        return 1;
+    }
+
+    // needed to initialize f16 tables
+    {
+        struct ggml_init_params params = {0, NULL, false};
+        struct ggml_context * ctx = ggml_init(params);
+        ggml_free(ctx);
+    }
+
+    const std::string fname_inp = argv[1];
+    const std::string fname_out = argv[2];
+
+    const ggml_ftype ftype = ggml_parse_ftype(argv[3]);
+
+    const int64_t t_main_start_us = ggml_time_us();
+
+    int64_t t_quantize_us = 0;
+
+    // load the model
+    {
+        const int64_t t_start_us = ggml_time_us();
+
+        if (!mpt_model_quantize(fname_inp, fname_out, ggml_ftype(ftype))) {
+            fprintf(stderr, "%s: failed to quantize model from '%s'\n",
+                    __func__, fname_inp.c_str());
+            return 1;
+        }
+
+        t_quantize_us = ggml_time_us() - t_start_us;
+    }
+
+    // report timing
+    {
+        const int64_t t_main_end_us = ggml_time_us();
+
+        printf("\n");
+        printf("%s: quantize time = %8.2f ms\n", __func__,
+               t_quantize_us / 1000.0f);
+        printf("%s:    total time = %8.2f ms\n", __func__,
+               (t_main_end_us - t_main_start_us) / 1000.0f);
+    }
+
+    return 0;
+}

ggml/examples/sam/CMakeLists.txt

@@ -0,0 +1,13 @@
+#
+# sam
+
+set(TEST_TARGET sam)
+add_executable(${TEST_TARGET} main.cpp)
+target_link_libraries(${TEST_TARGET} PRIVATE ggml common)
+
+#
+# sam-quantize
+
+#set(TEST_TARGET sam-quantize)
+#add_executable(${TEST_TARGET} quantize.cpp)
+#target_link_libraries(${TEST_TARGET} PRIVATE ggml common)

ggml/examples/sam/README.md

@@ -0,0 +1,105 @@
+# SAM.cpp
+
+Inference of Meta's [Segment Anything Model](https://github.com/facebookresearch/segment-anything/) in pure C/C++
+
+## Description
+
+The example currently supports only the [ViT-B SAM model checkpoint](https://huggingface.co/facebook/sam-vit-base).
+
+## Next steps
+
+- [X] Reduce memory usage by utilizing the new ggml-alloc
+- [X] Remove redundant graph nodes
+- [ ] Make inference faster
+- [X] Fix the difference in output masks compared to the PyTorch implementation
+- [X] Filter masks based on stability score
+- [ ] Add support for user input
+- [ ] Support F16 for heavy F32 ops
+- [ ] Test quantization
+- [X] Support bigger model checkpoints
+- [ ] GPU support
+
+## Quick start
+```bash
+git clone https://github.com/ggerganov/ggml
+cd ggml
+
+# Install Python dependencies
+python3 -m pip install -r requirements.txt
+
+# Convert PTH model to ggml
+python convert-pth-to-ggml.py examples/sam/sam_vit_b_01ec64.pth . 1
+
+# Build ggml + examples
+mkdir build && cd build
+cmake .. && make -j4
+
+# run inference
+./bin/sam -t 16 -i ../img.jpg -m examples/sam/ggml-model-f16.bin
+```
+
+## Downloading and converting the model checkpoints
+
+You can download a [model checkpoint](https://github.com/facebookresearch/segment-anything/tree/main#model-checkpoints) and convert it to `ggml` format using the script `convert-pth-to-ggml.py`:
+
+```
+# Convert PTH model to ggml
+python convert-pth-to-ggml.py examples/sam/sam_vit_b_01ec64.pth . 1
+```
+
+## Example output on M2 Ultra
+```
+ $ ▶ make -j sam && time ./bin/sam -t 8 -i img.jpg
+[ 28%] Built target common
+[ 71%] Built target ggml
+[100%] Built target sam
+main: seed = 1693224265
+main: loaded image 'img.jpg' (680 x 453)
+sam_image_preprocess: scale = 0.664062
+main: preprocessed image (1024 x 1024)
+sam_model_load: loading model from 'models/sam-vit-b/ggml-model-f16.bin' - please wait ...
+sam_model_load: n_enc_state      = 768
+sam_model_load: n_enc_layer      = 12
+sam_model_load: n_enc_head       = 12
+sam_model_load: n_enc_out_chans  = 256
+sam_model_load: n_pt_embd        = 4
+sam_model_load: ftype            = 1
+sam_model_load: qntvr            = 0
+operator(): ggml ctx size = 202.32 MB
+sam_model_load: ...................................... done
+sam_model_load: model size =   185.05 MB / num tensors = 304
+embd_img
+dims: 64 64 256 1 f32
+First & Last 10 elements:
+-0.05117 -0.06408 -0.07154 -0.06991 -0.07212 -0.07690 -0.07508 -0.07281 -0.07383 -0.06779
+0.01589 0.01775 0.02250 0.01675 0.01766 0.01661 0.01811 0.02051 0.02103 0.03382
+sum:  12736.272313
+
+Skipping mask 0 with iou 0.705935 below threshold 0.880000
+Skipping mask 1 with iou 0.762136 below threshold 0.880000
+Mask 2: iou = 0.947081, stability_score = 0.955437, bbox (371, 436), (144, 168)
+
+
+main:     load time =    51.28 ms
+main:    total time =  2047.49 ms
+
+real	0m2.068s
+user	0m16.343s
+sys	0m0.214s
+```
+
+Input point is (414.375, 162.796875) (currently hardcoded)
+
+Input image:
+
+![llamas](https://user-images.githubusercontent.com/8558655/261301565-37b7bf4b-bf91-40cf-8ec1-1532316e1612.jpg)
+
+Output mask (mask_out_2.png in build folder):
+
+![mask_glasses](https://user-images.githubusercontent.com/8558655/263706800-47eeea30-1457-4c87-938b-8f11536c5aa7.png)
+
+## References
+
+- [ggml](https://github.com/ggerganov/ggml)
+- [SAM](https://segment-anything.com/)
+- [SAM demo](https://segment-anything.com/demo)

ggml/examples/sam/convert-pth-to-ggml.py

@@ -0,0 +1,147 @@
+# Convert a SAM model checkpoint to a ggml compatible file
+#
+
+import sys
+import torch
+import struct
+import numpy as np
+
+if len(sys.argv) < 3:
+    print("Usage: convert-pth-to-ggml.py file-model dir-output [ftype]\n")
+    print("  ftype == 0 -> float32")
+    print("  ftype == 1 -> float16")
+    sys.exit(1)
+
+# output in the same directory as the model
+fname_model = sys.argv[1]
+dir_out     = sys.argv[2]
+fname_out   = dir_out + "/ggml-model.bin"
+
+# possible data types
+#   ftype == 0 -> float32
+#   ftype == 1 -> float16
+#
+# map from ftype to string
+ftype_str = ["f32", "f16"]
+
+ftype = 1
+if len(sys.argv) > 3:
+    ftype = int(sys.argv[3])
+
+if ftype < 0 or ftype > 1:
+    print("Invalid ftype: " + str(ftype))
+    sys.exit(1)
+
+fname_out = fname_out.replace(".bin", "-" + ftype_str[ftype] + ".bin")
+
+# Default params are set to sam_vit_b checkpoint
+n_enc_state = 768
+n_enc_layers = 12
+n_enc_heads = 12
+n_enc_out_chans = 256
+n_pt_embd = 4
+
+model = torch.load(fname_model, map_location="cpu")
+for k, v in model.items():
+    print(k, v.shape)
+    if k == "image_encoder.blocks.0.norm1.weight":
+        n_enc_state = v.shape[0]
+
+if n_enc_state == 1024: # sam_vit_l
+    n_enc_layers = 24
+    n_enc_heads  = 16
+elif n_enc_state == 1280: # sam_vit_h
+    n_enc_layers = 32
+    n_enc_heads  = 16
+
+hparams = {
+    "n_enc_state":      n_enc_state,
+    "n_enc_layers":     n_enc_layers,
+    "n_enc_heads":      n_enc_heads,
+    "n_enc_out_chans":  n_enc_out_chans,
+    "n_pt_embd":        n_pt_embd,
+}
+
+print(hparams)
+
+for k, v in model.items():
+    print(k, v.shape)
+
+#exit()
+#code.interact(local=locals())
+
+fout = open(fname_out, "wb")
+
+fout.write(struct.pack("i", 0x67676d6c)) # magic: ggml in hex
+fout.write(struct.pack("i", hparams["n_enc_state"]))
+fout.write(struct.pack("i", hparams["n_enc_layers"]))
+fout.write(struct.pack("i", hparams["n_enc_heads"]))
+fout.write(struct.pack("i", hparams["n_enc_out_chans"]))
+fout.write(struct.pack("i", hparams["n_pt_embd"]))
+fout.write(struct.pack("i", ftype))
+
+for k, v in model.items():
+    name = k
+    shape = v.shape
+
+    if name[:19] == "prompt_encoder.mask":
+        continue
+
+    print("Processing variable: " + name + " with shape: ", shape, " and type: ", v.dtype)
+
+    #data = tf.train.load_variable(dir_model, name).squeeze()
+    #data = v.numpy().squeeze()
+    data = v.numpy()
+    n_dims = len(data.shape)
+
+    # for efficiency - transpose some matrices
+    # "model/h.*/attn/c_attn/w"
+    # "model/h.*/attn/c_proj/w"
+    # "model/h.*/mlp/c_fc/w"
+    # "model/h.*/mlp/c_proj/w"
+    #if name[-14:] == "/attn/c_attn/w" or \
+    #   name[-14:] == "/attn/c_proj/w" or \
+    #   name[-11:] == "/mlp/c_fc/w" or \
+    #   name[-13:] == "/mlp/c_proj/w":
+    #    print("  Transposing")
+    #    data = data.transpose()
+
+    dshape = data.shape
+
+    # default type is fp16
+    ftype_cur = 1
+    if ftype == 0 or n_dims == 1 or \
+            name == "image_encoder.pos_embed" or \
+            name.startswith("prompt_encoder") or \
+            name.startswith("mask_decoder.iou_token") or \
+            name.startswith("mask_decoder.mask_tokens"):
+        print("  Converting to float32")
+        data = data.astype(np.float32)
+        ftype_cur = 0
+    else:
+        print("  Converting to float16")
+        data = data.astype(np.float16)
+
+    # reshape the 1D bias into a 4D tensor so we can use ggml_repeat
+    # keep it in F32 since the data is small
+    if name == "image_encoder.patch_embed.proj.bias":
+        data = data.reshape(1, data.shape[0], 1, 1)
+        n_dims = len(data.shape)
+        dshape = data.shape
+
+    print("  New shape: ", dshape)
+
+    # header
+    str = name.encode('utf-8')
+    fout.write(struct.pack("iii", n_dims, len(str), ftype_cur))
+    for i in range(n_dims):
+        fout.write(struct.pack("i", dshape[n_dims - 1 - i]))
+    fout.write(str)
+
+    # data
+    data.tofile(fout)
+
+fout.close()
+
+print("Done. Output file: " + fname_out)
+print("")

ggml/examples/sam/main.cpp

@@ -0,0 +1,2201 @@
+#define _USE_MATH_DEFINES // for M_PI
+#define _CRT_SECURE_NO_DEPRECATE // Disables ridiculous "unsafe" warnigns on Windows
+
+#include "ggml.h"
+#include "ggml-alloc.h"
+#define STB_IMAGE_IMPLEMENTATION
+#include "stb_image.h"
+#define STB_IMAGE_WRITE_IMPLEMENTATION
+#include "stb_image_write.h"
+
+#include <cassert>
+#include <cmath>
+#include <cstddef>
+#include <cstdio>
+#include <cstring>
+#include <fstream>
+#include <map>
+#include <string>
+#include <vector>
+#include <thread>
+
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
+// default hparams (ViT-B SAM)
+struct sam_hparams {
+    int32_t n_enc_state               = 768;
+    int32_t n_enc_layer               = 12;
+    int32_t n_enc_head                = 12;
+    int32_t n_enc_out_chans           = 256;
+    int32_t n_pt_embd                 = 4;
+    int32_t n_dec_heads               = 8;
+    int32_t ftype                     = 1;
+    float   mask_threshold            = 0.f;
+    float   iou_threshold             = 0.88f;
+    float   stability_score_threshold = 0.95f;
+    float   stability_score_offset    = 1.0f;
+    float   eps                       = 1e-6f;
+    float   eps_decoder_transformer   = 1e-5f;
+
+    int32_t n_enc_head_dim() const { return n_enc_state / n_enc_head; }
+    int32_t n_img_size()     const { return 1024; }
+    int32_t n_window_size()  const { return 14; }
+    int32_t n_patch_size()   const { return 16; }
+    int32_t n_img_embd()     const { return n_img_size() / n_patch_size(); }
+
+    std::vector<int32_t> global_attn_indices() const {
+        switch (n_enc_state) {
+            case  768: return {  2,  5,  8, 11 };
+            case 1024: return {  5, 11, 17, 23 };
+            case 1280: return {  7, 15, 23, 31 };
+            default:
+                {
+                    fprintf(stderr, "%s: unsupported n_enc_state = %d\n", __func__, n_enc_state);
+                } break;
+        };
+
+        return {};
+    }
+
+    bool is_global_attn(int32_t layer) const {
+        const auto indices = global_attn_indices();
+
+        for (const auto & idx : indices) {
+            if (layer == idx) {
+                return true;
+            }
+        }
+
+        return false;
+    }
+};
+
+struct sam_layer_enc {
+    struct ggml_tensor * norm1_w;
+    struct ggml_tensor * norm1_b;
+
+    struct ggml_tensor * rel_pos_w;
+    struct ggml_tensor * rel_pos_h;
+
+    struct ggml_tensor * qkv_w;
+    struct ggml_tensor * qkv_b;
+
+    struct ggml_tensor * proj_w;
+    struct ggml_tensor * proj_b;
+
+    struct ggml_tensor * norm2_w;
+    struct ggml_tensor * norm2_b;
+
+    struct ggml_tensor * mlp_lin1_w;
+    struct ggml_tensor * mlp_lin1_b;
+
+    struct ggml_tensor * mlp_lin2_w;
+    struct ggml_tensor * mlp_lin2_b;
+};
+
+struct sam_encoder_image {
+    struct ggml_tensor * pe;
+
+    struct ggml_tensor * proj_w;
+    struct ggml_tensor * proj_b;
+
+    struct ggml_tensor * neck_conv_0;
+    struct ggml_tensor * neck_norm_0_w;
+    struct ggml_tensor * neck_norm_0_b;
+    struct ggml_tensor * neck_conv_1;
+    struct ggml_tensor * neck_norm_1_w;
+    struct ggml_tensor * neck_norm_1_b;
+
+    std::vector<sam_layer_enc> layers;
+};
+
+struct sam_encoder_prompt {
+    struct ggml_tensor * pe;
+
+    struct ggml_tensor * not_a_pt_embd_w;
+    std::vector<struct ggml_tensor *> pt_embd;
+
+    struct ggml_tensor * no_mask_embd_w;
+    //std::vector<struct ggml_tensor *> mask_down_w;
+    //std::vector<struct ggml_tensor *> mask_down_b;
+};
+
+struct  sam_layer_dec_transformer_attn {
+    // q_proj
+    struct ggml_tensor * q_w;
+    struct ggml_tensor * q_b;
+
+    // k_proj
+    struct ggml_tensor * k_w;
+    struct ggml_tensor * k_b;
+
+    // v_proj
+    struct ggml_tensor * v_w;
+    struct ggml_tensor * v_b;
+
+    // out_proj
+    struct ggml_tensor * out_w;
+    struct ggml_tensor * out_b;
+};
+
+struct sam_layer_dec_transformer {
+    sam_layer_dec_transformer_attn self_attn;
+
+    // norm1
+    struct ggml_tensor * norm1_w;
+    struct ggml_tensor * norm1_b;
+
+    sam_layer_dec_transformer_attn cross_attn_token_to_img;
+
+    // norm2
+    struct ggml_tensor * norm2_w;
+    struct ggml_tensor * norm2_b;
+
+    // mlp.lin1
+    struct ggml_tensor * mlp_lin1_w;
+    struct ggml_tensor * mlp_lin1_b;
+
+    // mlp.lin2
+    struct ggml_tensor * mlp_lin2_w;
+    struct ggml_tensor * mlp_lin2_b;
+
+    // norm3
+    struct ggml_tensor * norm3_w;
+    struct ggml_tensor * norm3_b;
+
+    // norm4
+    struct ggml_tensor * norm4_w;
+    struct ggml_tensor * norm4_b;
+
+    sam_layer_dec_transformer_attn cross_attn_img_to_token;
+};
+
+struct sam_layer_dec_output_hypernet_mlps {
+    // mlps_*.layers.0
+    struct ggml_tensor * w_0;
+    struct ggml_tensor * b_0;
+
+    // mlps_*.layers.1
+    struct ggml_tensor * w_1;
+    struct ggml_tensor * b_1;
+
+    // mlps_*.layers.2
+    struct ggml_tensor * w_2;
+    struct ggml_tensor * b_2;
+};
+
+struct sam_decoder_mask {
+    std::vector<sam_layer_dec_transformer> transformer_layers;
+
+    // trasnformer.final_attn_token_to_image
+    sam_layer_dec_transformer_attn transformer_final_attn_token_to_img;
+
+    // transformer.norm_final
+    struct ggml_tensor * transformer_norm_final_w;
+    struct ggml_tensor * transformer_norm_final_b;
+
+    // output_upscaling.0
+    struct ggml_tensor * output_upscaling_0_w;
+    struct ggml_tensor * output_upscaling_0_b;
+
+    // output_upscaling.1
+    struct ggml_tensor * output_upscaling_1_w;
+    struct ggml_tensor * output_upscaling_1_b;
+
+    // output_upscaling.3
+    struct ggml_tensor * output_upscaling_3_w;
+    struct ggml_tensor * output_upscaling_3_b;
+
+    // output_hypernetworks_mlps
+    std::vector<sam_layer_dec_output_hypernet_mlps> output_hypernet_mlps;
+
+    // iou_prediction_head.0
+    struct ggml_tensor * iou_prediction_head_0_w;
+    struct ggml_tensor * iou_prediction_head_0_b;
+
+    // iou_prediction_head.1
+    struct ggml_tensor * iou_prediction_head_1_w;
+    struct ggml_tensor * iou_prediction_head_1_b;
+
+    // iou_prediction_head.2
+    struct ggml_tensor * iou_prediction_head_2_w;
+    struct ggml_tensor * iou_prediction_head_2_b;
+
+    // iou_token.weight
+    struct ggml_tensor * iou_token_w;
+
+    // mask_tokens.weight
+    struct ggml_tensor * mask_tokens_w;
+};
+
+
+struct sam_state {
+    struct ggml_tensor * embd_img;
+
+    struct ggml_tensor * low_res_masks;
+    struct ggml_tensor * iou_predictions;
+
+    //struct ggml_tensor * tmp_save = {};
+
+    struct ggml_context * ctx;
+
+    // buffer for `ggml_graph_plan.work_data`
+    std::vector<uint8_t> work_buffer;
+    // buffers to evaluate the model
+    std::vector<uint8_t> buf_alloc_img_enc;
+    std::vector<uint8_t> buf_compute_img_enc;
+
+    std::vector<uint8_t> buf_alloc_fast;
+    std::vector<uint8_t> buf_compute_fast;
+
+    struct ggml_allocr  * allocr = {};
+};
+
+// void save_tensor(sam_state& state, struct ggml_tensor * t, struct ggml_cgraph * gf) {
+//     if (!state.tmp_save) {
+//         state.tmp_save = ggml_new_tensor(state.ctx, t->type, t->n_dims, t->ne);
+//     }
+//     struct ggml_tensor * tmp0 = ggml_cpy(state.ctx, t, state.tmp_save);
+//     ggml_build_forward_expand(gf, tmp0);
+// }
+
+struct sam_model {
+    sam_hparams hparams;
+
+    sam_encoder_image  enc_img;
+    sam_encoder_prompt enc_prompt;
+    sam_decoder_mask   dec;
+
+    //
+    struct ggml_context * ctx;
+    std::map<std::string, struct ggml_tensor *> tensors;
+};
+
+struct sam_point {
+    float x;
+    float y;
+};
+
+// RGB uint8 image
+struct sam_image_u8 {
+    int nx;
+    int ny;
+
+    std::vector<uint8_t> data;
+};
+
+// RGB float32 image
+// Memory layout: RGBRGBRGB...
+struct sam_image_f32 {
+    int nx;
+    int ny;
+
+    std::vector<float> data;
+};
+
+void print_t_f32(const char* title, struct ggml_tensor * t, int n = 10) {
+    printf("%s\n", title);
+    float * data = (float *)t->data;
+    printf("dims: %jd %jd %jd %jd f32\n", t->ne[0], t->ne[1], t->ne[2], t->ne[3]);
+    printf("First & Last %d elements:\n", n);
+    for (int i = 0; i < std::min((int) (t->ne[0]*t->ne[1]), n); i++) {
+        printf("%.5f ", data[i]);
+        if (i != 0 && i % t->ne[0] == 0) {
+            printf("\n");
+        }
+    }
+    printf("\n");
+    for (int i = 0; i < std::min((int) (t->ne[0]*t->ne[1]), n); i++) {
+        printf("%.5f ", data[ggml_nelements(t) - n + i]);
+        if ((ggml_nelements(t) - n + i) % t->ne[0] == 0) {
+            printf("\n");
+        }
+    }
+    printf("\n");
+    double sum = 0.0;
+    for (int i = 0; i < ggml_nelements(t); i++) {
+        sum += data[i];
+    }
+    printf("sum:  %f\n\n", sum);
+}
+
+static void ggml_disconnect_node_from_graph(ggml_tensor * t) {
+    t->op = GGML_OP_NONE;
+    for (int i = 0; i < GGML_MAX_SRC; i++) {
+        t->src[i] = NULL;
+    }
+}
+
+static void ggml_graph_compute_helper(std::vector<uint8_t> & buf, ggml_cgraph * graph, int n_threads) {
+    struct ggml_cplan plan = ggml_graph_plan(graph, n_threads);
+
+    if (plan.work_size > 0) {
+        buf.resize(plan.work_size);
+        plan.work_data = buf.data();
+    }
+
+    ggml_graph_compute(graph, &plan);
+}
+
+static void ggml_sam_sin(struct ggml_tensor * dst , const struct ggml_tensor * src, int ith, int nth, void * userdata) {
+    GGML_ASSERT(userdata == NULL);
+    GGML_ASSERT(ggml_are_same_shape(dst, src));
+    GGML_ASSERT(ggml_is_contiguous(dst));
+    GGML_ASSERT(ggml_is_contiguous(src));
+
+    const float * src_data = ggml_get_data_f32(src);
+    float * dst_data = ggml_get_data_f32(dst);
+
+    const int ne = (int)ggml_nelements(dst);
+    const int dr = (ne + nth - 1) / nth;
+    const int ie0 = dr * ith;
+    const int ie1 = std::min(ie0 + dr, ne);
+
+    for (int i = ie0; i < ie1; ++i) {
+        dst_data[i] = sinf(src_data[i]);
+    }
+}
+
+static void ggml_sam_cos(struct ggml_tensor * dst , const struct ggml_tensor * src, int ith, int nth, void * userdata) {
+    GGML_ASSERT(userdata == NULL);
+    GGML_ASSERT(ggml_are_same_shape(dst, src));
+    GGML_ASSERT(ggml_is_contiguous(dst));
+    GGML_ASSERT(ggml_is_contiguous(src));
+
+    const float * src_data = ggml_get_data_f32(src);
+    float * dst_data = ggml_get_data_f32(dst);
+
+    const int ne = (int)ggml_nelements(dst);
+    const int dr = (ne + nth - 1) / nth;
+    const int ie0 = dr * ith;
+    const int ie1 = std::min(ie0 + dr, ne);
+
+    for (int i = ie0; i < ie1; ++i) {
+        dst_data[i] = cosf(src_data[i]);
+    }
+}
+
+bool sam_image_load_from_file(const std::string & fname, sam_image_u8 & img) {
+    int nx, ny, nc;
+    auto data = stbi_load(fname.c_str(), &nx, &ny, &nc, 3);
+    if (!data) {
+        fprintf(stderr, "%s: failed to load '%s'\n", __func__, fname.c_str());
+        return false;
+    }
+
+    img.nx = nx;
+    img.ny = ny;
+    img.data.resize(nx * ny * 3);
+    memcpy(img.data.data(), data, nx * ny * 3);
+
+    stbi_image_free(data);
+
+    return true;
+}
+
+// ref: https://github.com/facebookresearch/segment-anything/blob/efeab7296ab579d4a261e554eca80faf6b33924a/segment_anything/modeling/sam.py#L164
+// resize largest dimension to 1024
+// normalize: x = (x - mean) / std
+//     mean = [123.675, 116.28, 103.53]
+//     std  = [58.395, 57.12, 57.375]
+//     TODO: why are these hardcoded !?
+// pad to 1024x1024
+// TODO: for some reason, this is not numerically identical to pytorch's interpolation
+bool sam_image_preprocess(const sam_image_u8 & img, sam_image_f32 & res) {
+    const int nx = img.nx;
+    const int ny = img.ny;
+
+    const int nx2 = 1024;
+    const int ny2 = 1024;
+
+    res.nx = nx2;
+    res.ny = ny2;
+    res.data.resize(3*nx2*ny2);
+
+    const float scale = std::max(nx, ny) / 1024.0f;
+
+    fprintf(stderr, "%s: scale = %f\n", __func__, scale);
+
+    const int nx3 = int(nx/scale + 0.5f);
+    const int ny3 = int(ny/scale + 0.5f);
+
+    const float m3[3] = { 123.675f, 116.280f, 103.530f };
+    const float s3[3] = {  58.395f,  57.120f,  57.375f };
+
+    for (int y = 0; y < ny3; y++) {
+        for (int x = 0; x < nx3; x++) {
+            for (int c = 0; c < 3; c++) {
+                // linear interpolation
+                const float sx = (x + 0.5f)*scale - 0.5f;
+                const float sy = (y + 0.5f)*scale - 0.5f;
+
+                const int x0 = std::max(0, (int) std::floor(sx));
+                const int y0 = std::max(0, (int) std::floor(sy));
+
+                const int x1 = std::min(x0 + 1, nx - 1);
+                const int y1 = std::min(y0 + 1, ny - 1);
+
+                const float dx = sx - x0;
+                const float dy = sy - y0;
+
+                const int j00 = 3*(y0*nx + x0) + c;
+                const int j01 = 3*(y0*nx + x1) + c;
+                const int j10 = 3*(y1*nx + x0) + c;
+                const int j11 = 3*(y1*nx + x1) + c;
+
+                const float v00 = img.data[j00];
+                const float v01 = img.data[j01];
+                const float v10 = img.data[j10];
+                const float v11 = img.data[j11];
+
+                const float v0 = v00*(1.0f - dx) + v01*dx;
+                const float v1 = v10*(1.0f - dx) + v11*dx;
+
+                const float v = v0*(1.0f - dy) + v1*dy;
+
+                const uint8_t v2 = std::min(std::max(std::round(v), 0.0f), 255.0f);
+
+                const int i = 3*(y*nx3 + x) + c;
+
+                res.data[i] = (float(v2) - m3[c]) / s3[c];
+            }
+        }
+    }
+
+    return true;
+}
+
+// load the model's weights from a file
+bool sam_model_load(const std::string & fname, sam_model & model) {
+    fprintf(stderr, "%s: loading model from '%s' - please wait ...\n", __func__, fname.c_str());
+
+    auto fin = std::ifstream(fname, std::ios::binary);
+    if (!fin) {
+        fprintf(stderr, "%s: failed to open '%s'\n", __func__, fname.c_str());
+        return false;
+    }
+
+    // verify magic
+    {
+        uint32_t magic;
+        fin.read((char *) &magic, sizeof(magic));
+        if (magic != 0x67676d6c) {
+            fprintf(stderr, "%s: invalid model file '%s' (bad magic)\n", __func__, fname.c_str());
+            return false;
+        }
+    }
+
+    // load hparams
+    {
+        auto & hparams = model.hparams;
+
+        fin.read((char *) &hparams.n_enc_state,     sizeof(hparams.n_enc_state));
+        fin.read((char *) &hparams.n_enc_layer,     sizeof(hparams.n_enc_layer));
+        fin.read((char *) &hparams.n_enc_head,      sizeof(hparams.n_enc_head));
+        fin.read((char *) &hparams.n_enc_out_chans, sizeof(hparams.n_enc_out_chans));
+        fin.read((char *) &hparams.n_pt_embd,       sizeof(hparams.n_pt_embd));
+        fin.read((char *) &hparams.ftype,           sizeof(hparams.ftype));
+
+        const int32_t qntvr = hparams.ftype / GGML_QNT_VERSION_FACTOR;
+
+        printf("%s: n_enc_state      = %d\n", __func__, hparams.n_enc_state);
+        printf("%s: n_enc_layer      = %d\n", __func__, hparams.n_enc_layer);
+        printf("%s: n_enc_head       = %d\n", __func__, hparams.n_enc_head);
+        printf("%s: n_enc_out_chans  = %d\n", __func__, hparams.n_enc_out_chans);
+        printf("%s: n_pt_embd        = %d\n", __func__, hparams.n_pt_embd);
+        printf("%s: ftype            = %d\n", __func__, hparams.ftype);
+        printf("%s: qntvr            = %d\n", __func__, qntvr);
+
+        hparams.ftype %= GGML_QNT_VERSION_FACTOR;
+    }
+
+    // for the big tensors, we have the option to store the data in 16-bit floats or quantized
+    // in order to save memory and also to speed up the computation
+    ggml_type wtype = ggml_ftype_to_ggml_type((ggml_ftype) (model.hparams.ftype));
+    if (wtype == GGML_TYPE_COUNT) {
+        fprintf(stderr, "%s: invalid model file '%s' (bad ftype value %d)\n",
+                __func__, fname.c_str(), model.hparams.ftype);
+        return false;
+    }
+
+    auto & ctx = model.ctx;
+
+    const size_t ctx_size = [&]() {
+        size_t ctx_size = 0;
+
+        const auto & hparams = model.hparams;
+
+        const int32_t n_enc_state     = hparams.n_enc_state;
+        const int32_t n_enc_layer     = hparams.n_enc_layer;
+        const int32_t n_enc_head_dim  = hparams.n_enc_head_dim();
+        const int32_t n_enc_out_chans = hparams.n_enc_out_chans;
+        const int32_t n_pt_embd       = hparams.n_pt_embd;
+
+        const int32_t n_enc_layer_local  = hparams.global_attn_indices().size();
+        const int32_t n_enc_layer_global = n_enc_layer - n_enc_layer_local;
+
+        const int32_t n_img_embd    = hparams.n_img_embd();
+        const int32_t n_window_size = hparams.n_window_size();
+        const int32_t n_patch_size  = hparams.n_patch_size();
+
+        // image encoder
+        {
+            ctx_size += n_enc_state*n_img_embd*n_img_embd*ggml_type_sizef(GGML_TYPE_F32);
+
+            ctx_size += n_enc_state*3*n_patch_size*n_patch_size*ggml_type_sizef(GGML_TYPE_F16);
+            ctx_size += n_enc_state*ggml_type_sizef(GGML_TYPE_F32);
+
+            ctx_size +=     n_enc_state*n_enc_out_chans*1*1*ggml_type_sizef(GGML_TYPE_F16);
+            ctx_size += n_enc_out_chans*n_enc_out_chans*3*3*ggml_type_sizef(GGML_TYPE_F16);
+
+            ctx_size += n_enc_out_chans*ggml_type_sizef(GGML_TYPE_F32);
+            ctx_size += n_enc_out_chans*ggml_type_sizef(GGML_TYPE_F32);
+
+            ctx_size += n_enc_out_chans*ggml_type_sizef(GGML_TYPE_F32);
+            ctx_size += n_enc_out_chans*ggml_type_sizef(GGML_TYPE_F32);
+        }
+
+        // image encoder layers
+        {
+            ctx_size += n_enc_layer*n_enc_state*ggml_type_sizef(GGML_TYPE_F32);
+            ctx_size += n_enc_layer*n_enc_state*ggml_type_sizef(GGML_TYPE_F32);
+
+            ctx_size += n_enc_layer_global*n_enc_head_dim*(2*n_img_embd - 1)*ggml_type_sizef(GGML_TYPE_F16);
+            ctx_size += n_enc_layer_global*n_enc_head_dim*(2*n_img_embd - 1)*ggml_type_sizef(GGML_TYPE_F16);
+
+            ctx_size += n_enc_layer_local*n_enc_head_dim*(2*n_window_size - 1)*ggml_type_sizef(GGML_TYPE_F16);
+            ctx_size += n_enc_layer_local*n_enc_head_dim*(2*n_window_size - 1)*ggml_type_sizef(GGML_TYPE_F16);
+
+            ctx_size += n_enc_layer*3*n_enc_state*n_enc_state*ggml_type_sizef(GGML_TYPE_F16);
+            ctx_size += n_enc_layer*3*n_enc_state*            ggml_type_sizef(GGML_TYPE_F32);
+
+            ctx_size += n_enc_layer*n_enc_state*n_enc_state*ggml_type_sizef(GGML_TYPE_F16);
+            ctx_size += n_enc_layer*n_enc_state*            ggml_type_sizef(GGML_TYPE_F32);
+
+            ctx_size += n_enc_layer*n_enc_state*ggml_type_sizef(GGML_TYPE_F32);
+            ctx_size += n_enc_layer*n_enc_state*ggml_type_sizef(GGML_TYPE_F32);
+
+            ctx_size += n_enc_layer*4*n_enc_state*n_enc_state*ggml_type_sizef(GGML_TYPE_F16);
+            ctx_size += n_enc_layer*4*n_enc_state*            ggml_type_sizef(GGML_TYPE_F32);
+
+            ctx_size += n_enc_layer*4*n_enc_state*n_enc_state*ggml_type_sizef(GGML_TYPE_F16);
+            ctx_size += n_enc_layer*4*n_enc_state*            ggml_type_sizef(GGML_TYPE_F32);
+        }
+
+        ctx_size += (8 + 14*n_enc_layer)*ggml_tensor_overhead();
+
+        // prompt encoder
+        {
+            ctx_size += n_enc_out_chans*ggml_type_sizef(GGML_TYPE_F16); // 2*(n_enc_out_chans/2)
+
+            ctx_size += n_enc_out_chans*ggml_type_sizef(GGML_TYPE_F32);
+            ctx_size += n_pt_embd*n_enc_out_chans*ggml_type_sizef(GGML_TYPE_F32);
+        }
+
+        ctx_size += (2 + n_pt_embd)*ggml_tensor_overhead();
+
+        // mask decoder
+        {
+            //transformer
+            {
+                const int tfm_layers_count = 2;
+                const int qkv_count = 3;
+                const int norm_count = 4;
+                const int n_hypernet_mpls_count = 4;
+
+                // self_attn
+                ctx_size += tfm_layers_count*qkv_count*n_enc_state*n_enc_state*ggml_type_sizef(GGML_TYPE_F16);
+                ctx_size += tfm_layers_count*qkv_count*n_enc_state*            ggml_type_sizef(GGML_TYPE_F32);
+                ctx_size += tfm_layers_count*n_enc_state*                      ggml_type_sizef(GGML_TYPE_F32);
+
+                // all norms
+                ctx_size += tfm_layers_count*norm_count*n_enc_state*ggml_type_sizef(GGML_TYPE_F32);
+                ctx_size += tfm_layers_count*norm_count*n_enc_state*ggml_type_sizef(GGML_TYPE_F32);
+
+                // cross_attn_token_to_img
+                ctx_size += tfm_layers_count*qkv_count*n_enc_state*(n_enc_state/2)*ggml_type_sizef(GGML_TYPE_F16);
+                ctx_size += tfm_layers_count*qkv_count*(n_enc_state/2)*            ggml_type_sizef(GGML_TYPE_F32);
+                ctx_size += tfm_layers_count*n_enc_state*                          ggml_type_sizef(GGML_TYPE_F32);
+
+                // mlp
+                ctx_size += tfm_layers_count*8*n_enc_out_chans*n_enc_out_chans*ggml_type_sizef(GGML_TYPE_F16);
+                ctx_size += tfm_layers_count*8*n_enc_out_chans*                ggml_type_sizef(GGML_TYPE_F32);
+                ctx_size += tfm_layers_count*n_enc_out_chans*8*n_enc_out_chans*ggml_type_sizef(GGML_TYPE_F16);
+                ctx_size += tfm_layers_count*n_enc_out_chans*                  ggml_type_sizef(GGML_TYPE_F32);
+
+                // cross_attn_img_to_token
+                ctx_size += tfm_layers_count*qkv_count*n_enc_state*(n_enc_state/2)*ggml_type_sizef(GGML_TYPE_F16);
+                ctx_size += tfm_layers_count*qkv_count*(n_enc_state/2)*            ggml_type_sizef(GGML_TYPE_F32);
+                ctx_size += tfm_layers_count*n_enc_state*                          ggml_type_sizef(GGML_TYPE_F32);
+
+                // transformer_final_attn_token_to_img
+                ctx_size += qkv_count*n_enc_state*(n_enc_state/2)*ggml_type_sizef(GGML_TYPE_F16);
+                ctx_size += qkv_count*(n_enc_state/2)*            ggml_type_sizef(GGML_TYPE_F32);
+                ctx_size += n_enc_state*                          ggml_type_sizef(GGML_TYPE_F32);
+
+                // transformer_norm_final
+                ctx_size += norm_count*n_enc_state*ggml_type_sizef(GGML_TYPE_F32);
+                ctx_size += norm_count*n_enc_state*ggml_type_sizef(GGML_TYPE_F32);
+
+                // output_upscaling
+                ctx_size += n_enc_out_chans*n_img_embd*2*2*ggml_type_sizef(GGML_TYPE_F16);
+                ctx_size += 3*n_img_embd*                  ggml_type_sizef(GGML_TYPE_F32);
+                ctx_size += n_enc_out_chans*n_img_embd*(n_img_embd/2)*2*2*ggml_type_sizef(GGML_TYPE_F16);
+                ctx_size += (n_img_embd/2)*                               ggml_type_sizef(GGML_TYPE_F32);
+
+                // output_hypernetworks_mlps
+                ctx_size += n_hypernet_mpls_count*2*n_enc_out_chans*n_enc_out_chans*ggml_type_sizef(GGML_TYPE_F16);
+                ctx_size += n_hypernet_mpls_count*2*n_enc_out_chans*                ggml_type_sizef(GGML_TYPE_F32);
+                ctx_size += n_hypernet_mpls_count*n_enc_out_chans*(n_img_embd/2)*ggml_type_sizef(GGML_TYPE_F16);
+                ctx_size += n_hypernet_mpls_count*(n_img_embd/2)*                ggml_type_sizef(GGML_TYPE_F32);
+
+                // iou_prediction_head
+                ctx_size += 2*n_enc_out_chans*n_enc_out_chans*ggml_type_sizef(GGML_TYPE_F16);
+                ctx_size += 2*n_enc_out_chans*                ggml_type_sizef(GGML_TYPE_F32);
+                ctx_size += n_pt_embd*n_enc_out_chans*ggml_type_sizef(GGML_TYPE_F16);
+                ctx_size += n_pt_embd*                ggml_type_sizef(GGML_TYPE_F32);
+
+                // iou_token_w
+                ctx_size += n_enc_out_chans*ggml_type_sizef(GGML_TYPE_F32);
+
+                // mask_tokens_w
+                ctx_size += n_pt_embd*n_enc_out_chans*ggml_type_sizef(GGML_TYPE_F32);
+            }
+        }
+        fprintf(stderr, "%s: ggml ctx size = %6.2f MB\n", __func__, ctx_size/(1024.0*1024.0));
+
+        return ctx_size;
+    }();
+
+    // create the ggml context
+    {
+        struct ggml_init_params params = {
+            /*.mem_size   =*/ ctx_size,
+            /*.mem_buffer =*/ NULL,
+            /*.no_alloc   =*/ false,
+        };
+
+        ctx = ggml_init(params);
+        if (!ctx) {
+            fprintf(stderr, "%s: ggml_init() failed\n", __func__);
+            return false;
+        }
+    }
+
+    // prepare memory for the weights
+    {
+        const auto & hparams = model.hparams;
+
+        const int32_t n_enc_state      = hparams.n_enc_state;
+        const int32_t n_enc_layer      = hparams.n_enc_layer;
+        const int32_t n_enc_head_dim   = hparams.n_enc_head_dim();
+        const int32_t n_enc_out_chans  = hparams.n_enc_out_chans;
+        const int32_t n_pt_embd        = hparams.n_pt_embd;
+
+        const int32_t n_img_embd    = hparams.n_img_embd();
+        const int32_t n_window_size = hparams.n_window_size();
+        const int32_t n_patch_size  = hparams.n_patch_size();
+
+        model.enc_img.layers.resize(n_enc_layer);
+
+        // image encoder
+        {
+            auto & enc = model.enc_img;
+
+            enc.pe = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, n_enc_state, n_img_embd, n_img_embd, 1);
+
+            enc.proj_w = ggml_new_tensor_4d(ctx, GGML_TYPE_F16, n_patch_size, n_patch_size,           3, n_enc_state);
+            enc.proj_b = ggml_new_tensor_3d(ctx, GGML_TYPE_F32,            1,            1, n_enc_state);
+
+            enc.neck_conv_0 = ggml_new_tensor_4d(ctx, GGML_TYPE_F16, 1, 1, n_enc_state,     n_enc_out_chans);
+            enc.neck_conv_1 = ggml_new_tensor_4d(ctx, GGML_TYPE_F16, 3, 3, n_enc_out_chans, n_enc_out_chans);
+
+            enc.neck_norm_0_w = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_enc_out_chans);
+            enc.neck_norm_0_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_enc_out_chans);
+
+            enc.neck_norm_1_w = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_enc_out_chans);
+            enc.neck_norm_1_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_enc_out_chans);
+
+            model.tensors["image_encoder.pos_embed"] = enc.pe;
+
+            model.tensors["image_encoder.patch_embed.proj.weight"] = enc.proj_w;
+            model.tensors["image_encoder.patch_embed.proj.bias"]   = enc.proj_b;
+
+            model.tensors["image_encoder.neck.0.weight"] = enc.neck_conv_0;
+            model.tensors["image_encoder.neck.2.weight"] = enc.neck_conv_1;
+
+            model.tensors["image_encoder.neck.1.weight"] = enc.neck_norm_0_w;
+            model.tensors["image_encoder.neck.1.bias"]   = enc.neck_norm_0_b;
+
+            model.tensors["image_encoder.neck.3.weight"] = enc.neck_norm_1_w;
+            model.tensors["image_encoder.neck.3.bias"]   = enc.neck_norm_1_b;
+
+            for (int i = 0; i < n_enc_layer; ++i) {
+                auto & layer = enc.layers[i];
+
+                layer.norm1_w = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_enc_state);
+                layer.norm1_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_enc_state);
+
+                if (hparams.is_global_attn(i)) {
+                    layer.rel_pos_w = ggml_new_tensor_2d(ctx, GGML_TYPE_F16, n_enc_head_dim, 2*n_img_embd - 1);
+                    layer.rel_pos_h = ggml_new_tensor_2d(ctx, GGML_TYPE_F16, n_enc_head_dim, 2*n_img_embd - 1);
+                } else {
+                    layer.rel_pos_w = ggml_new_tensor_2d(ctx, GGML_TYPE_F16, n_enc_head_dim, 2*n_window_size - 1);
+                    layer.rel_pos_h = ggml_new_tensor_2d(ctx, GGML_TYPE_F16, n_enc_head_dim, 2*n_window_size - 1);
+                }
+
+                layer.qkv_w = ggml_new_tensor_2d(ctx, GGML_TYPE_F16,   n_enc_state, 3*n_enc_state);
+                layer.qkv_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 3*n_enc_state);
+
+                layer.proj_w = ggml_new_tensor_2d(ctx, GGML_TYPE_F16,  n_enc_state,   n_enc_state);
+                layer.proj_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,  n_enc_state);
+
+                layer.norm2_w = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_enc_state);
+                layer.norm2_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_enc_state);
+
+                layer.mlp_lin1_w = ggml_new_tensor_2d(ctx, GGML_TYPE_F16,   n_enc_state, 4*n_enc_state);
+                layer.mlp_lin1_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 4*n_enc_state);
+
+                layer.mlp_lin2_w = ggml_new_tensor_2d(ctx, GGML_TYPE_F16, 4*n_enc_state,   n_enc_state);
+                layer.mlp_lin2_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_enc_state);
+
+                model.tensors["image_encoder.blocks." + std::to_string(i) + ".norm1.weight"] = layer.norm1_w;
+                model.tensors["image_encoder.blocks." + std::to_string(i) + ".norm1.bias"]   = layer.norm1_b;
+
+                model.tensors["image_encoder.blocks." + std::to_string(i) + ".attn.rel_pos_w"] = layer.rel_pos_w;
+                model.tensors["image_encoder.blocks." + std::to_string(i) + ".attn.rel_pos_h"] = layer.rel_pos_h;
+
+                model.tensors["image_encoder.blocks." + std::to_string(i) + ".attn.qkv.weight"] = layer.qkv_w;
+                model.tensors["image_encoder.blocks." + std::to_string(i) + ".attn.qkv.bias"]   = layer.qkv_b;
+
+                model.tensors["image_encoder.blocks." + std::to_string(i) + ".attn.proj.weight"] = layer.proj_w;
+                model.tensors["image_encoder.blocks." + std::to_string(i) + ".attn.proj.bias"]   = layer.proj_b;
+
+                model.tensors["image_encoder.blocks." + std::to_string(i) + ".norm2.weight"] = layer.norm2_w;
+                model.tensors["image_encoder.blocks." + std::to_string(i) + ".norm2.bias"]   = layer.norm2_b;
+
+                model.tensors["image_encoder.blocks." + std::to_string(i) + ".mlp.lin1.weight"] = layer.mlp_lin1_w;
+                model.tensors["image_encoder.blocks." + std::to_string(i) + ".mlp.lin1.bias"]   = layer.mlp_lin1_b;
+
+                model.tensors["image_encoder.blocks." + std::to_string(i) + ".mlp.lin2.weight"] = layer.mlp_lin2_w;
+                model.tensors["image_encoder.blocks." + std::to_string(i) + ".mlp.lin2.bias"]   = layer.mlp_lin2_b;
+            }
+        }
+
+        // prompt encoder
+        {
+            auto & enc = model.enc_prompt;
+
+            enc.pe = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_enc_out_chans/2, 2);
+
+            enc.not_a_pt_embd_w = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_enc_out_chans);
+            enc.no_mask_embd_w  = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_enc_out_chans);
+
+            model.tensors["prompt_encoder.pe_layer.positional_encoding_gaussian_matrix"] = enc.pe;
+            model.tensors["prompt_encoder.not_a_point_embed.weight"] = enc.not_a_pt_embd_w;
+            model.tensors["prompt_encoder.no_mask_embed.weight"]     = enc.no_mask_embd_w;
+
+            enc.pt_embd.resize(n_pt_embd);
+            for (int i = 0; i < n_pt_embd; i++) {
+                enc.pt_embd[i] = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_enc_out_chans);
+
+                model.tensors["prompt_encoder.point_embeddings." + std::to_string(i) + ".weight"] = enc.pt_embd[i];
+            }
+        }
+
+        // mask decoder
+        {
+            auto & dec = model.dec;
+            auto & tfm_layers = dec.transformer_layers;
+
+            const int tfm_layers_count = 2;
+            tfm_layers.resize(tfm_layers_count);
+            for (int i = 0; i < tfm_layers_count; ++i) {
+                auto& l = tfm_layers[i];
+                l.self_attn.q_w = ggml_new_tensor_2d(ctx, GGML_TYPE_F16, n_enc_out_chans, n_enc_out_chans);
+                l.self_attn.q_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_enc_out_chans);
+                l.self_attn.k_w = ggml_new_tensor_2d(ctx, GGML_TYPE_F16, n_enc_out_chans, n_enc_out_chans);
+                l.self_attn.k_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_enc_out_chans);
+                l.self_attn.v_w = ggml_new_tensor_2d(ctx, GGML_TYPE_F16, n_enc_out_chans, n_enc_out_chans);
+                l.self_attn.v_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_enc_out_chans);
+                l.self_attn.out_w = ggml_new_tensor_2d(ctx, GGML_TYPE_F16, n_enc_out_chans, n_enc_out_chans);
+                l.self_attn.out_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_enc_out_chans);
+
+                l.norm1_w = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_enc_out_chans);
+                l.norm1_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_enc_out_chans);
+
+                l.cross_attn_token_to_img.q_w = ggml_new_tensor_2d(ctx, GGML_TYPE_F16, n_enc_out_chans, n_enc_out_chans/2);
+                l.cross_attn_token_to_img.q_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_enc_out_chans/2);
+                l.cross_attn_token_to_img.k_w = ggml_new_tensor_2d(ctx, GGML_TYPE_F16, n_enc_out_chans, n_enc_out_chans/2);
+                l.cross_attn_token_to_img.k_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_enc_out_chans/2);
+                l.cross_attn_token_to_img.v_w = ggml_new_tensor_2d(ctx, GGML_TYPE_F16, n_enc_out_chans, n_enc_out_chans/2);
+                l.cross_attn_token_to_img.v_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_enc_out_chans/2);
+                l.cross_attn_token_to_img.out_w = ggml_new_tensor_2d(ctx, GGML_TYPE_F16, n_enc_out_chans/2, n_enc_out_chans);
+                l.cross_attn_token_to_img.out_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_enc_out_chans);
+
+                l.norm2_w = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_enc_out_chans);
+                l.norm2_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_enc_out_chans);
+
+                l.mlp_lin1_w = ggml_new_tensor_2d(ctx, GGML_TYPE_F16, n_enc_out_chans, 8*n_enc_out_chans);
+                l.mlp_lin1_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 8*n_enc_out_chans);
+                l.mlp_lin2_w = ggml_new_tensor_2d(ctx, GGML_TYPE_F16, 8*n_enc_out_chans, n_enc_out_chans);
+                l.mlp_lin2_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_enc_out_chans);
+
+                l.norm3_w = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_enc_out_chans);
+                l.norm3_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_enc_out_chans);
+
+                l.norm4_w = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_enc_out_chans);
+                l.norm4_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_enc_out_chans);
+
+                l.cross_attn_img_to_token.q_w = ggml_new_tensor_2d(ctx, GGML_TYPE_F16, n_enc_out_chans, n_enc_out_chans/2);
+                l.cross_attn_img_to_token.q_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_enc_out_chans/2);
+                l.cross_attn_img_to_token.k_w = ggml_new_tensor_2d(ctx, GGML_TYPE_F16, n_enc_out_chans, n_enc_out_chans/2);
+                l.cross_attn_img_to_token.k_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_enc_out_chans/2);
+                l.cross_attn_img_to_token.v_w = ggml_new_tensor_2d(ctx, GGML_TYPE_F16, n_enc_out_chans, n_enc_out_chans/2);
+                l.cross_attn_img_to_token.v_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_enc_out_chans/2);
+                l.cross_attn_img_to_token.out_w = ggml_new_tensor_2d(ctx, GGML_TYPE_F16, n_enc_out_chans/2, n_enc_out_chans);
+                l.cross_attn_img_to_token.out_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_enc_out_chans);
+
+                const auto prefix = "mask_decoder.transformer.layers." + std::to_string(i) + ".";
+                model.tensors[prefix + "self_attn.q_proj.weight"] = l.self_attn.q_w;
+                model.tensors[prefix + "self_attn.q_proj.bias"]   = l.self_attn.q_b;
+                model.tensors[prefix + "self_attn.k_proj.weight"] = l.self_attn.k_w;
+                model.tensors[prefix + "self_attn.k_proj.bias"]   = l.self_attn.k_b;
+                model.tensors[prefix + "self_attn.v_proj.weight"] = l.self_attn.v_w;
+                model.tensors[prefix + "self_attn.v_proj.bias"]   = l.self_attn.v_b;
+                model.tensors[prefix + "self_attn.out_proj.weight"] = l.self_attn.out_w;
+                model.tensors[prefix + "self_attn.out_proj.bias"]   = l.self_attn.out_b;
+
+                model.tensors[prefix + "norm1.weight"] = l.norm1_w;
+                model.tensors[prefix + "norm1.bias"]   = l.norm1_b;
+
+                model.tensors[prefix + "cross_attn_token_to_image.q_proj.weight"] = l.cross_attn_token_to_img.q_w;
+                model.tensors[prefix + "cross_attn_token_to_image.q_proj.bias"]   = l.cross_attn_token_to_img.q_b;
+                model.tensors[prefix + "cross_attn_token_to_image.k_proj.weight"] = l.cross_attn_token_to_img.k_w;
+                model.tensors[prefix + "cross_attn_token_to_image.k_proj.bias"]   = l.cross_attn_token_to_img.k_b;
+                model.tensors[prefix + "cross_attn_token_to_image.v_proj.weight"] = l.cross_attn_token_to_img.v_w;
+                model.tensors[prefix + "cross_attn_token_to_image.v_proj.bias"]   = l.cross_attn_token_to_img.v_b;
+                model.tensors[prefix + "cross_attn_token_to_image.out_proj.weight"] = l.cross_attn_token_to_img.out_w;
+                model.tensors[prefix + "cross_attn_token_to_image.out_proj.bias"]   = l.cross_attn_token_to_img.out_b;
+
+                model.tensors[prefix + "norm2.weight"] = l.norm2_w;
+                model.tensors[prefix + "norm2.bias"]   = l.norm2_b;
+
+                model.tensors[prefix + "mlp.lin1.weight"] = l.mlp_lin1_w;
+                model.tensors[prefix + "mlp.lin1.bias"]   = l.mlp_lin1_b;
+                model.tensors[prefix + "mlp.lin2.weight"] = l.mlp_lin2_w;
+                model.tensors[prefix + "mlp.lin2.bias"]   = l.mlp_lin2_b;
+
+                model.tensors[prefix + "norm3.weight"] = l.norm3_w;
+                model.tensors[prefix + "norm3.bias"]   = l.norm3_b;
+                model.tensors[prefix + "norm4.weight"] = l.norm4_w;
+                model.tensors[prefix + "norm4.bias"]   = l.norm4_b;
+
+                model.tensors[prefix + "cross_attn_image_to_token.q_proj.weight"] = l.cross_attn_img_to_token.q_w;
+                model.tensors[prefix + "cross_attn_image_to_token.q_proj.bias"]   = l.cross_attn_img_to_token.q_b;
+                model.tensors[prefix + "cross_attn_image_to_token.k_proj.weight"] = l.cross_attn_img_to_token.k_w;
+                model.tensors[prefix + "cross_attn_image_to_token.k_proj.bias"]   = l.cross_attn_img_to_token.k_b;
+                model.tensors[prefix + "cross_attn_image_to_token.v_proj.weight"] = l.cross_attn_img_to_token.v_w;
+                model.tensors[prefix + "cross_attn_image_to_token.v_proj.bias"]   = l.cross_attn_img_to_token.v_b;
+                model.tensors[prefix + "cross_attn_image_to_token.out_proj.weight"] = l.cross_attn_img_to_token.out_w;
+                model.tensors[prefix + "cross_attn_image_to_token.out_proj.bias"]   = l.cross_attn_img_to_token.out_b;
+            }
+
+            dec.transformer_final_attn_token_to_img.q_w = ggml_new_tensor_2d(ctx, GGML_TYPE_F16, n_enc_out_chans, n_enc_out_chans/2);
+            dec.transformer_final_attn_token_to_img.q_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_enc_out_chans/2);
+            dec.transformer_final_attn_token_to_img.k_w = ggml_new_tensor_2d(ctx, GGML_TYPE_F16, n_enc_out_chans, n_enc_out_chans/2);
+            dec.transformer_final_attn_token_to_img.k_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_enc_out_chans/2);
+            dec.transformer_final_attn_token_to_img.v_w = ggml_new_tensor_2d(ctx, GGML_TYPE_F16, n_enc_out_chans, n_enc_out_chans/2);
+            dec.transformer_final_attn_token_to_img.v_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_enc_out_chans/2);
+            dec.transformer_final_attn_token_to_img.out_w = ggml_new_tensor_2d(ctx, GGML_TYPE_F16, n_enc_out_chans/2, n_enc_out_chans);
+            dec.transformer_final_attn_token_to_img.out_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_enc_out_chans);
+
+            model.tensors["mask_decoder.transformer.final_attn_token_to_image.q_proj.weight"] = dec.transformer_final_attn_token_to_img.q_w;
+            model.tensors["mask_decoder.transformer.final_attn_token_to_image.q_proj.bias"]   = dec.transformer_final_attn_token_to_img.q_b;
+            model.tensors["mask_decoder.transformer.final_attn_token_to_image.k_proj.weight"] = dec.transformer_final_attn_token_to_img.k_w;
+            model.tensors["mask_decoder.transformer.final_attn_token_to_image.k_proj.bias"]   = dec.transformer_final_attn_token_to_img.k_b;
+            model.tensors["mask_decoder.transformer.final_attn_token_to_image.v_proj.weight"] = dec.transformer_final_attn_token_to_img.v_w;
+            model.tensors["mask_decoder.transformer.final_attn_token_to_image.v_proj.bias"]   = dec.transformer_final_attn_token_to_img.v_b;
+            model.tensors["mask_decoder.transformer.final_attn_token_to_image.out_proj.weight"] = dec.transformer_final_attn_token_to_img.out_w;
+            model.tensors["mask_decoder.transformer.final_attn_token_to_image.out_proj.bias"]   = dec.transformer_final_attn_token_to_img.out_b;
+
+            dec.transformer_norm_final_w = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_enc_out_chans);
+            dec.transformer_norm_final_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_enc_out_chans);
+
+            model.tensors["mask_decoder.transformer.norm_final_attn.weight"] = dec.transformer_norm_final_w;
+            model.tensors["mask_decoder.transformer.norm_final_attn.bias"]   = dec.transformer_norm_final_b;
+
+            dec.output_upscaling_0_w = ggml_new_tensor_4d(ctx, GGML_TYPE_F16, 2, 2, n_img_embd, n_enc_out_chans);
+            dec.output_upscaling_0_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_img_embd);
+            dec.output_upscaling_1_w = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_img_embd);
+            dec.output_upscaling_1_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_img_embd);
+            dec.output_upscaling_3_w = ggml_new_tensor_4d(ctx, GGML_TYPE_F16,  2, 2, n_img_embd/2, n_img_embd);
+            dec.output_upscaling_3_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_img_embd/2);
+
+            model.tensors["mask_decoder.output_upscaling.0.weight"] = dec.output_upscaling_0_w;
+            model.tensors["mask_decoder.output_upscaling.0.bias"]   = dec.output_upscaling_0_b;
+            model.tensors["mask_decoder.output_upscaling.1.weight"] = dec.output_upscaling_1_w;
+            model.tensors["mask_decoder.output_upscaling.1.bias"]   = dec.output_upscaling_1_b;
+            model.tensors["mask_decoder.output_upscaling.3.weight"] = dec.output_upscaling_3_w;
+            model.tensors["mask_decoder.output_upscaling.3.bias"]   = dec.output_upscaling_3_b;
+
+            const int n_hypernet_mpls_count = 4;
+            dec.output_hypernet_mlps.resize(n_hypernet_mpls_count);
+            for (int i = 0; i < n_hypernet_mpls_count; ++i) {
+                auto& mlp = dec.output_hypernet_mlps[i];
+
+                mlp.w_0 = ggml_new_tensor_2d(ctx, GGML_TYPE_F16, n_enc_out_chans, n_enc_out_chans);
+                mlp.b_0 = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_enc_out_chans);
+                mlp.w_1 = ggml_new_tensor_2d(ctx, GGML_TYPE_F16, n_enc_out_chans, n_enc_out_chans);
+                mlp.b_1 = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_enc_out_chans);
+                mlp.w_2 = ggml_new_tensor_2d(ctx, GGML_TYPE_F16, n_enc_out_chans, n_img_embd/2);
+                mlp.b_2 = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_img_embd/2);
+
+                const auto prefix = "mask_decoder.output_hypernetworks_mlps." + std::to_string(i) + ".";
+                model.tensors[prefix + "layers.0.weight"] = mlp.w_0;
+                model.tensors[prefix + "layers.0.bias"]   = mlp.b_0;
+                model.tensors[prefix + "layers.1.weight"] = mlp.w_1;
+                model.tensors[prefix + "layers.1.bias"]   = mlp.b_1;
+                model.tensors[prefix + "layers.2.weight"] = mlp.w_2;
+                model.tensors[prefix + "layers.2.bias"]   = mlp.b_2;
+            }
+
+            dec.iou_prediction_head_0_w = ggml_new_tensor_2d(ctx, GGML_TYPE_F16, n_enc_out_chans, n_enc_out_chans);
+            dec.iou_prediction_head_0_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_enc_out_chans);
+            dec.iou_prediction_head_1_w = ggml_new_tensor_2d(ctx, GGML_TYPE_F16, n_enc_out_chans, n_enc_out_chans);
+            dec.iou_prediction_head_1_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_enc_out_chans);
+            dec.iou_prediction_head_2_w = ggml_new_tensor_2d(ctx, GGML_TYPE_F16, n_enc_out_chans, n_pt_embd);
+            dec.iou_prediction_head_2_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_pt_embd);
+
+            dec.iou_token_w = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_enc_out_chans, 1);
+            dec.mask_tokens_w = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_enc_out_chans, n_pt_embd);
+
+            model.tensors["mask_decoder.iou_prediction_head.layers.0.weight"] = dec.iou_prediction_head_0_w;
+            model.tensors["mask_decoder.iou_prediction_head.layers.0.bias"]   = dec.iou_prediction_head_0_b;
+            model.tensors["mask_decoder.iou_prediction_head.layers.1.weight"] = dec.iou_prediction_head_1_w;
+            model.tensors["mask_decoder.iou_prediction_head.layers.1.bias"]   = dec.iou_prediction_head_1_b;
+            model.tensors["mask_decoder.iou_prediction_head.layers.2.weight"] = dec.iou_prediction_head_2_w;
+            model.tensors["mask_decoder.iou_prediction_head.layers.2.bias"]   = dec.iou_prediction_head_2_b;
+
+            model.tensors["mask_decoder.iou_token.weight"] = dec.iou_token_w;
+            model.tensors["mask_decoder.mask_tokens.weight"] = dec.mask_tokens_w;
+        }
+    }
+
+    // load weights
+    {
+        int n_tensors = 0;
+        size_t total_size = 0;
+
+        fprintf(stderr, "%s: ", __func__);
+
+        while (true) {
+            int32_t n_dims;
+            int32_t length;
+            int32_t ftype;
+
+            fin.read(reinterpret_cast<char *>(&n_dims), sizeof(n_dims));
+            fin.read(reinterpret_cast<char *>(&length), sizeof(length));
+            fin.read(reinterpret_cast<char *>(&ftype),  sizeof(ftype));
+
+            if (fin.eof()) {
+                break;
+            }
+
+            int64_t nelements = 1;
+            int64_t ne[4] = { 1, 1, 1, 1 };
+            for (int i = 0; i < n_dims; ++i) {
+                int32_t ne_cur;
+                fin.read(reinterpret_cast<char *>(&ne_cur), sizeof(ne_cur));
+                ne[i] = ne_cur;
+                nelements *= ne[i];
+            }
+
+            std::string name(length, 0);
+            fin.read(&name[0], length);
+
+            if (model.tensors.find(name.data()) == model.tensors.end()) {
+                fprintf(stderr, "%s: unknown tensor '%s' in model file\n", __func__, name.data());
+                return false;
+            }
+
+            auto tensor = model.tensors[name.data()];
+            //printf("ne0 = %jd, ne1 = %jd, ne2 = %jd, ne3 = %jd\n", ne[0], ne[1], ne[2], ne[3]);
+
+            if (ggml_nelements(tensor) != nelements) {
+                fprintf(stderr, "%s: tensor '%s' has wrong size in model file: got %d, expected %d\n",
+                        __func__, name.data(), (int) nelements, (int) ggml_nelements(tensor));
+                return false;
+            }
+
+            if (tensor->ne[0] != ne[0] || tensor->ne[1] != ne[1] || tensor->ne[2] != ne[2] || tensor->ne[3] != ne[3]) {
+                fprintf(stderr, "%s: tensor '%s' has wrong shape in model file: got [%d, %d, %d, %d], expected [%d, %d, %d, %d]\n",
+                        __func__, name.data(),
+                        (int) ne[0], (int) ne[1], (int) ne[2], (int) ne[3],
+                        (int) tensor->ne[0], (int) tensor->ne[1], (int) tensor->ne[2], (int) tensor->ne[3]);
+                return false;
+            }
+
+            size_t bpe = 0;
+
+            switch (ftype) {
+                case 0: bpe = ggml_type_size(GGML_TYPE_F32);  break;
+                case 1: bpe = ggml_type_size(GGML_TYPE_F16);  break;
+                case 2: bpe = ggml_type_size(GGML_TYPE_Q4_0); assert(ne[0] % 64 == 0); break;
+                case 3: bpe = ggml_type_size(GGML_TYPE_Q4_1); assert(ne[0] % 64 == 0); break;
+                default:
+                        {
+                            fprintf(stderr, "%s: unknown ftype %d in model file\n", __func__, ftype);
+                            return false;
+                        }
+            };
+
+            if ((nelements*bpe)/ggml_blck_size(tensor->type) != ggml_nbytes(tensor)) {
+                fprintf(stderr, "%s: tensor '%s' has wrong size in model file: got %zu, expected %zu\n",
+                        __func__, name.data(), ggml_nbytes(tensor), (size_t) nelements*bpe);
+                return false;
+            }
+
+            fin.read(reinterpret_cast<char *>(tensor->data), ggml_nbytes(tensor));
+
+            total_size += ggml_nbytes(tensor);
+            if (++n_tensors % 8 == 0) {
+                fprintf(stderr, ".");
+                fflush(stdout);
+            }
+        }
+
+        if (n_tensors != int(model.tensors.size())) {
+            fprintf(stderr, "%s: model file has %d tensors, but %d tensors were expected\n", __func__, n_tensors, (int) model.tensors.size());
+            return false;
+        }
+
+        fprintf(stderr, " done\n");
+
+        fprintf(stderr, "%s: model size = %8.2f MB / num tensors = %d\n", __func__, total_size/1024.0/1024.0, n_tensors);
+    }
+
+    fin.close();
+
+    return true;
+}
+
+struct ggml_tensor * sam_fill_dense_pe(
+            const sam_model   & model,
+          struct ggml_context * ctx0,
+          struct ggml_cgraph  * gf,
+                  sam_state   & state) {
+    const auto & hparams = model.hparams;
+    const auto & enc     = model.enc_prompt;
+
+    const int32_t n_img_embd = hparams.n_img_embd();
+    const float n_img_embd_inv = 1.0f / n_img_embd;
+
+    struct ggml_tensor * xy_embed_stacked = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, 2, n_img_embd, n_img_embd);
+    ggml_allocr_alloc(state.allocr, xy_embed_stacked);
+
+    if (!ggml_allocr_is_measure(state.allocr)) {
+        float * data = (float *) ggml_get_data(xy_embed_stacked);
+        for (int i = 0; i < n_img_embd; ++i) {
+            const int row = 2*i*n_img_embd;
+            const float y_val = 2 * (i + 0.5f) * n_img_embd_inv - 1;
+            for (int j = 0; j < n_img_embd; ++j) {
+                const float x_val = 2 * (j + 0.5f) * n_img_embd_inv - 1;
+                data[row + 2*j + 0] = x_val;
+                data[row + 2*j + 1] = y_val;
+            }
+        }
+    }
+
+    struct ggml_tensor * cur = ggml_mul_mat(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, enc.pe)), xy_embed_stacked);
+
+    cur = ggml_scale(ctx0, cur, ggml_new_f32(ctx0, float(2.0*M_PI)));
+
+    // concat
+    // ref: https://github.com/facebookresearch/segment-anything/blob/main/segment_anything/modeling/prompt_encoder.py#L192
+    {
+        struct ggml_tensor * t_sin = ggml_map_custom1(ctx0, cur, ggml_sam_sin, GGML_N_TASKS_MAX, NULL);
+        struct ggml_tensor * t_cos = ggml_map_custom1(ctx0, cur, ggml_sam_cos, GGML_N_TASKS_MAX, NULL);
+
+        cur = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, t_sin->ne[0] + t_cos->ne[0], cur->ne[1], cur->ne[2]);
+
+        ggml_build_forward_expand(gf, ggml_cpy(ctx0, t_sin, ggml_view_3d(ctx0, cur, t_sin->ne[0], t_sin->ne[1], t_sin->ne[2], cur->nb[1], cur->nb[2], 0)));
+        ggml_build_forward_expand(gf, ggml_cpy(ctx0, t_cos, ggml_view_3d(ctx0, cur, t_sin->ne[0], t_sin->ne[1], t_sin->ne[2], cur->nb[1], cur->nb[2], t_sin->nb[1])));
+    }
+
+    struct ggml_tensor * pe_img_dense = ggml_cont(ctx0, ggml_permute(ctx0, cur, 2, 0, 1, 3));
+    ggml_build_forward_expand(gf, pe_img_dense);
+
+    return pe_img_dense;
+}
+
+struct ggml_tensor* sam_layer_norm_2d(
+                    struct ggml_context * ctx0,
+                    struct ggml_tensor  * layer,
+                    int                   n_channels,
+                    struct ggml_tensor  * w,
+                    struct ggml_tensor  * b,
+                    float                 eps) {
+    // LayerNorm2d
+    // normalize along channel dimmension
+    // TODO: better implementation
+    layer = ggml_permute(ctx0,
+                ggml_norm(ctx0, ggml_cont(ctx0, ggml_permute(ctx0, layer, 1, 2, 0, 3)), eps),
+                2, 0, 1, 3);
+
+    layer = ggml_add(ctx0,
+              ggml_mul(ctx0,
+                  ggml_repeat(ctx0, ggml_reshape_3d(ctx0, w, 1, 1, n_channels), layer),
+                  layer),
+              ggml_repeat(ctx0, ggml_reshape_3d(ctx0, b, 1, 1, n_channels), layer));
+
+    return layer;
+}
+
+struct ggml_cgraph  * sam_encode_image(
+            const sam_model & model,
+                  sam_state & state,
+        const sam_image_f32 & img) {
+
+    const auto & hparams = model.hparams;
+    const auto & enc     = model.enc_img;
+
+    const int32_t n_enc_state     = hparams.n_enc_state;
+    const int32_t n_enc_layer     = hparams.n_enc_layer;
+    const int32_t n_enc_head      = hparams.n_enc_head;
+    const int32_t n_enc_head_dim  = hparams.n_enc_head_dim();
+    const int32_t n_enc_out_chans = hparams.n_enc_out_chans;
+    const int32_t n_img_size    = hparams.n_img_size();
+    const int32_t n_window_size = hparams.n_window_size();
+
+    struct ggml_init_params ggml_params = {
+        /*.mem_size   =*/ state.buf_compute_img_enc.size(),
+        /*.mem_buffer =*/ state.buf_compute_img_enc.data(),
+        /*.no_alloc   =*/ true, // skip allocating as we use ggml_alloc to allocate exact memory requirements
+    };
+
+    struct ggml_context * ctx0   = ggml_init(ggml_params);
+    struct ggml_cgraph  * gf     = ggml_new_graph(ctx0);
+
+    struct ggml_tensor * inp = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_img_size, n_img_size, 3, 1);
+    ggml_allocr_alloc(state.allocr, inp);
+    if (!ggml_allocr_is_measure(state.allocr)) {
+        float * data = (float *) ggml_get_data(inp);
+
+        const int nx = img.nx;
+        const int ny = img.ny;
+        const int n  = nx*ny;
+
+        GGML_ASSERT(nx == n_img_size && ny == n_img_size);
+
+        for (int k = 0; k < 3; k++) {
+            for (int y = 0; y < ny; y++) {
+                for (int x = 0; x < nx; x++) {
+                    data[k*n + y*nx + x] = img.data[3*(y*nx + x) + k];
+                }
+            }
+        }
+    }
+
+    // ref: https://github.com/facebookresearch/segment-anything/blob/main/segment_anything/modeling/image_encoder.py#L392
+    struct ggml_tensor * cur = ggml_conv_2d_sk_p0(ctx0, enc.proj_w, inp);
+    cur = ggml_add_inplace(ctx0,
+            cur,
+            ggml_repeat(ctx0, enc.proj_b, cur));
+
+    // ref: https://github.com/facebookresearch/segment-anything/blob/main/segment_anything/modeling/image_encoder.py#L394
+    // keep in F32
+    cur = ggml_cont(ctx0,
+            ggml_permute(ctx0, cur, 1, 2, 0, 3));
+
+    // convert to F16
+    //cur = ggml_cpy(ctx0,
+    //        ggml_permute(ctx0, cur, 1, 2, 0, 3),
+    //        ggml_new_tensor_3d(ctx0, GGML_TYPE_F16, n_enc_state, n_img_embd, n_img_embd));
+
+    // ref: https://github.com/facebookresearch/segment-anything/blob/main/segment_anything/modeling/image_encoder.py#L108-L109
+    cur = ggml_add_inplace(ctx0, cur, enc.pe);
+
+    struct ggml_tensor * inpL = cur;
+
+    for (int il = 0; il < n_enc_layer; ++il) {
+        const auto & layer = enc.layers[il];
+
+        // norm
+        // ref: https://github.com/facebookresearch/segment-anything/blob/main/segment_anything/modeling/image_encoder.py#L168
+        {
+            cur = ggml_norm(ctx0, inpL, hparams.eps);
+
+            // cur = ln_0_w*cur + ln_0_b
+            cur = ggml_mul(ctx0, cur, layer.norm1_w);
+            cur = ggml_add_inplace(ctx0, cur, layer.norm1_b);
+        }
+
+        const int64_t w0 = cur->ne[1];
+        const int64_t h0 = cur->ne[2];
+
+        if (hparams.is_global_attn(il) == false) {
+            // local attention layer - apply window partition
+            // ref: https://github.com/facebookresearch/segment-anything/blob/main/segment_anything/modeling/image_encoder.py#L169-L172
+            cur = ggml_win_part(ctx0, cur, n_window_size);
+        }
+
+        const int64_t W = cur->ne[1];
+        const int64_t H = cur->ne[2];
+
+        // self-attention
+        {
+            cur = ggml_mul_mat(ctx0, layer.qkv_w, cur);
+            cur = ggml_add_inplace(ctx0, cur, layer.qkv_b);
+
+            // split qkv into separate tensors
+            // ref: https://github.com/facebookresearch/segment-anything/blob/main/segment_anything/modeling/image_encoder.py#L225-L229
+            const int B = cur->ne[3];
+
+            cur = ggml_reshape_4d(ctx0, cur, n_enc_state, 3, W*H, B);
+            cur = ggml_cont(ctx0, ggml_permute(ctx0, cur, 0, 3, 1, 2));
+
+            struct ggml_tensor * Q;
+            struct ggml_tensor * K;
+            struct ggml_tensor * V;
+
+            Q = ggml_view_3d   (ctx0, cur, n_enc_state, W*H, B, cur->nb[1], cur->nb[2], 0*cur->nb[3]);
+            Q = ggml_reshape_4d(ctx0, Q,   n_enc_head_dim, n_enc_head, W*H, B);
+            Q = ggml_cont      (ctx0, ggml_permute(ctx0, Q, 0, 2, 1, 3));
+            Q = ggml_reshape_3d(ctx0, Q,   n_enc_head_dim, W*H, B*n_enc_head);
+
+            K = ggml_view_3d   (ctx0, cur, n_enc_state, W*H, B, cur->nb[1], cur->nb[2], 1*cur->nb[3]);
+            K = ggml_reshape_4d(ctx0, K,   n_enc_head_dim, n_enc_head, W*H, B);
+            K = ggml_cont      (ctx0, ggml_permute(ctx0, K, 0, 2, 1, 3));
+            K = ggml_reshape_3d(ctx0, K,   n_enc_head_dim, W*H, B*n_enc_head);
+
+            V = ggml_view_3d   (ctx0, cur, n_enc_state, W*H, B, cur->nb[1], cur->nb[2], 2*cur->nb[3]);
+            V = ggml_reshape_4d(ctx0, V,   n_enc_head_dim, n_enc_head, W*H, B);
+            V = ggml_cont      (ctx0, ggml_permute(ctx0, V, 1, 2, 0, 3)); // transposed
+            V = ggml_reshape_3d(ctx0, V,   W*H, n_enc_head_dim, B*n_enc_head);
+
+            struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);
+
+            struct ggml_tensor * KQ_scaled =
+                ggml_scale_inplace(ctx0,
+                        KQ,
+                        ggml_new_f32(ctx0, 1.0f/sqrtf(n_enc_head_dim))
+                        );
+
+            struct ggml_tensor * rw = ggml_get_rel_pos(ctx0, layer.rel_pos_w, W, W);
+            struct ggml_tensor * rh = ggml_get_rel_pos(ctx0, layer.rel_pos_h, H, H);
+
+            struct ggml_tensor * q_r = ggml_reshape_4d(ctx0, Q, n_enc_head_dim, W, H, B*n_enc_head);
+
+            struct ggml_tensor * rel_w = ggml_cont(ctx0, ggml_permute(ctx0,
+                        ggml_mul_mat(ctx0,
+                            rw,
+                            ggml_cont(ctx0, ggml_permute(ctx0, q_r, 0, 2, 1, 3))),
+                        0, 2, 1, 3));
+            struct ggml_tensor * rel_h = ggml_mul_mat(ctx0, rh, q_r);
+
+            struct ggml_tensor * attn = ggml_add_rel_pos_inplace(ctx0, KQ_scaled, rel_w, rel_h);
+
+            struct ggml_tensor * KQ_soft_max = ggml_soft_max_inplace(ctx0, attn);
+
+            struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max);
+
+            cur =
+                ggml_reshape_4d(ctx0,
+                        ggml_cont(ctx0,
+                            ggml_permute(ctx0,
+                                ggml_reshape_4d(ctx0, KQV, n_enc_head_dim, W*H, n_enc_head, B),
+                                0, 2, 1, 3)),
+                        n_enc_state, W, H, B);
+
+            cur = ggml_mul_mat(ctx0, layer.proj_w, cur);
+            cur = ggml_add_inplace(ctx0, cur, layer.proj_b);
+        }
+
+        if (hparams.is_global_attn(il) == false) {
+            // local attention layer - reverse window partition
+            cur = ggml_win_unpart(ctx0, cur, w0, h0, n_window_size);
+        }
+
+        cur = ggml_add_inplace(ctx0, cur, inpL);
+
+        struct ggml_tensor * inpFF = cur;
+
+        // feed-forward network
+        {
+            // norm
+            {
+                cur = ggml_norm(ctx0, inpFF, hparams.eps);
+
+                // cur = mlp_ln_w*cur + mlp_ln_b
+                cur = ggml_mul(ctx0, cur, layer.norm2_w);
+                cur = ggml_add_inplace(ctx0, cur, layer.norm2_b);
+            }
+
+            // fully connected
+            cur = ggml_mul_mat(ctx0, layer.mlp_lin1_w, cur);
+            cur = ggml_add_inplace(ctx0, cur, layer.mlp_lin1_b);
+
+            // GELU activation
+            cur = ggml_gelu(ctx0, cur);
+
+            // projection
+            cur = ggml_mul_mat(ctx0, layer.mlp_lin2_w, cur);
+            cur = ggml_add_inplace(ctx0, cur, layer.mlp_lin2_b);
+        }
+
+        inpL = ggml_add(ctx0, cur, inpFF);
+    }
+
+    cur = ggml_cont(ctx0, ggml_permute(ctx0, inpL, 2, 0, 1, 3));
+
+    cur = ggml_conv_2d_sk_p0(ctx0, enc.neck_conv_0, cur);
+
+    cur = sam_layer_norm_2d(ctx0, cur, n_enc_out_chans, enc.neck_norm_0_w, enc.neck_norm_0_b, hparams.eps);
+
+    cur = ggml_conv_2d_s1_ph(ctx0, enc.neck_conv_1, cur);
+
+    cur = sam_layer_norm_2d(ctx0, cur, n_enc_out_chans, enc.neck_norm_1_w, enc.neck_norm_1_b, hparams.eps);
+
+    cur = ggml_cpy(ctx0, cur, state.embd_img);
+
+    ggml_build_forward_expand(gf, cur);
+    ggml_disconnect_node_from_graph(state.embd_img);
+
+    //ggml_graph_print(&gf);
+
+    ggml_free(ctx0);
+
+    return gf;
+}
+
+
+struct prompt_encoder_result {
+    struct ggml_tensor * embd_prompt_sparse = {};
+    struct ggml_tensor * embd_prompt_dense = {};
+};
+
+// encode a prompt
+//
+// - points
+// - boxes
+// - masks
+//
+// TODO: currently just encode a single point for simplicity
+//
+prompt_encoder_result sam_encode_prompt(
+        const sam_model     & model,
+        struct ggml_context * ctx0,
+        struct ggml_cgraph  * gf,
+                  sam_state & state,
+                        int   nx,
+                        int   ny,
+                  sam_point   point) {
+
+    const auto & hparams = model.hparams;
+    const auto & enc = model.enc_prompt;
+
+    // transform points
+    // ref: https://github.com/facebookresearch/segment-anything/blob/main/segment_anything/automatic_mask_generator.py#L276
+    {
+        const int nmax = std::max(nx, ny);
+
+        const float scale = hparams.n_img_size() / (float) nmax;
+
+        const int nx_new = int(nx*scale + 0.5f);
+        const int ny_new = int(ny*scale + 0.5f);
+
+        point.x = point.x*(float(nx_new)/nx) + 0.5f;
+        point.y = point.y*(float(ny_new)/ny) + 0.5f;
+    }
+
+    struct ggml_tensor * inp = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, 2, 2);
+
+    ggml_allocr_alloc(state.allocr, inp);
+    if (!ggml_allocr_is_measure(state.allocr)) {
+        // set the input by converting the [0, 1] coordinates to [-1, 1]
+        float * data = (float *) inp->data;
+
+        data[0] = 2.0f*(point.x / hparams.n_img_size()) - 1.0f;
+        data[1] = 2.0f*(point.y / hparams.n_img_size()) - 1.0f;
+
+        // padding
+        // ref: https://github.com/facebookresearch/segment-anything/blob/main/segment_anything/modeling/prompt_encoder.py#L81-L85
+        data[2] = 2.0f*(0.0f) - 1.0f;
+        data[3] = 2.0f*(0.0f) - 1.0f;
+    }
+
+    struct ggml_tensor * cur = ggml_mul_mat(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, enc.pe)), inp);
+
+    cur = ggml_scale(ctx0, cur, ggml_new_f32(ctx0, float(2.0*M_PI)));
+
+    // concat
+    // ref: https://github.com/facebookresearch/segment-anything/blob/main/segment_anything/modeling/prompt_encoder.py#L192
+    {
+        struct ggml_tensor * t_sin = ggml_map_custom1(ctx0, cur, ggml_sam_sin, GGML_N_TASKS_MAX, NULL);
+        struct ggml_tensor * t_cos = ggml_map_custom1(ctx0, cur, ggml_sam_cos, GGML_N_TASKS_MAX, NULL);
+
+        cur = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, t_sin->ne[0] + t_cos->ne[0], cur->ne[1]);
+
+        ggml_build_forward_expand(gf, ggml_cpy(ctx0, t_sin, ggml_view_2d(ctx0, cur, t_sin->ne[0], t_sin->ne[1], cur->nb[1], 0)));
+        ggml_build_forward_expand(gf, ggml_cpy(ctx0, t_cos, ggml_view_2d(ctx0, cur, t_sin->ne[0], t_sin->ne[1], cur->nb[1], t_sin->nb[1])));
+
+        // overwrite label == -1 with not_a_point_embed.weight
+        // ref: https://github.com/facebookresearch/segment-anything/blob/main/segment_anything/modeling/prompt_encoder.py#L86
+        // TODO: extend for multiple points
+        ggml_build_forward_expand(gf, ggml_cpy(ctx0, enc.not_a_pt_embd_w, ggml_view_2d(ctx0, cur, cur->ne[0], 1, cur->nb[1], cur->nb[1])));
+    }
+
+    // add point_embeddings[1] to label == 1
+    // ref: https://github.com/facebookresearch/segment-anything/blob/main/segment_anything/modeling/prompt_encoder.py#L90
+    struct ggml_tensor * v = ggml_view_2d(ctx0, cur, cur->ne[0], 1, cur->nb[1], 0);
+    ggml_build_forward_expand(gf, ggml_cpy(ctx0, ggml_add_inplace(ctx0, v, enc.pt_embd[1]), v));
+
+    struct ggml_tensor * embd_prompt_sparse = cur;
+    ggml_build_forward_expand(gf, embd_prompt_sparse);
+
+    struct ggml_tensor * embd_prompt_dense = ggml_repeat(ctx0,
+            ggml_cont(ctx0,
+                ggml_view_3d(ctx0, enc.no_mask_embd_w,
+                    1, 1, enc.no_mask_embd_w->ne[0], enc.no_mask_embd_w->nb[0], enc.no_mask_embd_w->nb[0], 0)),
+            ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, hparams.n_img_embd(), hparams.n_img_embd(), hparams.n_enc_out_chans));
+
+    ggml_build_forward_expand(gf, embd_prompt_dense);
+
+    //printf("used_mem = %zu\n", ggml_used_mem(ctx0));
+
+    prompt_encoder_result res;
+    res.embd_prompt_sparse = embd_prompt_sparse;
+    res.embd_prompt_dense  = embd_prompt_dense;
+    return res;
+}
+
+struct ggml_tensor* sam_decode_mask_transformer_attn(
+    const sam_layer_dec_transformer_attn & attn,
+                      struct ggml_tensor * queries,
+                      struct ggml_tensor * keys,
+                      struct ggml_tensor * values,
+                     struct ggml_context * ctx0,
+                         const sam_model & model) {
+    const auto & hparams = model.hparams;
+    const int n_head = hparams.n_dec_heads;
+
+    struct ggml_tensor * Qcur = {};
+    struct ggml_tensor * Kcur = {};
+    struct ggml_tensor * Vcur = {};
+
+    Qcur = ggml_mul_mat(ctx0, attn.q_w, queries);
+    Qcur = ggml_add_inplace(ctx0, Qcur, attn.q_b);
+
+    Kcur = ggml_mul_mat(ctx0, attn.k_w, keys);
+    Kcur = ggml_add_inplace(ctx0, Kcur, attn.k_b);
+
+    Vcur = ggml_mul_mat(ctx0, attn.v_w, values);
+    Vcur = ggml_add_inplace(ctx0, Vcur, attn.v_b);
+
+    struct ggml_tensor * Q = {};
+    struct ggml_tensor * K = {};
+    struct ggml_tensor * V = {};
+
+    Q = ggml_reshape_4d(ctx0, Qcur, Qcur->ne[0]/n_head, n_head, Qcur->ne[1], Qcur->ne[2]);
+    Q = ggml_cont(ctx0, ggml_permute(ctx0, Q, 0, 2, 1, 3));
+
+    K = ggml_reshape_4d(ctx0, Kcur, Kcur->ne[0]/n_head, n_head, Kcur->ne[1], Kcur->ne[2]);
+    K = ggml_cont(ctx0, ggml_permute(ctx0, K, 0, 2, 1, 3));
+
+    V = ggml_reshape_4d(ctx0, Vcur, Vcur->ne[0]/n_head, n_head, Vcur->ne[1], Vcur->ne[2]);
+    V = ggml_cont(ctx0, ggml_permute(ctx0, V, 0, 2, 1, 3));
+
+    // Q * K
+    struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);
+
+    struct ggml_tensor * KQ_scaled =
+        ggml_scale_inplace(ctx0,
+                KQ,
+                ggml_new_f32(ctx0, 1.0f/sqrt(float(Q->ne[0]))));
+
+    struct ggml_tensor * KQ_soft_max = ggml_soft_max_inplace(ctx0, KQ_scaled);
+
+    struct ggml_tensor * KQV = ggml_mul_mat(ctx0, KQ_soft_max, ggml_cont(ctx0, ggml_transpose(ctx0, V)));
+
+    struct ggml_tensor * KQV_merged = ggml_cont(ctx0, ggml_transpose(ctx0, KQV));
+    KQV_merged = ggml_cont(ctx0, ggml_permute(ctx0, KQV_merged, 0, 2, 1, 3));
+    KQV_merged = ggml_reshape_3d(ctx0, KQV_merged, KQV_merged->ne[0]*KQV_merged->ne[1], KQV_merged->ne[2], KQV_merged->ne[3]);
+    KQV_merged = ggml_mul_mat(ctx0, attn.out_w, KQV_merged);
+    KQV_merged = ggml_add_inplace(ctx0, KQV_merged, attn.out_b);
+
+    return KQV_merged;
+}
+
+struct ggml_tensor * sam_decode_mask_mlp_relu_3(
+     struct ggml_tensor * in,
+     struct ggml_tensor * w_0,
+     struct ggml_tensor * b_0,
+     struct ggml_tensor * w_1,
+     struct ggml_tensor * b_1,
+     struct ggml_tensor * w_2,
+     struct ggml_tensor * b_2,
+    struct ggml_context * ctx0) {
+
+    struct ggml_tensor * cur = {};
+    cur = ggml_mul_mat(ctx0, w_0, in);
+    cur = ggml_add_inplace(ctx0, cur, b_0);
+
+    cur = ggml_relu_inplace(ctx0, cur);
+
+    cur = ggml_mul_mat(ctx0, w_1, cur);
+    cur = ggml_add_inplace(ctx0, cur, b_1);
+
+    cur = ggml_relu_inplace(ctx0, cur);
+
+    cur = ggml_mul_mat(ctx0, w_2, cur);
+    cur = ggml_add_inplace(ctx0, cur, b_2);
+
+    return cur;
+}
+
+bool sam_decode_mask(
+                    const sam_model & model,
+        const prompt_encoder_result & prompt,
+                 struct ggml_tensor * pe_img,
+                struct ggml_context * ctx0,
+                struct ggml_cgraph  * gf,
+                          sam_state & state) {
+
+    const auto & hparams = model.hparams;
+    const auto & dec = model.dec;
+    const int n_img_embd = hparams.n_img_embd();
+
+    struct ggml_tensor * tokens = {};
+    {
+        // Concatenate output tokens
+        // ref: https://github.com/facebookresearch/segment-anything/blob/6fdee8f2727f4506cfbbe553e23b895e27956588/segment_anything/modeling/mask_decoder.py#L120
+        const auto& sparse = prompt.embd_prompt_sparse;
+
+        tokens = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, dec.iou_token_w->ne[0], dec.iou_token_w->ne[1] + dec.mask_tokens_w->ne[1] + sparse->ne[1], sparse->ne[2]);
+
+        const size_t offsets[3] = { 0, dec.iou_token_w->ne[1]*tokens->nb[1], dec.iou_token_w->ne[1]*tokens->nb[1] + dec.mask_tokens_w->ne[1]*tokens->nb[1] };
+        ggml_build_forward_expand(gf, ggml_cpy(ctx0, dec.iou_token_w,   ggml_view_2d(ctx0, tokens, tokens->ne[0], dec.iou_token_w->ne[1],   tokens->nb[1], offsets[0])));
+        ggml_build_forward_expand(gf, ggml_cpy(ctx0, dec.mask_tokens_w, ggml_view_2d(ctx0, tokens, tokens->ne[0], dec.mask_tokens_w->ne[1], tokens->nb[1], offsets[1])));
+        ggml_build_forward_expand(gf, ggml_cpy(ctx0, sparse,            ggml_view_2d(ctx0, tokens, tokens->ne[0], sparse->ne[1],            tokens->nb[1], offsets[2])));
+        // TODO: Sparse prompt embeddings can have more than one point
+    }
+
+
+    struct ggml_tensor * src = {};
+    struct ggml_tensor * pos_src = {};
+    int srcNE[4] = { 0, 0, 0, 0 };
+    {
+        // Expand per-image data in the batch direction to be per-mask
+        // ref: https://github.com/facebookresearch/segment-anything/blob/6fdee8f2727f4506cfbbe553e23b895e27956588/segment_anything/modeling/mask_decoder.py#L125
+        src = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, state.embd_img->ne[0], state.embd_img->ne[1], state.embd_img->ne[2], tokens->ne[2]);
+
+        src = ggml_add(ctx0,
+            ggml_repeat(ctx0,
+                state.embd_img,
+                src),
+            prompt.embd_prompt_dense);
+
+        srcNE[0] = src->ne[0];
+        srcNE[1] = src->ne[1];
+        srcNE[2] = src->ne[2];
+        srcNE[3] = src->ne[3];
+
+        // flatten & permute
+        // ref: https://github.com/facebookresearch/segment-anything/blob/6fdee8f2727f4506cfbbe553e23b895e27956588/segment_anything/modeling/transformer.py#L83
+        src = ggml_cont(ctx0, ggml_permute(ctx0,
+            ggml_view_3d(ctx0,
+                src,
+                src->ne[0]*src->ne[1],
+                src->ne[2],
+                src->ne[3],
+                src->nb[2],
+                src->nb[3],
+                0),
+            1, 0, 2, 3));
+
+        pos_src = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, pe_img->ne[0], pe_img->ne[1], pe_img->ne[2], tokens->ne[2]);
+        pos_src = ggml_repeat(ctx0,
+            pe_img,
+            pos_src);
+
+        // flatten & permute
+        // ref: https://github.com/facebookresearch/segment-anything/blob/6fdee8f2727f4506cfbbe553e23b895e27956588/segment_anything/modeling/transformer.py#L83
+        pos_src = ggml_cont(ctx0, ggml_permute(ctx0,
+            ggml_view_3d(ctx0,
+                pos_src,
+                pos_src->ne[0]*pos_src->ne[1],
+                pos_src->ne[2],
+                pos_src->ne[3],
+                pos_src->nb[2],
+                pos_src->nb[3],
+                0),
+            1, 0, 2, 3));
+    }
+
+    struct ggml_tensor * queries = tokens;
+    struct ggml_tensor * keys = src;
+    {
+        // Run the transformer
+        // ref: https://github.com/facebookresearch/segment-anything/blob/6fdee8f2727f4506cfbbe553e23b895e27956588/segment_anything/modeling/transformer.py#L62
+        for (int i = 0; i < int(model.dec.transformer_layers.size()); ++i) {
+            const auto& tfm_layer = model.dec.transformer_layers[i];
+
+            // Self attention block
+            // ref: https://github.com/facebookresearch/segment-anything/blob/6fdee8f2727f4506cfbbe553e23b895e27956588/segment_anything/modeling/transformer.py#L154
+            const bool skip_first_layer_pe = i == 0;
+            if (skip_first_layer_pe) {
+                queries = sam_decode_mask_transformer_attn(tfm_layer.self_attn, queries, queries, queries, ctx0, model);
+            }
+            else {
+                struct ggml_tensor * q_0 = ggml_add(ctx0, queries, tokens);
+
+                struct ggml_tensor * self_attn = sam_decode_mask_transformer_attn(tfm_layer.self_attn, q_0, q_0, queries, ctx0, model);
+                queries = ggml_add(ctx0, queries, self_attn);
+            }
+
+            queries = ggml_norm(ctx0, queries, hparams.eps_decoder_transformer);
+            queries = ggml_add_inplace(ctx0,
+                    ggml_mul(ctx0, queries, tfm_layer.norm1_w),
+                    tfm_layer.norm1_b);
+
+            // Cross attention block, tokens attending to image embedding
+            // ref: https://github.com/facebookresearch/segment-anything/blob/6fdee8f2727f4506cfbbe553e23b895e27956588/segment_anything/modeling/transformer.py#L163
+            struct ggml_tensor * q_1 = ggml_add(ctx0, queries, tokens);
+            struct ggml_tensor * k_1 = ggml_add(ctx0, keys, pos_src);
+
+            struct ggml_tensor * cross_attn_token_to_img = sam_decode_mask_transformer_attn(tfm_layer.cross_attn_token_to_img, q_1, k_1, keys, ctx0, model);
+
+            queries = ggml_add_inplace(ctx0, queries, cross_attn_token_to_img);
+            queries = ggml_norm_inplace(ctx0, queries, hparams.eps_decoder_transformer);
+            queries = ggml_add_inplace(ctx0,
+                    ggml_mul(ctx0, queries, tfm_layer.norm2_w),
+                    tfm_layer.norm2_b);
+
+            // MLP block
+            // ref: https://github.com/facebookresearch/segment-anything/blob/6fdee8f2727f4506cfbbe553e23b895e27956588/segment_anything/modeling/transformer.py#L170
+            struct ggml_tensor * mlp_out = ggml_mul_mat(ctx0,
+                tfm_layer.mlp_lin1_w,
+                queries);
+
+            mlp_out = ggml_add_inplace(ctx0, mlp_out, tfm_layer.mlp_lin1_b);
+
+            // RELU activation
+            mlp_out = ggml_relu_inplace(ctx0, mlp_out);
+            mlp_out = ggml_mul_mat(ctx0, tfm_layer.mlp_lin2_w, mlp_out);
+
+            mlp_out = ggml_add_inplace(ctx0, mlp_out, tfm_layer.mlp_lin2_b);
+
+            queries = ggml_add_inplace(ctx0, queries, mlp_out);
+            queries = ggml_norm_inplace(ctx0, queries, hparams.eps_decoder_transformer);
+            queries = ggml_add_inplace(ctx0,
+                    ggml_mul(ctx0, queries, tfm_layer.norm3_w),
+                    tfm_layer.norm3_b);
+
+            // Cross attention block, image embedding attending to tokens
+            // ref: https://github.com/facebookresearch/segment-anything/blob/6fdee8f2727f4506cfbbe553e23b895e27956588/segment_anything/modeling/transformer.py#L175
+            struct ggml_tensor * q_2 = ggml_add(ctx0, queries, tokens);
+            struct ggml_tensor * k_2 = ggml_add(ctx0, keys, pos_src);
+
+            struct ggml_tensor * cross_attn_img_to_token = sam_decode_mask_transformer_attn(tfm_layer.cross_attn_img_to_token, k_2, q_2, queries, ctx0, model);
+            keys = ggml_add_inplace(ctx0, keys, cross_attn_img_to_token);
+            keys = ggml_norm_inplace(ctx0, keys, hparams.eps_decoder_transformer);
+            keys = ggml_add_inplace(ctx0,
+                    ggml_mul(ctx0, keys, tfm_layer.norm4_w),
+                    tfm_layer.norm4_b);
+        }
+
+        // Apply the final attention layer from the points to the image
+        // ref: https://github.com/facebookresearch/segment-anything/blob/6fdee8f2727f4506cfbbe553e23b895e27956588/segment_anything/modeling/transformer.py#L99
+        struct ggml_tensor * q = ggml_add(ctx0, queries, tokens);
+        struct ggml_tensor * k = ggml_add(ctx0, keys, pos_src);
+
+        struct ggml_tensor * final_attn_token_to_img = sam_decode_mask_transformer_attn(dec.transformer_final_attn_token_to_img, q, k, keys, ctx0, model);
+
+        queries = ggml_add_inplace(ctx0, queries, final_attn_token_to_img);
+        queries = ggml_norm_inplace(ctx0, queries, hparams.eps_decoder_transformer);
+        queries = ggml_add_inplace(ctx0,
+                ggml_mul(ctx0, queries, dec.transformer_norm_final_w),
+                dec.transformer_norm_final_b);
+    }
+
+
+    struct ggml_tensor * iou_pred = ggml_view_2d(ctx0, queries, queries->ne[0], queries->ne[2], queries->nb[2], 0);
+    const int num_mask_tokens = 4; // num_multimask_outputs + 1
+    struct ggml_tensor * mask_tokens_out = ggml_view_3d(ctx0, queries, queries->ne[0], num_mask_tokens, queries->ne[2], queries->nb[1], num_mask_tokens*queries->nb[1], queries->nb[1]);
+
+    // Upscale mask embeddings and predict masks using the mask tokens
+    // ref: https://github.com/facebookresearch/segment-anything/blob/6fdee8f2727f4506cfbbe553e23b895e27956588/segment_anything/modeling/mask_decoder.py#L136
+    keys = ggml_cont(ctx0, ggml_transpose(ctx0, keys));
+    keys = ggml_view_4d(ctx0, keys, srcNE[0], srcNE[1], srcNE[2], srcNE[3], srcNE[0]*keys->nb[0], keys->nb[1], keys->nb[2], 0);
+    // ggml_build_forward_expand(gf, keys);
+    struct ggml_tensor * upscaled_embedding = {};
+    {
+        // ConvTranspose2d
+        keys = ggml_conv_transpose_2d_p0(ctx0, dec.output_upscaling_0_w, keys, 2);
+        ggml_allocr_alloc(state.allocr, keys); // TODO: This alloc shouldn't be needed
+        keys = ggml_add_inplace(ctx0, keys, ggml_repeat(ctx0,
+                                     ggml_reshape_3d(ctx0, dec.output_upscaling_0_b, 1, 1, dec.output_upscaling_0_b->ne[0]),
+                                     keys));
+
+        keys = sam_layer_norm_2d(ctx0, keys, n_img_embd, dec.output_upscaling_1_w, dec.output_upscaling_1_b, hparams.eps);
+
+        // GELU activation
+        keys = ggml_gelu_inplace(ctx0, keys);
+
+        // ConvTranspose2d
+        keys = ggml_conv_transpose_2d_p0(ctx0, dec.output_upscaling_3_w, keys, 2);
+        ggml_allocr_alloc(state.allocr, keys); // TODO: This alloc shouldn't be needed
+        keys = ggml_add_inplace(ctx0, ggml_repeat(ctx0,
+                                ggml_reshape_3d(ctx0, dec.output_upscaling_3_b, 1, 1, dec.output_upscaling_3_b->ne[0]),
+                                keys), keys);
+        // GELU activation
+        keys = ggml_gelu_inplace(ctx0, keys);
+        upscaled_embedding = ggml_reshape_3d(ctx0, keys, keys->ne[0]*keys->ne[1], keys->ne[2], keys->ne[3]);
+        upscaled_embedding = ggml_cont(ctx0, ggml_transpose(ctx0, upscaled_embedding)); // TODO: Shouldn't be needed
+    }
+
+    struct ggml_tensor * hyper_in = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_img_embd/2, num_mask_tokens, mask_tokens_out->ne[2]);
+
+    for (int i = 0; i < num_mask_tokens; ++i) {
+        const auto& mlp = dec.output_hypernet_mlps[i];
+        struct ggml_tensor * in = ggml_view_2d(ctx0, mask_tokens_out, mask_tokens_out->ne[0], mask_tokens_out->ne[2], mask_tokens_out->nb[1], i*mask_tokens_out->nb[1]);
+        struct ggml_tensor * out = sam_decode_mask_mlp_relu_3(in, mlp.w_0, mlp.b_0, mlp.w_1, mlp.b_1, mlp.w_2, mlp.b_2, ctx0);
+        ggml_build_forward_expand(gf, ggml_cpy(ctx0, out, ggml_view_2d(ctx0, hyper_in, hyper_in->ne[0], hyper_in->ne[2], hyper_in->nb[1], i*hyper_in->nb[1])));
+    }
+
+    struct ggml_tensor * masks = ggml_mul_mat(ctx0, hyper_in, upscaled_embedding);
+    masks = ggml_cont(ctx0, ggml_transpose(ctx0, masks)); // TODO: Shouldn't be needed
+    masks = ggml_reshape_4d(ctx0, masks, keys->ne[0], keys->ne[1], masks->ne[1], keys->ne[3]);
+
+    // Generate mask quality predictions
+    // ref: https://github.com/facebookresearch/segment-anything/blob/6fdee8f2727f4506cfbbe553e23b895e27956588/segment_anything/modeling/mask_decoder.py#L146
+    iou_pred = sam_decode_mask_mlp_relu_3(iou_pred, dec.iou_prediction_head_0_w, dec.iou_prediction_head_0_b, dec.iou_prediction_head_1_w, dec.iou_prediction_head_1_b, dec.iou_prediction_head_2_w, dec.iou_prediction_head_2_b, ctx0);
+
+    // Select the correct mask or masks for output
+    // ref: https://github.com/facebookresearch/segment-anything/blob/6fdee8f2727f4506cfbbe553e23b895e27956588/segment_anything/modeling/mask_decoder.py#L101
+    iou_pred = ggml_cpy(state.ctx, ggml_view_1d(ctx0, iou_pred, iou_pred->ne[0] - 1, iou_pred->nb[0]), state.iou_predictions);
+    masks = ggml_view_4d(ctx0, masks, masks->ne[0], masks->ne[1], masks->ne[2] - 1, masks->ne[3],
+                                      masks->nb[1], masks->nb[2], masks->nb[3], masks->nb[2] /* offset*/);
+    masks = ggml_cpy(state.ctx, masks, state.low_res_masks);
+
+    ggml_build_forward_expand(gf, masks);
+    ggml_build_forward_expand(gf, iou_pred);
+
+    ggml_disconnect_node_from_graph(state.low_res_masks);
+    ggml_disconnect_node_from_graph(state.iou_predictions);
+
+    return true;
+}
+
+bool sam_write_masks(const sam_hparams& hparams, int nx, int ny, const sam_state & state) {
+    if (state.low_res_masks->ne[2] == 0) return true;
+    if (state.low_res_masks->ne[2] != state.iou_predictions->ne[0]) {
+        printf("Error: number of masks (%d) does not match number of iou predictions (%d)\n", (int)state.low_res_masks->ne[2], (int)state.iou_predictions->ne[0]);
+        return false;
+    }
+
+    const int n_img_size = hparams.n_img_size();
+    const float mask_threshold = hparams.mask_threshold;
+    const float iou_threshold = hparams.iou_threshold;
+    const float stability_score_threshold = hparams.stability_score_threshold;
+    const float intersection_threshold = mask_threshold + hparams.stability_score_offset;
+    const float union_threshold = mask_threshold - hparams.stability_score_offset;
+
+    const int ne0 = state.low_res_masks->ne[0];
+    const int ne1 = state.low_res_masks->ne[1];
+    const int ne2 = state.low_res_masks->ne[2];
+
+    // Remove padding and upscale masks to the original image size.
+    // ref: https://github.com/facebookresearch/segment-anything/blob/efeab7296ab579d4a261e554eca80faf6b33924a/segment_anything/modeling/sam.py#L140
+
+    const float preprocess_scale = std::max(nx, ny) / float(n_img_size);
+    const int cropped_nx = int(nx / preprocess_scale + 0.5f);
+    const int cropped_ny = int(ny / preprocess_scale + 0.5f);
+
+    const float scale_x_1 = (float)ne0 / (float)n_img_size;
+    const float scale_y_1 = (float)ne1 / (float)n_img_size;
+
+    const float scale_x_2 = float(cropped_nx) / float(nx);
+    const float scale_y_2 = float(cropped_ny) / float(ny);
+
+    const auto iou_data = (float*)state.iou_predictions->data;
+
+    for (int i = 0; i < ne2; ++i) {
+        if (iou_threshold > 0.f && iou_data[i] < iou_threshold) {
+            printf("Skipping mask %d with iou %f below threshold %f\n", i, iou_data[i], iou_threshold);
+            continue; // Filtering masks with iou below the threshold
+        }
+
+        std::vector<float> mask_data(n_img_size*n_img_size);
+        {
+            const float* data = (float *) state.low_res_masks->data + i*ne0*ne1;
+
+            for (int iy = 0; iy < n_img_size; ++iy) {
+                for (int ix = 0; ix < n_img_size; ++ix) {
+                    const float sx = std::max(scale_x_1*(ix + 0.5f) - 0.5f, 0.0f);
+                    const float sy = std::max(scale_y_1*(iy + 0.5f) - 0.5f, 0.0f);
+
+                    const int x0 = std::max(0, (int)sx);
+                    const int y0 = std::max(0, (int)sy);
+
+                    const int x1 = std::min(x0 + 1, ne0 - 1);
+                    const int y1 = std::min(y0 + 1, ne1 - 1);
+
+                    const float dx = sx - x0;
+                    const float dy = sy - y0;
+
+                    const int j00 = y0*ne0 + x0;
+                    const int j01 = y0*ne0 + x1;
+                    const int j10 = y1*ne0 + x0;
+                    const int j11 = y1*ne0 + x1;
+
+                    const float v00 = data[j00];
+                    const float v01 = data[j01];
+                    const float v10 = data[j10];
+                    const float v11 = data[j11];
+
+                    const float v0 = (1-dx)*v00 + dx*v01;
+                    const float v1 = (1-dx)*v10 + dx*v11;
+
+                    const float v = (1-dy)*v0 + dy*v1;
+
+                    mask_data[iy*n_img_size + ix] = v;
+                }
+            }
+        }
+
+        int intersections = 0;
+        int unions = 0;
+        sam_image_u8 res;
+        int min_iy = ny;
+        int max_iy = 0;
+        int min_ix = nx;
+        int max_ix = 0;
+        {
+            const float* data = mask_data.data();
+
+            res.nx = nx;
+            res.ny = ny;
+            res.data.resize(nx*ny);
+
+            for (int iy = 0; iy < ny; ++iy) {
+                for (int ix = 0; ix < nx; ++ix) {
+                    const float sx = std::max(scale_x_2*(ix + 0.5f) - 0.5f, 0.0f);
+                    const float sy = std::max(scale_y_2*(iy + 0.5f) - 0.5f, 0.0f);
+
+                    const int x0 = std::max(0, (int)sx);
+                    const int y0 = std::max(0, (int)sy);
+
+                    const int x1 = std::min(x0 + 1, cropped_nx - 1);
+                    const int y1 = std::min(y0 + 1, cropped_ny - 1);
+
+                    const float dx = sx - x0;
+                    const float dy = sy - y0;
+
+                    const int j00 = y0*n_img_size + x0;
+                    const int j01 = y0*n_img_size + x1;
+                    const int j10 = y1*n_img_size + x0;
+                    const int j11 = y1*n_img_size + x1;
+
+                    const float v00 = data[j00];
+                    const float v01 = data[j01];
+                    const float v10 = data[j10];
+                    const float v11 = data[j11];
+
+                    const float v0 = (1-dx)*v00 + dx*v01;
+                    const float v1 = (1-dx)*v10 + dx*v11;
+
+                    const float v = (1-dy)*v0 + dy*v1;
+
+                    if (v > intersection_threshold) {
+                        intersections++;
+                    }
+                    if (v > union_threshold) {
+                        unions++;
+                    }
+                    if (v > mask_threshold) {
+                        min_iy = std::min(min_iy, iy);
+                        max_iy = std::max(max_iy, iy);
+                        min_ix = std::min(min_ix, ix);
+                        max_ix = std::max(max_ix, ix);
+
+                        res.data[iy*nx + ix] = 255;
+                    }
+                }
+            }
+        }
+
+        const float stability_score = float(intersections) / float(unions);
+        if (stability_score_threshold > 0.f && stability_score < stability_score_threshold) {
+            printf("Skipping mask %d with stability score %f below threshold %f\n", i, stability_score, stability_score_threshold);
+            continue; // Filtering masks with stability score below the threshold
+        }
+
+        printf("Mask %d: iou = %f, stability_score = %f, bbox (%d, %d), (%d, %d)\n",
+                i, iou_data[i], stability_score, min_ix, max_ix, min_iy, max_iy);
+
+        std::string filename = "mask_out_" + std::to_string(i) + ".png";
+        if (!stbi_write_png(filename.c_str(), res.nx, res.ny, 1, res.data.data(), res.nx)) {
+            printf("%s: failed to write mask %s\n", __func__, filename.c_str());
+            return false;
+        }
+    }
+
+
+    return true;
+}
+
+struct ggml_cgraph  * sam_build_fast_graph(
+        const sam_model     & model,
+                  sam_state & state,
+                        int   nx,
+                        int   ny,
+                  sam_point   point) {
+
+    struct ggml_init_params ggml_params = {
+        /*.mem_size   =*/ state.buf_compute_fast.size(),
+        /*.mem_buffer =*/ state.buf_compute_fast.data(),
+        /*.no_alloc   =*/ true, // skip allocating as we use ggml_alloc to allocate exact memory requirements
+    };
+
+    struct ggml_context * ctx0   = ggml_init(ggml_params);
+    struct ggml_cgraph  * gf     = ggml_new_graph(ctx0);
+
+    prompt_encoder_result enc_res = sam_encode_prompt(model, ctx0, gf, state, nx, ny, point);
+    if (!enc_res.embd_prompt_sparse || !enc_res.embd_prompt_dense) {
+        fprintf(stderr, "%s: failed to encode prompt\n", __func__);
+        return {};
+    }
+
+    struct ggml_tensor * pe_img_dense = sam_fill_dense_pe(model, ctx0, gf, state);
+    if (!pe_img_dense) {
+        fprintf(stderr, "%s: failed to get dense positional encoding\n", __func__);
+        return {};
+    }
+
+    if (!sam_decode_mask(model, enc_res, pe_img_dense, ctx0, gf, state)) {
+         fprintf(stderr, "%s: failed to decode mask\n", __func__);
+         return {};
+    }
+
+    ggml_free(ctx0);
+
+    return gf;
+}
+struct sam_params {
+    int32_t seed      = -1; // RNG seed
+    int32_t n_threads = std::min(4, (int32_t) std::thread::hardware_concurrency());
+
+    std::string model     = "models/sam-vit-b/ggml-model-f16.bin"; // model path
+    std::string fname_inp = "img.jpg";
+    std::string fname_out = "img.out";
+};
+
+void sam_print_usage(int argc, char ** argv, const sam_params & params) {
+    fprintf(stderr, "usage: %s [options]\n", argv[0]);
+    fprintf(stderr, "\n");
+    fprintf(stderr, "options:\n");
+    fprintf(stderr, "  -h, --help            show this help message and exit\n");
+    fprintf(stderr, "  -s SEED, --seed SEED  RNG seed (default: -1)\n");
+    fprintf(stderr, "  -t N, --threads N     number of threads to use during computation (default: %d)\n", params.n_threads);
+    fprintf(stderr, "  -m FNAME, --model FNAME\n");
+    fprintf(stderr, "                        model path (default: %s)\n", params.model.c_str());
+    fprintf(stderr, "  -i FNAME, --inp FNAME\n");
+    fprintf(stderr, "                        input file (default: %s)\n", params.fname_inp.c_str());
+    fprintf(stderr, "  -o FNAME, --out FNAME\n");
+    fprintf(stderr, "                        output file (default: %s)\n", params.fname_out.c_str());
+    fprintf(stderr, "\n");
+}
+
+bool sam_params_parse(int argc, char ** argv, sam_params & params) {
+    for (int i = 1; i < argc; i++) {
+        std::string arg = argv[i];
+
+        if (arg == "-s" || arg == "--seed") {
+            params.seed = std::stoi(argv[++i]);
+        } else if (arg == "-t" || arg == "--threads") {
+            params.n_threads = std::stoi(argv[++i]);
+        } else if (arg == "-m" || arg == "--model") {
+            params.model = argv[++i];
+        } else if (arg == "-i" || arg == "--inp") {
+            params.fname_inp = argv[++i];
+        } else if (arg == "-o" || arg == "--out") {
+            params.fname_out = argv[++i];
+        } else if (arg == "-h" || arg == "--help") {
+            sam_print_usage(argc, argv, params);
+            exit(0);
+        } else {
+            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
+            sam_print_usage(argc, argv, params);
+            exit(0);
+        }
+    }
+
+    return true;
+}
+
+int main(int argc, char ** argv) {
+    const int64_t t_main_start_us = ggml_time_us();
+
+    sam_params params;
+    params.model = "models/sam-vit-b/ggml-model-f16.bin";
+
+    sam_model model;
+    sam_state state;
+    int64_t t_load_us = 0;
+
+    if (sam_params_parse(argc, argv, params) == false) {
+        return 1;
+    }
+
+    if (params.seed < 0) {
+        params.seed = time(NULL);
+    }
+    fprintf(stderr, "%s: seed = %d\n", __func__, params.seed);
+
+    // load the image
+    sam_image_u8 img0;
+    if (!sam_image_load_from_file(params.fname_inp, img0)) {
+        fprintf(stderr, "%s: failed to load image from '%s'\n", __func__, params.fname_inp.c_str());
+        return 1;
+    }
+    fprintf(stderr, "%s: loaded image '%s' (%d x %d)\n", __func__, params.fname_inp.c_str(), img0.nx, img0.ny);
+
+    // preprocess to f32
+    sam_image_f32 img1;
+    if (!sam_image_preprocess(img0, img1)) {
+        fprintf(stderr, "%s: failed to preprocess image\n", __func__);
+        return 1;
+    }
+    fprintf(stderr, "%s: preprocessed image (%d x %d)\n", __func__, img1.nx, img1.ny);
+
+
+    // load the model
+    {
+        const int64_t t_start_us = ggml_time_us();
+
+        if (!sam_model_load(params.model, model)) {
+            fprintf(stderr, "%s: failed to load model from '%s'\n", __func__, params.model.c_str());
+            return 1;
+        }
+
+        t_load_us = ggml_time_us() - t_start_us;
+    }
+
+    {
+        static size_t buf_size = 256u*1024*1024;
+
+        struct ggml_init_params ggml_params = {
+            /*.mem_size   =*/ buf_size,
+            /*.mem_buffer =*/ NULL,
+            /*.no_alloc   =*/ false,
+        };
+
+        state.ctx = ggml_init(ggml_params);
+
+        state.embd_img = ggml_new_tensor_3d(state.ctx, GGML_TYPE_F32,
+                model.hparams.n_img_embd(), model.hparams.n_img_embd(), model.hparams.n_enc_out_chans);
+
+        state.low_res_masks = ggml_new_tensor_3d(state.ctx, GGML_TYPE_F32,
+                model.hparams.n_enc_out_chans, model.hparams.n_enc_out_chans, 3);
+
+        state.iou_predictions = ggml_new_tensor_1d(state.ctx, GGML_TYPE_F32, 3);
+    }
+
+
+    static const size_t tensor_alignment = 32;
+    {
+        state.buf_compute_img_enc.resize(ggml_tensor_overhead()*GGML_MAX_NODES + ggml_graph_overhead());
+        state.allocr = ggml_allocr_new_measure(tensor_alignment);
+        struct ggml_cgraph * gf_measure = sam_encode_image(model, state, img1);
+        if (!gf_measure) {
+            fprintf(stderr, "%s: failed to encode image\n", __func__);
+            return 1;
+        }
+
+        size_t alloc_size = ggml_allocr_alloc_graph(state.allocr, gf_measure) + tensor_alignment;
+        ggml_allocr_free(state.allocr);
+
+        // recreate allocator with exact memory requirements
+        state.buf_alloc_img_enc.resize(alloc_size);
+        state.allocr = ggml_allocr_new(state.buf_alloc_img_enc.data(), state.buf_alloc_img_enc.size(), tensor_alignment);
+
+        // compute the graph with the measured exact memory requirements from above
+        ggml_allocr_reset(state.allocr);
+
+        struct ggml_cgraph  * gf = sam_encode_image(model, state, img1);
+        if (!gf) {
+            fprintf(stderr, "%s: failed to encode image\n", __func__);
+            return 1;
+        }
+
+        ggml_allocr_alloc_graph(state.allocr, gf);
+
+        ggml_graph_compute_helper(state.work_buffer, gf, params.n_threads);
+
+        print_t_f32("embd_img", state.embd_img);
+
+        ggml_allocr_free(state.allocr);
+        state.allocr = NULL;
+        state.work_buffer.clear();
+    }
+    {
+        state.buf_compute_fast.resize(ggml_tensor_overhead()*GGML_MAX_NODES + ggml_graph_overhead());
+        state.allocr = ggml_allocr_new_measure(tensor_alignment);
+
+        // TODO: user input
+        const sam_point pt = { 414.375f, 162.796875f, };
+        // measure memory requirements for the graph
+        struct ggml_cgraph  * gf_measure = sam_build_fast_graph(model, state, img0.nx, img0.ny, pt);
+        if (!gf_measure) {
+            fprintf(stderr, "%s: failed to build fast graph to measure\n", __func__);
+            return 1;
+        }
+
+        size_t alloc_size = ggml_allocr_alloc_graph(state.allocr, gf_measure) + tensor_alignment;
+        ggml_allocr_free(state.allocr);
+
+        // recreate allocator with exact memory requirements
+        state.buf_alloc_fast.resize(alloc_size);
+        state.allocr = ggml_allocr_new(state.buf_alloc_fast.data(), state.buf_alloc_fast.size(), tensor_alignment);
+
+        // compute the graph with the measured exact memory requirements from above
+        ggml_allocr_reset(state.allocr);
+
+        struct ggml_cgraph  * gf = sam_build_fast_graph(model, state, img0.nx, img0.ny, pt);
+        if (!gf) {
+            fprintf(stderr, "%s: failed to build fast graph\n", __func__);
+            return 1;
+        }
+
+        ggml_allocr_alloc_graph(state.allocr, gf);
+
+        ggml_graph_compute_helper(state.work_buffer, gf, params.n_threads);
+
+        //print_t_f32("iou_predictions", state.iou_predictions);
+        //print_t_f32("low_res_masks", state.low_res_masks);
+        ggml_allocr_free(state.allocr);
+        state.allocr = NULL;
+    }
+
+    if (!sam_write_masks(model.hparams, img0.nx, img0.ny, state)) {
+        fprintf(stderr, "%s: failed to write masks\n", __func__);
+        return 1;
+    }
+
+    // report timing
+    {
+        const int64_t t_main_end_us = ggml_time_us();
+
+        fprintf(stderr, "\n\n");
+        fprintf(stderr, "%s:     load time = %8.2f ms\n", __func__, t_load_us/1000.0f);
+        fprintf(stderr, "%s:    total time = %8.2f ms\n", __func__, (t_main_end_us - t_main_start_us)/1000.0f);
+    }
+
+    ggml_free(model.ctx);
+
+    return 0;
+}

ggml/examples/starcoder/CMakeLists.txt

@@ -0,0 +1,31 @@
+#
+# starcoder
+
+set(TEST_TARGET starcoder)
+add_executable(${TEST_TARGET} main.cpp)
+target_link_libraries(${TEST_TARGET} PRIVATE ggml common common-ggml)
+
+#
+# starcoder-mmap
+
+set(TEST_TARGET starcoder-mmap)
+add_executable(${TEST_TARGET} starcoder-mmap.cpp)
+target_link_libraries(${TEST_TARGET} PRIVATE ggml common common-ggml)
+
+#
+# starcoder-quantize
+
+set(TEST_TARGET starcoder-quantize)
+add_executable(${TEST_TARGET} quantize.cpp)
+target_link_libraries(${TEST_TARGET} PRIVATE ggml common common-ggml)
+
+#
+# For GPU offloading
+
+if (GGML_CUBLAS)
+    add_compile_definitions(GGML_USE_CUBLAS)
+endif()
+if (GGML_CLBLAST)
+    add_compile_definitions(GGML_USE_CLBLAST)
+endif()
+

ggml/examples/starcoder/README.md

@@ -0,0 +1,115 @@
+# 💫 StarCoder
+
+This is a C++ example running 💫 StarCoder inference using the [ggml](https://github.com/ggerganov/ggml) library.
+
+The program runs on the CPU - no video card is required.
+
+The example supports the following 💫 StarCoder models:
+
+- `bigcode/starcoder`
+- `bigcode/gpt_bigcode-santacoder` aka the smol StarCoder
+
+Sample performance on MacBook M1 Pro:
+
+TODO
+
+
+Sample output:
+
+```
+$ ./bin/starcoder -h
+usage: ./bin/starcoder [options]
+
+options:
+  -h, --help            show this help message and exit
+  -s SEED, --seed SEED  RNG seed (default: -1)
+  -t N, --threads N     number of threads to use during computation (default: 8)
+  -p PROMPT, --prompt PROMPT
+                        prompt to start generation with (default: random)
+  -n N, --n_predict N   number of tokens to predict (default: 200)
+  --top_k N             top-k sampling (default: 40)
+  --top_p N             top-p sampling (default: 0.9)
+  --temp N              temperature (default: 1.0)
+  -b N, --batch_size N  batch size for prompt processing (default: 8)
+  -m FNAME, --model FNAME
+                        model path (default: models/starcoder-117M/ggml-model.bin)
+
+$ ./bin/starcoder -m ../models/bigcode/gpt_bigcode-santacoder-ggml-q4_1.bin -p "def fibonnaci(" -t 4 --top_k 0 --top_p 0.95 --temp 0.2      
+main: seed = 1683881276
+starcoder_model_load: loading model from '../models/bigcode/gpt_bigcode-santacoder-ggml-q4_1.bin'
+starcoder_model_load: n_vocab = 49280
+starcoder_model_load: n_ctx   = 2048
+starcoder_model_load: n_embd  = 2048
+starcoder_model_load: n_head  = 16
+starcoder_model_load: n_layer = 24
+starcoder_model_load: ftype   = 3
+starcoder_model_load: ggml ctx size = 1794.90 MB
+starcoder_model_load: memory size =   768.00 MB, n_mem = 49152
+starcoder_model_load: model size  =  1026.83 MB
+main: prompt: 'def fibonnaci('
+main: number of tokens in prompt = 7, first 8 tokens: 563 24240 78 2658 64 2819 7 
+
+def fibonnaci(n):
+    if n == 0:
+        return 0
+    elif n == 1:
+        return 1
+    else:
+        return fibonacci(n-1) + fibonacci(n-2)
+
+print(fibo(10))
+
+main: mem per token =  9597928 bytes
+main:     load time =   480.43 ms
+main:   sample time =    26.21 ms
+main:  predict time =  3987.95 ms / 19.36 ms per token
+main:    total time =  4580.56 ms
+```
+
+## Quick start
+```bash
+git clone https://github.com/ggerganov/ggml
+cd ggml
+
+# Install Python dependencies
+python3 -m pip install -r requirements.txt
+
+# Convert HF model to ggml
+python examples/starcoder/convert-hf-to-ggml.py bigcode/gpt_bigcode-santacoder
+
+# Build ggml + examples
+mkdir build && cd build
+cmake .. && make -j4 starcoder starcoder-quantize
+
+# quantize the model
+./bin/starcoder-quantize ../models/bigcode/gpt_bigcode-santacoder-ggml.bin ../models/bigcode/gpt_bigcode-santacoder-ggml-q4_1.bin 3
+
+# run inference
+./bin/starcoder -m ../models/bigcode/gpt_bigcode-santacoder-ggml-q4_1.bin -p "def fibonnaci(" --top_k 0 --top_p 0.95 --temp 0.2
+```
+
+
+## Downloading and converting the original models (💫 StarCoder)
+
+You can download the original model and convert it to `ggml` format using the script `convert-hf-to-ggml.py`:
+
+```
+# Convert HF model to ggml
+python examples/starcoder/convert-hf-to-ggml.py bigcode/gpt_bigcode-santacoder
+```
+
+This conversion requires that you have python and Transformers installed on your computer.
+
+## Quantizing the models
+
+You can also try to quantize the `ggml` models via 4-bit integer quantization.
+
+```
+# quantize the model
+./bin/starcoder-quantize ../models/bigcode/gpt_bigcode-santacoder-ggml.bin ../models/bigcode/gpt_bigcode-santacoder-ggml-q4_1.bin 3
+```
+
+| Model | Original size | Quantized size | Quantization type |
+| --- | --- | --- | --- |
+| `bigcode/gpt_bigcode-santacoder` | 5396.45 MB | 1026.83 MB | 4-bit integer (q4_1) |
+| `bigcode/starcoder` | 71628.23 MB | 13596.23 MB | 4-bit integer (q4_1) |

ggml/examples/starcoder/convert-hf-to-ggml.py

@@ -0,0 +1,208 @@
+# Convert HF models to ggml format
+#
+
+import sys
+import struct
+import json
+import torch
+import numpy as np
+import re
+import os
+import argparse
+
+from transformers import AutoModelForCausalLM
+from transformers import AutoTokenizer, AutoModelForCausalLM, AutoConfig, BloomForCausalLM
+
+# ref: https://github.com/openai/gpt-2/blob/master/src/encoder.py
+def bytes_to_unicode():
+    """
+    Returns list of utf-8 byte and a corresponding list of unicode strings.
+    The reversible bpe codes work on unicode strings.
+    This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.
+    When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.
+    This is a signficant percentage of your normal, say, 32K bpe vocab.
+    To avoid that, we want lookup tables between utf-8 bytes and unicode strings.
+    And avoids mapping to whitespace/control characters the bpe code barfs on.
+    """
+    bs = list(range(ord("!"), ord("~")+1))+list(range(ord("¡"), ord("¬")+1))+list(range(ord("®"), ord("ÿ")+1))
+    cs = bs[:]
+    n = 0
+    for b in range(2**8):
+        if b not in bs:
+            bs.append(b)
+            cs.append(2**8+n)
+            n += 1
+    cs = [chr(n) for n in cs]
+    return dict(zip(bs, cs))
+
+parser = argparse.ArgumentParser(description='Convert starcoder HF model to GGML')
+parser.add_argument('model_name_or_path', type=str, help='Name of model on HF hub, or local model folder')
+parser.add_argument('--outfile', type=str, default='ggml-model.bin', help='Path of GGML file to write.')
+parser.add_argument('--use_f32', action="store_true", help='Save GGML file in fp32')
+
+args = parser.parse_args()
+
+# use 16-bit or 32-bit floats
+use_f16 = not args.use_f32
+
+fname_out = args.outfile
+fname_dir = os.path.dirname(fname_out)
+if fname_dir:
+    os.makedirs(fname_dir, exist_ok=True)
+
+print("Loading model: ", args.model_name_or_path)
+tokenizer = AutoTokenizer.from_pretrained(args.model_name_or_path)
+config = AutoConfig.from_pretrained(args.model_name_or_path, trust_remote_code=True)
+hparams = config.to_dict()
+model = AutoModelForCausalLM.from_pretrained(args.model_name_or_path, config=config, torch_dtype=torch.float16 if use_f16 else torch.float32, low_cpu_mem_usage=True, trust_remote_code=True, offload_state_dict=True)
+print("Model loaded: ", args.model_name_or_path)
+
+list_vars = model.state_dict()
+
+encoder = tokenizer.vocab
+# Add added_tokens (special tokens) to the encoder
+encoder.update(tokenizer.get_added_vocab())
+print(hparams)
+
+print("Saving ggml model to: ", fname_out)
+fout = open(fname_out, "wb")
+
+fout.write(struct.pack("i", 0x67676d6c)) # magic: ggml in hex
+vocab_size = hparams["vocab_size"]
+fout.write(struct.pack("i", vocab_size))
+# fout.write(struct.pack("i", len(encoder)))
+fout.write(struct.pack("i", hparams["n_positions"]))
+fout.write(struct.pack("i", hparams["n_embd"]))
+fout.write(struct.pack("i", hparams["n_head"]))
+fout.write(struct.pack("i", hparams["n_layer"]))
+fout.write(struct.pack("i", use_f16))
+
+byte_encoder = bytes_to_unicode()
+byte_decoder = {v:k for k, v in byte_encoder.items()}
+
+fout.write(struct.pack("i", vocab_size))
+
+counter = 0
+# sort by value
+for key in sorted(encoder, key=encoder.get):
+    text = bytearray([byte_decoder[c] for c in key])
+    fout.write(struct.pack("i", len(text)))
+    fout.write(text)
+    counter += 1
+
+# TODO: Repeat last token until vocab_size
+while counter < vocab_size:
+    fout.write(struct.pack("i", len(text)))
+    fout.write(text)
+    counter += 1
+# assert counter == config.vocab_size
+
+for name in list_vars.keys():
+    data = list_vars[name].squeeze().numpy()
+    print("Processing variable: " + name + " with shape: ", data.shape)
+
+    # rename headers to keep compatibility
+    if name == "transformer.ln_f.weight":
+        name = "model/ln_f/g"
+    elif name == "transformer.ln_f.bias":
+        name = "model/ln_f/b"
+    elif name == "transformer.wte.weight":
+        name = "model/wte"
+    elif name == "transformer.wpe.weight":
+        name = "model/wpe"
+    elif name == "lm_head.weight":
+        name = "model/lm_head"
+    elif re.match(r"transformer.h\.\d+\.ln_1\.weight", name):
+        i = re.findall("\d+", name)[0]
+        name = f"model/h{i}/ln_1/g"
+    elif re.match(r"transformer.h\.\d+\.ln_1\.bias", name):
+        i = re.findall("\d+", name)[0]
+        name = f"model/h{i}/ln_1/b"
+    elif re.match(r"transformer.h\.\d+\.attn\.c_attn\.weight", name):
+        i = re.findall("\d+", name)[0]
+        name = f"model/h{i}/attn/c_attn/w"
+    elif re.match(r"transformer.h\.\d+\.attn\.c_attn\.bias", name):
+        i = re.findall("\d+", name)[0]
+        name = f"model/h{i}/attn/c_attn/b"
+    elif re.match(r"transformer.h\.\d+\.attn\.c_proj\.weight", name):
+        i = re.findall("\d+", name)[0]
+        name = f"model/h{i}/attn/c_proj/w"
+    elif re.match(r"transformer.h.\d+.attn.c_proj.bias", name):
+        i = re.findall("\d+", name)[0]
+        name = f"model/h{i}/attn/c_proj/b"
+    elif re.match(r"transformer.h.\d+.ln_2.weight", name):
+        i = re.findall("\d+", name)[0]
+        name = f"model/h{i}/ln_2/g"
+    elif re.match(r"transformer.h.\d+.ln_2.bias", name):
+        i = re.findall("\d+", name)[0]
+        name = f"model/h{i}/ln_2/b"
+    elif re.match(r"transformer.h.\d+.mlp.c_fc.weight", name):
+        i = re.findall("\d+", name)[0]
+        name = f"model/h{i}/mlp/c_fc/w"
+    elif re.match(r"transformer.h.\d+.mlp.c_fc.bias", name):
+        i = re.findall("\d+", name)[0]
+        name = f"model/h{i}/mlp/c_fc/b"
+    elif re.match(r"transformer.h.\d+.mlp.c_proj.weight", name):
+        i = re.findall("\d+", name)[0]
+        name = f"model/h{i}/mlp/c_proj/w"
+    elif re.match(r"transformer.h.\d+.mlp.c_proj.bias", name):
+        i = re.findall("\d+", name)[0]
+        name = f"model/h{i}/mlp/c_proj/b"
+    else:
+        print("Unrecognized variable name. %s", name)
+
+    # we don't need these
+    if name.endswith("attn.masked_bias") or name.endswith(".attn.bias"):
+        print("  Skipping variable: " + name)
+        continue
+
+    n_dims = len(data.shape);
+
+    # ftype == 0 -> float32, ftype == 1 -> float16
+    ftype = 0;
+    if use_f16:
+        if (name == "model/wte" or name == "model/lm_head" or name[-2:] == "/g" or name[-2:] == "/w") and n_dims == 2:
+            print("  Converting to float16")
+            data = data.astype(np.float16)
+            ftype = 1
+        else:
+            print("  Converting to float32")
+            data = data.astype(np.float32)
+            ftype = 0
+
+    "model/h.*/attn/c_attn/w"
+    "model/h.*/attn/c_proj/w"
+    "model/h.*/mlp/c_fc/w"
+    "model/h.*/mlp/c_proj/w"
+    if name[-14:] == "/attn/c_attn/w" or name[-14:] == "/attn/c_attn/b":
+        print("  Duplicate K,V heads to use MHA instead of MQA")
+
+        embed_dim = hparams["n_embd"]
+        head_dim = embed_dim // hparams["n_head"]
+
+        # ((n_heads + 2) * head_dim, hidden_dim) -> (3 * n_heads * head_dim, hidden_dim)
+        q, k ,v = np.split(data, (hparams["n_head"] * head_dim, (hparams["n_head"] + 1) * head_dim), axis=0)
+        # duplicate k, v along the first axis (head_dim, hidden_dim) -> (n_heads * head_dim, hidden_dim)
+        if len(k.shape) == 2:
+            k = np.tile(k, (hparams["n_head"], 1))
+            v = np.tile(v, (hparams["n_head"], 1))
+        elif len(k.shape) == 1:
+            k = np.tile(k, (hparams["n_head"]))
+            v = np.tile(v, (hparams["n_head"]))
+        # concat q, k, v along the first axis (n_heads * head_dim, hidden_dim) -> (3 * n_heads * head_dim, hidden_dim)
+        data = np.concatenate((q, k, v), axis=0)
+
+    # header
+    str = name.encode('utf-8')
+    fout.write(struct.pack("iii", n_dims, len(str), ftype))
+    for i in range(n_dims):
+        fout.write(struct.pack("i", data.shape[n_dims - 1 - i]))
+    fout.write(str);
+
+    # data
+    data.tofile(fout)
+
+fout.close()
+
+print("Done. Output file: " + fname_out)
+print("")

ggml/examples/starcoder/main.cpp

@@ -0,0 +1,927 @@
+#include "ggml/ggml.h"
+
+#include "common.h"
+#include "common-ggml.h"
+
+#include <cassert>
+#include <cmath>
+#include <cstdio>
+#include <cstring>
+#include <fstream>
+#include <map>
+#include <string>
+#include <vector>
+
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
+// default hparams (GPT-2 117M)
+// https://huggingface.co/bigcode/gpt_bigcode-santacoder/blob/main/config.json
+struct starcoder_hparams {
+    int32_t n_vocab = 49280;
+    int32_t n_ctx   = 2048;
+    int32_t n_embd  = 2048;
+    int32_t n_head  = 16;
+    int32_t n_layer = 24;
+    int32_t ftype   = 1;
+    float   eps     = 1e-5f;
+};
+
+struct starcoder_layer {
+    // normalization
+    struct ggml_tensor * ln_1_g;
+    struct ggml_tensor * ln_1_b;
+
+    struct ggml_tensor * ln_2_g;
+    struct ggml_tensor * ln_2_b;
+
+    // attention
+    struct ggml_tensor * c_attn_attn_w;
+    struct ggml_tensor * c_attn_attn_b;
+
+    struct ggml_tensor * c_attn_proj_w;
+    struct ggml_tensor * c_attn_proj_b;
+
+    // mlp
+    struct ggml_tensor * c_mlp_fc_w;
+    struct ggml_tensor * c_mlp_fc_b;
+
+    struct ggml_tensor * c_mlp_proj_w;
+    struct ggml_tensor * c_mlp_proj_b;
+};
+
+struct starcoder_model {
+    starcoder_hparams hparams;
+
+    // normalization
+    struct ggml_tensor * ln_f_g;
+    struct ggml_tensor * ln_f_b;
+
+    struct ggml_tensor * wte;     // position embedding
+    struct ggml_tensor * wpe;     //    token embedding
+    struct ggml_tensor * lm_head; // language model head
+
+    std::vector<starcoder_layer> layers;
+
+    // key + value memory
+    struct ggml_tensor * memory_k;
+    struct ggml_tensor * memory_v;
+
+    //
+    struct ggml_context * ctx;
+    std::map<std::string, struct ggml_tensor *> tensors;
+};
+
+// load the model's weights from a file
+bool starcoder_model_load(const std::string & fname, starcoder_model & model, gpt_vocab & vocab) {
+    printf("%s: loading model from '%s'\n", __func__, fname.c_str());
+
+    auto fin = std::ifstream(fname, std::ios::binary);
+    if (!fin) {
+        fprintf(stderr, "%s: failed to open '%s'\n", __func__, fname.c_str());
+        return false;
+    }
+
+    // verify magic
+    {
+        uint32_t magic;
+        fin.read((char *) &magic, sizeof(magic));
+        if (magic != GGML_FILE_MAGIC) {
+            fprintf(stderr, "%s: invalid model file '%s' (bad magic)\n", __func__, fname.c_str());
+            return false;
+        }
+    }
+
+    // load hparams
+    {
+        auto & hparams = model.hparams;
+
+        fin.read((char *) &hparams.n_vocab, sizeof(hparams.n_vocab));
+        fin.read((char *) &hparams.n_ctx,   sizeof(hparams.n_ctx));
+        fin.read((char *) &hparams.n_embd,  sizeof(hparams.n_embd));
+        fin.read((char *) &hparams.n_head,  sizeof(hparams.n_head));
+        fin.read((char *) &hparams.n_layer, sizeof(hparams.n_layer));
+        fin.read((char *) &hparams.ftype,   sizeof(hparams.ftype));
+
+        const int32_t qntvr = hparams.ftype / GGML_QNT_VERSION_FACTOR;
+
+        printf("%s: n_vocab = %d\n", __func__, hparams.n_vocab);
+        printf("%s: n_ctx   = %d\n", __func__, hparams.n_ctx);
+        printf("%s: n_embd  = %d\n", __func__, hparams.n_embd);
+        printf("%s: n_head  = %d\n", __func__, hparams.n_head);
+        printf("%s: n_layer = %d\n", __func__, hparams.n_layer);
+        printf("%s: ftype   = %d\n", __func__, hparams.ftype);
+        printf("%s: qntvr   = %d\n", __func__, qntvr);
+
+        hparams.ftype %= GGML_QNT_VERSION_FACTOR;
+    }
+
+    // load vocab
+    {
+        int32_t n_vocab = 0;
+        fin.read((char *) &n_vocab, sizeof(n_vocab));
+
+        if (n_vocab != model.hparams.n_vocab) {
+            fprintf(stderr, "%s: invalid model file '%s' (bad vocab size %d != %d)\n",
+                    __func__, fname.c_str(), n_vocab, model.hparams.n_vocab);
+            return false;
+        }
+
+        std::string word;
+        std::vector<char> buf(128);
+
+        for (int i = 0; i < n_vocab; i++) {
+            uint32_t len;
+            fin.read((char *) &len, sizeof(len));
+
+            buf.resize(len);
+            fin.read((char *) buf.data(), len);
+            word.assign(buf.data(), len);
+
+            vocab.token_to_id[word] = i;
+            vocab.id_to_token[i] = word;
+
+            // if (i < 10) fprintf(stderr, "%.s: vocab[%d] = '%s'\n", __func__, i, word.c_str());
+        }
+
+        // Add StarChat special tokens.
+        for (std::string token : {
+                "<|system|>",
+                "<|user|>",
+                "<|assistant|>",
+                "<|end|>",
+                "<fim-prefix>",
+                "<fim-middle>",
+                "<fim-suffix>",
+                "<fim-pad>",
+                "<|end_of_turn|>"
+            }) {
+            if (vocab.token_to_id.find(token) != vocab.token_to_id.end()) {
+                vocab.add_special_token(token);
+            }
+        }
+    }
+
+    // for the big tensors, we have the option to store the data in 16-bit floats or quantized
+    // in order to save memory and also to speed up the computation
+    ggml_type wtype = ggml_ftype_to_ggml_type((ggml_ftype) (model.hparams.ftype));
+    if (wtype == GGML_TYPE_COUNT) {
+        fprintf(stderr, "%s: invalid model file '%s' (bad ftype value %d)\n",
+                __func__, fname.c_str(), model.hparams.ftype);
+        return false;
+    }
+
+    auto & ctx = model.ctx;
+
+    size_t ctx_size = 0;
+
+    {
+        const auto & hparams = model.hparams;
+
+        const int n_embd  = hparams.n_embd;
+        const int n_layer = hparams.n_layer;
+        const int n_ctx   = hparams.n_ctx;
+        const int n_vocab = hparams.n_vocab;
+
+        const int head_dim = n_embd / hparams.n_head;
+        const int kv_heads = hparams.n_head; // 1 if MQA else hparams.n_head
+        const int kv_dim   = kv_heads * head_dim;
+
+        ctx_size += n_embd*ggml_type_sizef(GGML_TYPE_F32); // ln_f_g
+        ctx_size += n_embd*ggml_type_sizef(GGML_TYPE_F32); // ln_f_b
+
+        ctx_size += n_vocab*n_embd*ggml_type_sizef(wtype);         // wte
+        ctx_size +=   n_ctx*n_embd*ggml_type_sizef(GGML_TYPE_F32); // wpe
+        ctx_size += n_vocab*n_embd*ggml_type_sizef(wtype);         // lm_head
+
+        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_1_g
+        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_1_b
+
+        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_2_g
+        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_2_b
+
+        ctx_size += n_layer*((n_embd + 2*kv_dim)*n_embd*ggml_type_sizef(wtype));         // c_attn_attn_w // TODO:
+        ctx_size += n_layer*(       (n_embd + 2*kv_dim)*ggml_type_sizef(GGML_TYPE_F32)); // c_attn_attn_b
+
+        ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype));           // c_attn_proj_w
+        ctx_size += n_layer*(       n_embd*ggml_type_sizef(GGML_TYPE_F32));   // c_attn_proj_b
+
+        ctx_size += n_layer*(4*n_embd*n_embd*ggml_type_sizef(wtype));         // c_mlp_fc_w
+        ctx_size += n_layer*(       4*n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_mlp_fc_b
+
+        ctx_size += n_layer*(4*n_embd*n_embd*ggml_type_sizef(wtype));         // c_mlp_proj_w
+        ctx_size += n_layer*(         n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_mlp_proj_b
+
+        ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(GGML_TYPE_F32); // memory_k
+        ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(GGML_TYPE_F32); // memory_v
+
+        ctx_size += (6 + 12*n_layer)*512; // object overhead
+
+        printf("%s: ggml ctx size = %6.2f MB\n", __func__, ctx_size/(1024.0*1024.0));
+    }
+
+    // create the ggml context
+    {
+        struct ggml_init_params params = {
+            /*.mem_size   =*/ ctx_size,
+            /*.mem_buffer =*/ NULL,
+            /*.no_alloc   =*/ false,
+        };
+
+        model.ctx = ggml_init(params);
+        if (!model.ctx) {
+            fprintf(stderr, "%s: ggml_init() failed\n", __func__);
+            return false;
+        }
+    }
+
+    // prepare memory for the weights
+    {
+        const auto & hparams = model.hparams;
+
+        const int n_embd  = hparams.n_embd;
+        const int n_layer = hparams.n_layer;
+        const int n_ctx   = hparams.n_ctx;
+        const int n_vocab = hparams.n_vocab;
+
+        const int head_dim = n_embd / hparams.n_head;
+        const int kv_heads = hparams.n_head; // 1 if MQA else hparams.n_head
+        const int kv_dim   = kv_heads * head_dim;
+
+        model.layers.resize(n_layer);
+
+        model.ln_f_g = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
+        model.ln_f_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
+
+        model.wte     = ggml_new_tensor_2d(ctx, wtype,         n_embd, n_vocab);
+        model.wpe     = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_ctx);
+        model.lm_head = ggml_new_tensor_2d(ctx, wtype,         n_embd, n_vocab);
+
+        // map by name
+        model.tensors["model/ln_f/g"] = model.ln_f_g;
+        model.tensors["model/ln_f/b"] = model.ln_f_b;
+
+        model.tensors["model/wte"]     = model.wte;
+        model.tensors["model/wpe"]     = model.wpe;
+        model.tensors["model/lm_head"] = model.lm_head;
+
+        for (int i = 0; i < n_layer; ++i) {
+            auto & layer = model.layers[i];
+
+            layer.ln_1_g        = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+            layer.ln_1_b        = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+
+            layer.ln_2_g        = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+            layer.ln_2_b        = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+
+            layer.c_attn_attn_w = ggml_new_tensor_2d(ctx, wtype,           n_embd, n_embd + 2*kv_dim);
+            layer.c_attn_attn_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd + 2*kv_dim);
+
+            layer.c_attn_proj_w = ggml_new_tensor_2d(ctx, wtype,           n_embd, n_embd);
+            layer.c_attn_proj_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+
+            layer.c_mlp_fc_w    = ggml_new_tensor_2d(ctx, wtype,           n_embd, 4*n_embd); //TODO: 4*n_embd = config.n_inner
+            layer.c_mlp_fc_b    = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 4*n_embd);
+
+            layer.c_mlp_proj_w  = ggml_new_tensor_2d(ctx, wtype,         4*n_embd, n_embd);
+            layer.c_mlp_proj_b  = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+
+            // map by name
+            model.tensors["model/h" + std::to_string(i) + "/ln_1/g"]        = layer.ln_1_g;
+            model.tensors["model/h" + std::to_string(i) + "/ln_1/b"]        = layer.ln_1_b;
+
+            model.tensors["model/h" + std::to_string(i) + "/ln_2/g"]        = layer.ln_2_g;
+            model.tensors["model/h" + std::to_string(i) + "/ln_2/b"]        = layer.ln_2_b;
+
+            model.tensors["model/h" + std::to_string(i) + "/attn/c_attn/w"] = layer.c_attn_attn_w;
+            model.tensors["model/h" + std::to_string(i) + "/attn/c_attn/b"] = layer.c_attn_attn_b;
+
+            model.tensors["model/h" + std::to_string(i) + "/attn/c_proj/w"] = layer.c_attn_proj_w;
+            model.tensors["model/h" + std::to_string(i) + "/attn/c_proj/b"] = layer.c_attn_proj_b;
+
+            model.tensors["model/h" + std::to_string(i) + "/mlp/c_fc/w"]    = layer.c_mlp_fc_w;
+            model.tensors["model/h" + std::to_string(i) + "/mlp/c_fc/b"]    = layer.c_mlp_fc_b;
+
+            model.tensors["model/h" + std::to_string(i) + "/mlp/c_proj/w"]  = layer.c_mlp_proj_w;
+            model.tensors["model/h" + std::to_string(i) + "/mlp/c_proj/b"]  = layer.c_mlp_proj_b;
+        }
+    }
+
+    // key + value memory
+    {
+        const auto & hparams = model.hparams;
+
+        const int n_embd  = hparams.n_embd;
+        const int n_layer = hparams.n_layer;
+        const int n_ctx   = hparams.n_ctx;
+
+        const int n_mem      = n_layer*n_ctx;
+        const int n_elements = n_embd*n_mem;
+
+        model.memory_k = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_elements);
+        model.memory_v = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_elements);
+
+        const size_t memory_size = ggml_nbytes(model.memory_k) + ggml_nbytes(model.memory_v);
+
+        printf("%s: memory size = %8.2f MB, n_mem = %d\n", __func__, memory_size/1024.0/1024.0, n_mem);
+    }
+
+    // load weights
+    {
+        size_t total_size = 0;
+
+        bool has_lm_head = false;
+
+        while (true) {
+            int32_t n_dims;
+            int32_t length;
+            int32_t ttype;
+
+            fin.read(reinterpret_cast<char *>(&n_dims), sizeof(n_dims));
+            fin.read(reinterpret_cast<char *>(&length), sizeof(length));
+            fin.read(reinterpret_cast<char *>(&ttype),  sizeof(ttype));
+
+            if (fin.eof()) {
+                break;
+            }
+
+            int32_t nelements = 1;
+            int32_t ne[2] = { 1, 1 };
+            for (int i = 0; i < n_dims; ++i) {
+                fin.read(reinterpret_cast<char *>(&ne[i]), sizeof(ne[i]));
+                nelements *= ne[i];
+            }
+
+            std::string name(length, 0);
+            fin.read(&name[0], length);
+
+            if (model.tensors.find(name) == model.tensors.end()) {
+                fprintf(stderr, "%s: unknown tensor '%s' in model file\n", __func__, name.c_str());
+                return false;
+            }
+
+            auto tensor = model.tensors[name];
+            if (tensor->ne[0] != ne[0] || tensor->ne[1] != ne[1]) {
+                fprintf(stderr, "%s: tensor '%s' has wrong shape in model file: got [%d, %d], expected [%d, %d]\n",
+                        __func__, name.c_str(), (int) tensor->ne[0], (int) tensor->ne[1], ne[0], ne[1]);
+                return false;
+            }
+            if (ggml_nelements(tensor) != nelements) {
+                fprintf(stderr, "%s: tensor '%s' has wrong size in model file. got %d, expected %d\n",
+                        __func__, name.c_str(), (int) ggml_nelements(tensor), nelements);
+                return false;
+            }
+
+            // for debugging
+            if (0) {
+                printf("%24s - [%5d, %5d], type = %6s, %6.2f MB, %9zu bytes\n", name.c_str(), ne[0], ne[1], ggml_type_name(ggml_type(ttype)), ggml_nbytes(tensor)/1024.0/1024.0, ggml_nbytes(tensor));
+            }
+
+            const size_t bpe = ggml_type_size(ggml_type(ttype));
+
+            if ((nelements*bpe)/ggml_blck_size(tensor->type) != ggml_nbytes(tensor)) {
+                fprintf(stderr, "%s: tensor '%s' has wrong size in model file: got %zu, expected %zu\n",
+                        __func__, name.c_str(), ggml_nbytes(tensor), nelements*bpe);
+                return false;
+            }
+
+            fin.read(reinterpret_cast<char *>(tensor->data), ggml_nbytes(tensor));
+
+            // GPT-2 models share the WTE tensor as the LM head
+            if (name == "model/wte" && has_lm_head == false) {
+                memcpy(model.lm_head->data, tensor->data, ggml_nbytes(tensor));
+            }
+
+            if (name == "model/lm_head") {
+                has_lm_head = true;
+            }
+
+            total_size += ggml_nbytes(tensor);
+        }
+
+        printf("%s: model size  = %8.2f MB\n", __func__, total_size/1024.0/1024.0);
+    }
+
+    fin.close();
+
+    return true;
+}
+
+// evaluate the transformer
+//
+//   - model:     the model
+//   - n_threads: number of threads to use
+//   - n_past:    the context size so far
+//   - embd_inp:  the embeddings of the tokens in the context
+//   - embd_w:    the predicted logits for the next token
+//
+bool starcoder_eval(
+        const starcoder_model & model,
+        const int n_threads,
+        const int n_past,
+        const std::vector<gpt_vocab::id> & embd_inp,
+              std::vector<float>         & embd_w,
+              size_t                     & mem_per_token) {
+    const int N = embd_inp.size();
+
+    const auto & hparams = model.hparams;
+
+    const int n_embd  = hparams.n_embd;
+    const int n_layer = hparams.n_layer;
+    const int n_ctx   = hparams.n_ctx;
+    const int n_head  = hparams.n_head;
+    const int n_vocab = hparams.n_vocab;
+
+    static size_t buf_size = 256u*1024*1024;
+    static void * buf = malloc(buf_size);
+
+    // use 2 scratch buffers
+    // TODO: very hacky solution - reimplement in a more elegant way
+    static size_t scr0_size = 256u*1024*1024;
+    static void * scr0 = malloc(scr0_size);
+
+    static size_t scr1_size = 256u*1024*1024;
+    static void * scr1 = malloc(scr1_size);
+
+    if (mem_per_token > 0 && mem_per_token*N > buf_size) {
+        const size_t buf_size_new = 1.1*(mem_per_token*N); // add 10% to account for ggml object overhead
+        //printf("\n%s: reallocating buffer from %zu to %zu bytes\n", __func__, buf_size, buf_size_new);
+
+        // reallocate
+        buf_size = buf_size_new;
+        buf = realloc(buf, buf_size);
+        if (buf == nullptr) {
+            fprintf(stderr, "%s: failed to allocate %zu bytes\n", __func__, buf_size);
+            return false;
+        }
+    }
+
+    struct ggml_init_params params = {
+        /*.mem_size   =*/ buf_size,
+        /*.mem_buffer =*/ buf,
+        /*.no_alloc   =*/ false,
+    };
+
+    struct ggml_context * ctx0 = ggml_init(params);
+    struct ggml_cgraph gf = {};
+
+    struct ggml_tensor * embd = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
+    memcpy(embd->data, embd_inp.data(), N*ggml_element_size(embd));
+
+    struct ggml_tensor * position = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
+    for (int i = 0; i < N; ++i) {
+        ((int32_t *) position->data)[i] = n_past + i;
+    }
+
+    // wte + wpe
+    struct ggml_tensor * inpL =
+        ggml_add(ctx0,
+                ggml_get_rows(ctx0, model.wte, embd),
+                ggml_get_rows(ctx0, model.wpe, position));
+
+    for (int il = 0; il < n_layer; ++il) {
+        struct ggml_tensor * cur;
+
+        ggml_set_scratch(ctx0, { 0, scr0_size, scr0, });
+
+        // norm
+        {
+            // [ 768, N]
+            cur = ggml_norm(ctx0, inpL, hparams.eps);
+
+            // cur = ln_1_g*cur + ln_1_b
+            // [ 768, N]
+            cur = ggml_add(ctx0,
+                    ggml_mul(ctx0,
+                        ggml_repeat(ctx0, model.layers[il].ln_1_g, cur),
+                        cur),
+                    ggml_repeat(ctx0, model.layers[il].ln_1_b, cur));
+        }
+
+        // attn
+        // [2304, 768] - model.layers[il].c_attn_attn_w
+        // [2304,   1] - model.layers[il].c_attn_attn_b
+        // [ 768,   N] - cur (in)
+        // [2304,   N] - cur (out)
+        //
+        // cur = attn_w*cur + attn_b
+        // [2304, N]
+        {
+            cur = ggml_mul_mat(ctx0,
+                    model.layers[il].c_attn_attn_w,
+                    cur);
+
+            cur = ggml_add(ctx0,
+                    ggml_repeat(ctx0, model.layers[il].c_attn_attn_b, cur),
+                    cur);
+        }
+
+        // self-attention
+        {
+            struct ggml_tensor * Qcur = ggml_view_2d(ctx0, cur, n_embd, N, cur->nb[1], 0*sizeof(float)*n_embd);
+            struct ggml_tensor * Kcur = ggml_view_2d(ctx0, cur, n_embd, N, cur->nb[1], 1*sizeof(float)*n_embd);
+            struct ggml_tensor * Vcur = ggml_view_2d(ctx0, cur, n_embd, N, cur->nb[1], 2*sizeof(float)*n_embd);
+
+            // store key and value to memory
+            if (N >= 1) {
+                struct ggml_tensor * k = ggml_view_1d(ctx0, model.memory_k, N*n_embd, (ggml_element_size(model.memory_k)*n_embd)*(il*n_ctx + n_past));
+                struct ggml_tensor * v = ggml_view_1d(ctx0, model.memory_v, N*n_embd, (ggml_element_size(model.memory_v)*n_embd)*(il*n_ctx + n_past));
+
+                ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Kcur, k));
+                ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Vcur, v));
+            }
+
+            // Q = Qcur.contiguous().view(n_embd/n_head, n_head, N).permute(0, 2, 1, 3)
+            // [64, N, 12]
+            struct ggml_tensor * Q =
+                ggml_permute(ctx0,
+                        ggml_cpy(ctx0,
+                            Qcur,
+                            ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_embd/n_head, n_head, N)),
+                        0, 2, 1, 3);
+
+            // K = Kmem.view(n_embd/n_head, n_head, n_past + N).permute(0, 2, 1, 3)
+            // [64, n_past + N, 12]
+            struct ggml_tensor * K =
+                ggml_permute(ctx0,
+                        ggml_reshape_3d(ctx0,
+                            ggml_view_1d(ctx0, model.memory_k, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(model.memory_k)*n_embd),
+                            n_embd/n_head, n_head, n_past + N),
+                        0, 2, 1, 3); //TODO: need to be tiled
+
+            // GG: flash attention
+            //struct ggml_tensor * V =
+            //    ggml_cpy(ctx0,
+            //            ggml_permute(ctx0,
+            //                ggml_reshape_3d(ctx0,
+            //                    ggml_view_1d(ctx0, model.memory_v, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(model.memory_v)*n_embd),
+            //                    n_embd/n_head, n_head, n_past + N),
+            //                1, 2, 0, 3),
+            //            ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_past + N, n_embd/n_head, n_head));
+
+            //struct ggml_tensor * KQV = ggml_flash_attn(ctx0, Q, K, V, true);
+
+            // K * Q
+            // [n_past + N, N, 12]
+            struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q); //TODO: check if it broadcasts
+
+            // KQ_scaled = KQ / sqrt(n_embd/n_head)
+            // [n_past + N, N, 12]
+            struct ggml_tensor * KQ_scaled =
+                ggml_scale_inplace(ctx0,
+                        KQ,
+                        ggml_new_f32(ctx0, 1.0f/sqrt(float(n_embd)/n_head))
+                        );
+
+            // KQ_masked = mask_past(KQ_scaled)
+            // [n_past + N, N, 12]
+            struct ggml_tensor * KQ_masked = ggml_diag_mask_inf_inplace(ctx0, KQ_scaled, n_past);
+
+            // KQ = soft_max(KQ_masked)
+            // [n_past + N, N, 12]
+            struct ggml_tensor * KQ_soft_max = ggml_soft_max_inplace(ctx0, KQ_masked);
+
+            // V_trans = Vmem.view(n_embd/n_head, n_head, n_past + N).permute(1, 2, 0, 3).contiguous()
+            // [n_past + N, 64, 12]
+            struct ggml_tensor * V_trans =
+                ggml_cpy(ctx0,
+                        ggml_permute(ctx0,
+                            ggml_reshape_3d(ctx0,
+                                ggml_view_1d(ctx0, model.memory_v, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(model.memory_v)*n_embd),
+                                n_embd/n_head, n_head, n_past + N),
+                            1, 2, 0, 3),
+                        ggml_new_tensor_3d(ctx0, model.memory_v->type, n_past + N, n_embd/n_head, n_head));
+
+            // KQV = transpose(V) * KQ_soft_max
+            // [64, N, 12]
+            struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V_trans, KQ_soft_max);
+
+            // KQV_merged = KQV.permute(0, 2, 1, 3)
+            // [64, 12, N]
+            struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
+
+            // cur = KQV_merged.contiguous().view(n_embd, N)
+            // [768, N]
+            cur = ggml_cpy(ctx0,
+                    KQV_merged,
+                    ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, N));
+        }
+
+        // projection
+        // [ 768, 768] - model.layers[il].c_attn_proj_w
+        // [ 768,   1] - model.layers[il].c_attn_proj_b
+        // [ 768,   N] - cur (in)
+        // [ 768,   N] - cur (out)
+        //
+        // cur = proj_w*cur + proj_b
+        // [768, N]
+        {
+            cur = ggml_mul_mat(ctx0,
+                    model.layers[il].c_attn_proj_w,
+                    cur);
+
+            cur = ggml_add(ctx0,
+                    ggml_repeat(ctx0, model.layers[il].c_attn_proj_b, cur),
+                    cur);
+        }
+
+        // add the input
+        cur = ggml_add(ctx0, cur, inpL);
+
+        struct ggml_tensor * inpFF = cur;
+
+        ggml_set_scratch(ctx0, { 0, scr1_size, scr1, });
+
+        // feed-forward network
+        {
+            // norm
+            {
+                cur = ggml_norm(ctx0, inpFF, hparams.eps);
+
+                // cur = ln_2_g*cur + ln_2_b
+                // [ 768, N]
+                cur = ggml_add(ctx0,
+                        ggml_mul(ctx0,
+                            ggml_repeat(ctx0, model.layers[il].ln_2_g, cur),
+                            cur),
+                        ggml_repeat(ctx0, model.layers[il].ln_2_b, cur));
+            }
+
+            // fully connected
+            // [3072, 768] - model.layers[il].c_mlp_fc_w
+            // [3072,   1] - model.layers[il].c_mlp_fc_b
+            // [ 768,   N] - cur (in)
+            // [3072,   N] - cur (out)
+            //
+            // cur = fc_w*cur + fc_b
+            // [3072, N]
+            cur = ggml_mul_mat(ctx0,
+                    model.layers[il].c_mlp_fc_w,
+                    cur);
+
+            cur = ggml_add(ctx0,
+                    ggml_repeat(ctx0, model.layers[il].c_mlp_fc_b, cur),
+                    cur);
+
+            // GELU activation
+            // [3072, N]
+            cur = ggml_gelu(ctx0, cur);
+
+            // projection
+            // [ 768, 3072] - model.layers[il].c_mlp_proj_w
+            // [ 768,    1] - model.layers[il].c_mlp_proj_b
+            // [3072,    N] - cur (in)
+            // [ 768,    N] - cur (out)
+            //
+            // cur = proj_w*cur + proj_b
+            // [768, N]
+            cur = ggml_mul_mat(ctx0,
+                    model.layers[il].c_mlp_proj_w,
+                    cur);
+
+            cur = ggml_add(ctx0,
+                    ggml_repeat(ctx0, model.layers[il].c_mlp_proj_b, cur),
+                    cur);
+        }
+
+        // input for next layer
+        inpL = ggml_add(ctx0, cur, inpFF);
+    }
+
+    ggml_set_scratch(ctx0, { 0, scr0_size, scr0, });
+
+    // norm
+    {
+        // [ 768, N]
+        inpL = ggml_norm(ctx0, inpL, hparams.eps);
+
+        // inpL = ln_f_g*inpL + ln_f_b
+        // [ 768, N]
+        inpL = ggml_add(ctx0,
+                ggml_mul(ctx0,
+                    ggml_repeat(ctx0, model.ln_f_g, inpL),
+                    inpL),
+                ggml_repeat(ctx0, model.ln_f_b, inpL));
+    }
+
+    ggml_set_scratch(ctx0, { 0, 0, nullptr, });
+
+    // inpL = WTE * inpL
+    // [ 768, 50257] - model.lm_head
+    // [ 768, N]     - inpL
+    inpL = ggml_mul_mat(ctx0, model.lm_head, inpL);
+
+    // logits -> probs
+    //inpL = ggml_soft_max_inplace(ctx0, inpL);
+
+    // run the computation
+    ggml_build_forward_expand(&gf, inpL);
+    ggml_graph_compute_with_ctx(ctx0, &gf, n_threads);
+
+    //if (n_past%100 == 0) {
+    //    ggml_graph_print   (&gf);
+    //    ggml_graph_dump_dot(&gf, NULL, "gpt-2.dot");
+    //}
+
+    //embd_w.resize(n_vocab*N);
+    //memcpy(embd_w.data(), ggml_get_data(inpL), sizeof(float)*n_vocab*N);
+
+    // return result just for the last token
+    embd_w.resize(n_vocab);
+    memcpy(embd_w.data(), (float *) ggml_get_data(inpL) + (n_vocab*(N-1)), sizeof(float)*n_vocab);
+
+    if (mem_per_token == 0) {
+        mem_per_token = ggml_used_mem(ctx0)/N;
+    }
+    //printf("used_mem = %zu MB\n", ggml_used_mem(ctx0)/(1024*1024));
+
+    ggml_free(ctx0);
+
+    return true;
+}
+
+int main(int argc, char ** argv) {
+    ggml_time_init();
+
+    const int64_t t_main_start_us = ggml_time_us();
+
+    gpt_params params;
+
+    if (gpt_params_parse(argc, argv, params) == false) {
+        return 1;
+    }
+
+    if (params.seed < 0) {
+        params.seed = time(NULL);
+    }
+
+    printf("%s: seed = %d\n", __func__, params.seed);
+
+    std::mt19937 rng(params.seed);
+    if (params.prompt.empty()) {
+        params.prompt = gpt_random_prompt(rng);
+    }
+
+    int64_t t_load_us = 0;
+
+    gpt_vocab vocab;
+    starcoder_model model;
+
+    // load the model
+    {
+        const int64_t t_start_us = ggml_time_us();
+
+        if (!starcoder_model_load(params.model, model, vocab)) {
+            fprintf(stderr, "%s: failed to load model from '%s'\n", __func__, params.model.c_str());
+            return 1;
+        }
+
+        t_load_us = ggml_time_us() - t_start_us;
+
+        test_gpt_tokenizer(vocab, params.token_test);
+    }
+
+    if (params.repeat_last_n == -1) {
+        params.repeat_last_n = model.hparams.n_ctx;
+    }
+    printf("\n");
+    printf("%s: temp           = %.3f\n", __func__, params.temp);
+    printf("%s: top_k          = %d\n",   __func__, params.top_k);
+    printf("%s: top_p          = %.3f\n", __func__, params.top_p);
+    printf("%s: repeat_last_n  = %d\n",   __func__, params.repeat_last_n);
+    printf("%s: repeat_penalty = %.3f\n", __func__, params.repeat_penalty);
+
+    int n_past = 0;
+
+    int64_t t_sample_us  = 0;
+    int64_t t_predict_us = 0;
+
+    std::vector<float> logits;
+
+    std::vector<int32_t> last_n_tokens(model.hparams.n_ctx);
+    std::fill(last_n_tokens.begin(), last_n_tokens.end(), 0);
+
+    // tokenize the prompt
+    std::vector<gpt_vocab::id> embd_inp = ::gpt_tokenize(vocab, params.prompt);
+
+    params.n_predict = std::min(params.n_predict, model.hparams.n_ctx - (int) embd_inp.size());
+
+    printf("%s: prompt: '%s'\n", __func__, params.prompt.c_str());
+    printf("%s: number of tokens in prompt = %zu\n", __func__, embd_inp.size());
+    for (size_t i = 0; i < embd_inp.size(); i++) {
+        printf("%s: token[%zu] = %6d, %s\n", __func__, i, embd_inp[i], vocab.id_to_token.at(embd_inp[i]).c_str());
+    }
+    printf("\n\n");
+
+    // Handle StarChat "<|end|>" and OpenCoder "<|end_of_turn>" tokens.
+    gpt_vocab::id starchat_end_token = -1;
+    {
+        const auto it = vocab.token_to_id.find("<|end|>");
+        if (it != vocab.token_to_id.end()) {
+            starchat_end_token = it->second;
+        } else {
+            const auto eot_token_id = vocab.token_to_id.find("<|end_of_turn|>");
+            if (eot_token_id != vocab.token_to_id.end()) {
+              starchat_end_token = eot_token_id->second;
+            }
+        }
+    }
+
+    // submit the input prompt token-by-token
+    // this reduces the memory usage during inference, at the cost of a bit of speed at the beginning
+    std::vector<gpt_vocab::id> embd;
+
+    // determine the required inference memory per token:
+    size_t mem_per_token = 0;
+    starcoder_eval(model, params.n_threads, 0, { 0, 1, 2, 3 }, logits, mem_per_token);
+
+    for (size_t i = embd.size(); i < embd_inp.size() + params.n_predict; i++) {
+        // predict
+        if (embd.size() > 0) {
+            const int64_t t_start_us = ggml_time_us();
+
+            if (!starcoder_eval(model, params.n_threads, n_past, embd, logits, mem_per_token)) {
+                printf("Failed to predict\n");
+                return 1;
+            }
+
+            t_predict_us += ggml_time_us() - t_start_us;
+        }
+
+        n_past += embd.size();
+        embd.clear();
+
+        if (i >= embd_inp.size()) {
+            // sample next token
+            const int   top_k = params.top_k;
+            const float top_p = params.top_p;
+            const float temp  = params.temp;
+
+            const int n_vocab = model.hparams.n_vocab;
+
+            gpt_vocab::id id = 0;
+
+            {
+                const int64_t t_start_sample_us = ggml_time_us();
+
+                id = gpt_sample_top_k_top_p_repeat(vocab, logits.data() + (logits.size() - n_vocab), last_n_tokens.data(), last_n_tokens.size(), top_k, top_p, temp, params.repeat_last_n, params.repeat_penalty, rng);
+                t_sample_us += ggml_time_us() - t_start_sample_us;
+            }
+
+            // add it to the context
+            embd.push_back(id);
+
+            last_n_tokens.erase(last_n_tokens.begin());
+            last_n_tokens.push_back(id);
+        } else {
+            // if here, it means we are still processing the input prompt
+            for (size_t k = i; k < embd_inp.size(); k++) {
+                embd.push_back(embd_inp[k]);
+
+                last_n_tokens.erase(last_n_tokens.begin());
+                last_n_tokens.push_back(embd_inp[k]);
+
+                if (int32_t(embd.size()) >= params.n_batch) {
+                    break;
+                }
+            }
+            i += embd.size() - 1;
+        }
+
+        // display text
+        for (auto id : embd) {
+            printf("%s", vocab.id_to_token[id].c_str());
+        }
+        fflush(stdout);
+
+        // check if model is santacoder
+        if (model.hparams.n_layer <= 30 && embd.back() == 49152) {
+            break;
+        }
+        // check if model is starcoder
+        else if (embd.back() == 0) { //TODO: this is only for starcoder
+            break;
+        }
+        // Handle StarChat "<|end|>" token.
+        else if (embd.back() == starchat_end_token && i >= embd_inp.size()) {
+            break;
+        }
+    }
+
+    // report timing
+    {
+        const int64_t t_main_end_us = ggml_time_us();
+
+        printf("\n\n");
+        printf("%s: mem per token = %8zu bytes\n", __func__, mem_per_token);
+        printf("%s:     load time = %8.2f ms\n", __func__, t_load_us/1000.0f);
+        printf("%s:   sample time = %8.2f ms\n", __func__, t_sample_us/1000.0f);
+        printf("%s:  predict time = %8.2f ms / %.2f ms per token\n", __func__, t_predict_us/1000.0f, t_predict_us/1000.0f/n_past);
+        printf("%s:    total time = %8.2f ms\n", __func__, (t_main_end_us - t_main_start_us)/1000.0f);
+    }
+
+    ggml_free(model.ctx);
+
+    return 0;
+}

ggml/examples/starcoder/quantize.cpp

@@ -0,0 +1,184 @@
+#include "ggml/ggml.h"
+
+#include "common.h"
+#include "common-ggml.h"
+
+#include <cassert>
+#include <cmath>
+#include <cstdio>
+#include <cstring>
+#include <fstream>
+#include <map>
+#include <string>
+#include <vector>
+#include <regex>
+
+// default hparams (GPT-2 117M)
+struct starcoder_hparams {
+    int32_t n_vocab = 49280;
+    int32_t n_ctx   = 2048;
+    int32_t n_embd  = 2048;
+    int32_t n_head  = 16;
+    int32_t n_layer = 24;
+    int32_t ftype   = 1;
+};
+
+// quantize a model
+bool starcoder_model_quantize(const std::string & fname_inp, const std::string & fname_out, ggml_ftype ftype) {
+    gpt_vocab vocab;
+
+    printf("%s: loading model from '%s'\n", __func__, fname_inp.c_str());
+
+    auto finp = std::ifstream(fname_inp, std::ios::binary);
+    if (!finp) {
+        fprintf(stderr, "%s: failed to open '%s' for reading\n", __func__, fname_inp.c_str());
+        return false;
+    }
+
+    auto fout = std::ofstream(fname_out, std::ios::binary);
+    if (!fout) {
+        fprintf(stderr, "%s: failed to open '%s' for writing\n", __func__, fname_out.c_str());
+        return false;
+    }
+
+    // verify magic
+    {
+        uint32_t magic;
+        finp.read((char *) &magic, sizeof(magic));
+        if (magic != GGML_FILE_MAGIC) {
+            fprintf(stderr, "%s: invalid model file '%s' (bad magic)\n", __func__, fname_inp.c_str());
+            return false;
+        }
+
+        fout.write((char *) &magic, sizeof(magic));
+    }
+
+    starcoder_hparams hparams;
+
+    // load hparams
+    {
+        finp.read((char *) &hparams.n_vocab, sizeof(hparams.n_vocab));
+        finp.read((char *) &hparams.n_ctx,   sizeof(hparams.n_ctx));
+        finp.read((char *) &hparams.n_embd,  sizeof(hparams.n_embd));
+        finp.read((char *) &hparams.n_head,  sizeof(hparams.n_head));
+        finp.read((char *) &hparams.n_layer, sizeof(hparams.n_layer));
+        finp.read((char *) &hparams.ftype,   sizeof(hparams.ftype));
+
+        const int32_t qntvr_src =    hparams.ftype / GGML_QNT_VERSION_FACTOR;
+        const int32_t ftype_dst = GGML_QNT_VERSION * GGML_QNT_VERSION_FACTOR + ftype;
+
+        printf("%s: n_vocab     = %d\n", __func__, hparams.n_vocab);
+        printf("%s: n_ctx       = %d\n", __func__, hparams.n_ctx);
+        printf("%s: n_embd      = %d\n", __func__, hparams.n_embd);
+        printf("%s: n_head      = %d\n", __func__, hparams.n_head);
+        printf("%s: n_layer     = %d\n", __func__, hparams.n_layer);
+        printf("%s: ftype (src) = %d\n", __func__, hparams.ftype);
+        printf("%s: qntvr (src) = %d\n", __func__, qntvr_src);
+        printf("%s: ftype (dst) = %d\n", __func__, ftype_dst);
+        printf("%s: qntvr (dst) = %d\n", __func__, GGML_QNT_VERSION);
+
+        fout.write((char *) &hparams.n_vocab, sizeof(hparams.n_vocab));
+        fout.write((char *) &hparams.n_ctx,   sizeof(hparams.n_ctx));
+        fout.write((char *) &hparams.n_embd,  sizeof(hparams.n_embd));
+        fout.write((char *) &hparams.n_head,  sizeof(hparams.n_head));
+        fout.write((char *) &hparams.n_layer, sizeof(hparams.n_layer));
+        fout.write((char *) &ftype_dst,       sizeof(ftype_dst));
+    }
+
+    // load vocab
+    {
+        int32_t n_vocab = 0;
+        finp.read ((char *) &n_vocab, sizeof(n_vocab));
+        fout.write((char *) &n_vocab, sizeof(n_vocab));
+
+        if (n_vocab != hparams.n_vocab) {
+            fprintf(stderr, "%s: invalid model file '%s' (bad vocab size %d != %d)\n",
+                    __func__, fname_inp.c_str(), n_vocab, hparams.n_vocab);
+            return false;
+        }
+
+        std::string word;
+        for (int i = 0; i < n_vocab; i++) {
+            uint32_t len;
+            finp.read ((char *) &len, sizeof(len));
+            fout.write((char *) &len, sizeof(len));
+
+            word.resize(len);
+            finp.read ((char *) word.data(), len);
+            fout.write((char *) word.data(), len);
+
+            vocab.token_to_id[word] = i;
+            vocab.id_to_token[i] = word;
+        }
+    }
+
+    // regexes of tensor names to be quantized
+    const std::vector<std::string> to_quant = {
+        "model/wte",
+        "model/lm_head",
+        "model/h.*/attn/c_attn/w",
+        "model/h.*/attn/c_proj/w",
+        "model/h.*/mlp/c_fc/w",
+        "model/h.*/mlp/c_proj/w",
+    };
+
+    if (!ggml_common_quantize_0(finp, fout, ftype, to_quant, {})) {
+        fprintf(stderr, "%s: failed to quantize model '%s'\n", __func__, fname_inp.c_str());
+        return false;
+    }
+
+    finp.close();
+    fout.close();
+
+    return true;
+}
+
+// usage:
+//  ./gpt-2-quantize models/gpt-2-117M/ggml-model.bin models/gpt-2-117M/ggml-model-quant.bin type
+//
+int main(int argc, char ** argv) {
+    if (argc != 4) {
+        fprintf(stderr, "usage: %s model-f32.bin model-quant.bin type\n", argv[0]);
+        ggml_print_ftypes(stderr);
+        return 1;
+    }
+
+    // needed to initialize f16 tables
+    {
+        struct ggml_init_params params = { 0, NULL, false };
+        struct ggml_context * ctx = ggml_init(params);
+        ggml_free(ctx);
+    }
+
+    const std::string fname_inp = argv[1];
+    const std::string fname_out = argv[2];
+
+    const ggml_ftype ftype = ggml_parse_ftype(argv[3]);
+
+    const int64_t t_main_start_us = ggml_time_us();
+
+    int64_t t_quantize_us = 0;
+
+    // load the model
+    {
+        const int64_t t_start_us = ggml_time_us();
+
+        if (!starcoder_model_quantize(fname_inp, fname_out, ggml_ftype(ftype))) {
+            fprintf(stderr, "%s: failed to quantize model from '%s'\n", __func__, fname_inp.c_str());
+            return 1;
+        }
+
+        t_quantize_us = ggml_time_us() - t_start_us;
+    }
+
+    // report timing
+    {
+        const int64_t t_main_end_us = ggml_time_us();
+
+        printf("\n");
+        printf("%s: quantize time = %8.2f ms\n", __func__, t_quantize_us/1000.0f);
+        printf("%s:    total time = %8.2f ms\n", __func__, (t_main_end_us - t_main_start_us)/1000.0f);
+    }
+
+    return 0;
+}

ggml/examples/starcoder/starcoder-mmap.cpp

@@ -0,0 +1,1127 @@
+#include "ggml/ggml.h"
+
+#include "common.h"
+#include "common-ggml.h"
+
+#include <cassert>
+#include <cmath>
+#include <cstdio>
+#include <cstring>
+#include <fstream>
+#include <map>
+#include <string>
+#include <vector>
+
+#if !defined(_WIN32)
+// mmap
+#include <sys/types.h>
+#include <sys/mman.h>
+#include <unistd.h>
+#include <fcntl.h>
+#else
+#define NOMINMAX
+#include <Windows.h>
+#endif
+
+#ifdef GGML_USE_CUBLAS
+#include "ggml-cuda.h"
+#endif
+
+#ifdef GGML_USE_CLBLAST
+#include "ggml-opencl.h"
+#endif
+
+// default hparams (GPT-2 117M)
+// https://huggingface.co/bigcode/gpt_bigcode-santacoder/blob/main/config.json
+struct starcoder_hparams {
+    int32_t n_vocab = 49280;
+    int32_t n_ctx   = 2048;
+    int32_t n_embd  = 2048;
+    int32_t n_head  = 16;
+    int32_t n_layer = 24;
+    int32_t ftype   = 1;
+    float   eps     = 1e-5f;
+};
+
+struct starcoder_layer {
+    // normalization
+    struct ggml_tensor * ln_1_g;
+    struct ggml_tensor * ln_1_b;
+
+    struct ggml_tensor * ln_2_g;
+    struct ggml_tensor * ln_2_b;
+
+    // attention
+    struct ggml_tensor * c_attn_attn_w;
+    struct ggml_tensor * c_attn_attn_b;
+
+    struct ggml_tensor * c_attn_proj_w;
+    struct ggml_tensor * c_attn_proj_b;
+
+    // mlp
+    struct ggml_tensor * c_mlp_fc_w;
+    struct ggml_tensor * c_mlp_fc_b;
+
+    struct ggml_tensor * c_mlp_proj_w;
+    struct ggml_tensor * c_mlp_proj_b;
+};
+
+struct llama_buffer {
+    uint8_t * addr = NULL;
+    size_t size = 0;
+
+    llama_buffer() = default;
+
+    void resize(size_t len) {
+#ifdef GGML_USE_METAL
+        free(addr);
+        int result = posix_memalign((void **) &addr, getpagesize(), len);
+        if (result == 0) {
+            memset(addr, 0, len);
+        }
+        else {
+            addr = NULL;
+        }
+#else
+        delete[] addr;
+        addr = new uint8_t[len];
+#endif
+        size = len;
+    }
+
+    ~llama_buffer() {
+#ifdef GGML_USE_METAL
+        free(addr);
+#else
+        delete[] addr;
+#endif
+        addr = NULL;
+    }
+
+    // disable copy and move
+    llama_buffer(const llama_buffer&) = delete;
+    llama_buffer(llama_buffer&&) = delete;
+    llama_buffer& operator=(const llama_buffer&) = delete;
+    llama_buffer& operator=(llama_buffer&&) = delete;
+};
+
+
+struct kv_cache {
+    struct ggml_tensor * k;
+    struct ggml_tensor * v;
+
+    struct ggml_context * ctx = NULL;
+
+    //std::vector<uint8_t> buf;
+    llama_buffer buf;
+
+    int n;
+};
+
+struct starcoder_model {
+    starcoder_hparams hparams;
+
+    // normalization
+    struct ggml_tensor * ln_f_g;
+    struct ggml_tensor * ln_f_b;
+
+    struct ggml_tensor * wte;     // position embedding
+    struct ggml_tensor * wpe;     //    token embedding
+    struct ggml_tensor * lm_head; // language model head
+
+    std::vector<starcoder_layer> layers;
+
+    // key + value memory
+    //struct ggml_tensor * memory_k;
+    //struct ggml_tensor * memory_v;
+    struct kv_cache cache;
+
+    // model memory mapped file
+    void * mm_addr = NULL;
+    uint64_t mm_length = 0;
+
+    //
+    struct ggml_context * ctx;
+    std::map<std::string, struct ggml_tensor *> tensors;
+};
+
+// From PR #613 (https://github.com/ggerganov/llama.cpp/pull/613)
+static void *mmap_file(const char *fname, uint64_t *mm_length) {
+#if defined(_WIN32) && !defined(_POSIX_MAPPED_FILES)
+    HANDLE hFile = CreateFileA(fname,
+                               GENERIC_READ,
+                               FILE_SHARE_READ | FILE_SHARE_WRITE | FILE_SHARE_DELETE,
+                               NULL,
+                               OPEN_EXISTING,
+                               FILE_ATTRIBUTE_NORMAL | FILE_ATTRIBUTE_NOT_CONTENT_INDEXED,
+                               NULL);
+    if (hFile == INVALID_HANDLE_VALUE) return 0;
+    LARGE_INTEGER fileSize;
+    fileSize.QuadPart = -1;
+    GetFileSizeEx(hFile, &fileSize);
+    int64_t length = fileSize.QuadPart;
+    HANDLE hMapping = CreateFileMappingA(hFile, NULL, PAGE_READONLY, 0, 0, NULL);
+    CloseHandle(hFile);
+    if (!hMapping) return 0;
+    void *addr = MapViewOfFile(hMapping, FILE_MAP_READ, 0, 0, 0);
+    CloseHandle(hMapping);
+    if (!addr) return 0;
+#else
+    int fd = open(fname, O_RDONLY);
+    if (fd == -1) return 0;
+    int64_t length = lseek(fd, 0, SEEK_END);
+    void *addr = mmap(NULL, length, PROT_READ, MAP_SHARED, fd, 0);
+    close(fd);
+    if (addr == MAP_FAILED) return 0;
+#endif
+    *mm_length = length;
+    return addr;
+}
+
+static void munmap_file(void * addr, size_t length) {
+#if defined(_WIN32) && !defined(_POSIX_MAPPED_FILES)
+    UnmapViewOfFile(addr);
+#else
+    munmap(addr, length);
+#endif
+}
+
+// load the model's weights from a file
+bool starcoder_model_load(const std::string & fname, starcoder_model & model, gpt_vocab & vocab, int32_t n_gpu_layers) {
+    printf("%s: loading model from '%s'\n", __func__, fname.c_str());
+
+    auto fin = std::ifstream(fname, std::ios::binary);
+    if (!fin) {
+        fprintf(stderr, "%s: failed to open '%s'\n", __func__, fname.c_str());
+        return false;
+    }
+
+    std::vector<char> f_buf(1024*1024);
+    fin.rdbuf()->pubsetbuf(f_buf.data(), f_buf.size());
+
+
+    // verify magic
+    {
+        uint32_t magic;
+        fin.read((char *) &magic, sizeof(magic));
+        //if (magic != 0x67676a74) {
+        if (magic != 0x67676d6c) {
+            fprintf(stderr, "%s: invalid model file '%s' (bad magic)\n", __func__, fname.c_str());
+            return false;
+        }
+    }
+
+    // load hparams
+    {
+        auto & hparams = model.hparams;
+
+        fin.read((char *) &hparams.n_vocab, sizeof(hparams.n_vocab));
+        fin.read((char *) &hparams.n_ctx,   sizeof(hparams.n_ctx));
+        fin.read((char *) &hparams.n_embd,  sizeof(hparams.n_embd));
+        fin.read((char *) &hparams.n_head,  sizeof(hparams.n_head));
+        fin.read((char *) &hparams.n_layer, sizeof(hparams.n_layer));
+        fin.read((char *) &hparams.ftype,   sizeof(hparams.ftype));
+
+        const int32_t qntvr = hparams.ftype / GGML_QNT_VERSION_FACTOR;
+
+        printf("%s: n_vocab = %d\n", __func__, hparams.n_vocab);
+        printf("%s: n_ctx   = %d\n", __func__, hparams.n_ctx);
+        printf("%s: n_embd  = %d\n", __func__, hparams.n_embd);
+        printf("%s: n_head  = %d\n", __func__, hparams.n_head);
+        printf("%s: n_layer = %d\n", __func__, hparams.n_layer);
+        printf("%s: ftype   = %d\n", __func__, hparams.ftype);
+        printf("%s: qntvr   = %d\n", __func__, qntvr);
+
+        hparams.ftype %= GGML_QNT_VERSION_FACTOR;
+    }
+
+    // load vocab
+    {
+        int32_t n_vocab = 0;
+        fin.read((char *) &n_vocab, sizeof(n_vocab));
+
+        if (n_vocab != model.hparams.n_vocab) {
+            fprintf(stderr, "%s: invalid model file '%s' (bad vocab size %d != %d)\n",
+                    __func__, fname.c_str(), n_vocab, model.hparams.n_vocab);
+            return false;
+        }
+
+        std::string word;
+        std::vector<char> buf(128);
+
+        for (int i = 0; i < n_vocab; i++) {
+            uint32_t len;
+            fin.read((char *) &len, sizeof(len));
+
+            buf.resize(len);
+            fin.read((char *) buf.data(), len);
+            word.assign(buf.data(), len);
+
+            vocab.token_to_id[word] = i;
+            vocab.id_to_token[i] = word;
+
+            // if (i < 10) fprintf(stderr, "%.s: vocab[%d] = '%s'\n", __func__, i, word.c_str());
+        }
+
+        // Add StarChat special tokens.
+        for (std::string token : {
+                "<|system|>",
+                "<|user|>",
+                "<|assistant|>",
+                "<|end|>",
+                }) {
+            if (vocab.token_to_id.find(token) != vocab.token_to_id.end()) {
+                vocab.add_special_token(token);
+            }
+        }
+    }
+
+    char *mm_addr = NULL;
+    model.mm_addr = mmap_file(fname.c_str(), &model.mm_length);
+    if (model.mm_addr == NULL) {
+        fprintf(stderr, "%s: failed to mmap '%s'\n", __func__, fname.c_str());
+        return false;
+    }
+    mm_addr = (char *)model.mm_addr;
+    fprintf(stderr, "%s: ggml map size = %6.2f MB\n", __func__, model.mm_length/(1024.0*1024.0));
+
+    // for the big tensors, we have the option to store the data in 16-bit floats or quantized
+    // in order to save memory and also to speed up the computation
+    ggml_type wtype = ggml_ftype_to_ggml_type((ggml_ftype) (model.hparams.ftype));
+    if (wtype == GGML_TYPE_COUNT) {
+        fprintf(stderr, "%s: invalid model file '%s' (bad ftype value %d)\n",
+                __func__, fname.c_str(), model.hparams.ftype);
+        return false;
+    }
+
+    auto & ctx = model.ctx;
+
+    size_t ctx_size = 0;
+
+    {
+        const auto & hparams = model.hparams;
+
+        const int n_layer = hparams.n_layer;
+
+
+        /*
+        const int n_embd  = hparams.n_embd;
+        const int n_layer = hparams.n_layer;
+        const int n_ctx   = hparams.n_ctx;
+        const int n_vocab = hparams.n_vocab;
+
+        const int head_dim = n_embd / hparams.n_head;
+        const int kv_heads = hparams.n_head; // 1 if MQA else hparams.n_head
+        const int kv_dim   = kv_heads * head_dim;
+
+
+        ctx_size += n_embd*ggml_type_sizef(GGML_TYPE_F32); // ln_f_g
+        ctx_size += n_embd*ggml_type_sizef(GGML_TYPE_F32); // ln_f_b
+
+        ctx_size += n_vocab*n_embd*ggml_type_sizef(wtype);         // wte
+        ctx_size +=   n_ctx*n_embd*ggml_type_sizef(GGML_TYPE_F32); // wpe
+        ctx_size += n_vocab*n_embd*ggml_type_sizef(wtype);         // lm_head
+
+        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_1_g
+        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_1_b
+
+        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_2_g
+        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_2_b
+
+        ctx_size += n_layer*((n_embd + 2*kv_dim)*n_embd*ggml_type_sizef(wtype));         // c_attn_attn_w // TODO:
+        ctx_size += n_layer*(       (n_embd + 2*kv_dim)*ggml_type_sizef(GGML_TYPE_F32)); // c_attn_attn_b
+
+        ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype));           // c_attn_proj_w
+        ctx_size += n_layer*(       n_embd*ggml_type_sizef(GGML_TYPE_F32));   // c_attn_proj_b
+
+        ctx_size += n_layer*(4*n_embd*n_embd*ggml_type_sizef(wtype));         // c_mlp_fc_w
+        ctx_size += n_layer*(       4*n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_mlp_fc_b
+
+        ctx_size += n_layer*(4*n_embd*n_embd*ggml_type_sizef(wtype));         // c_mlp_proj_w
+        ctx_size += n_layer*(         n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_mlp_proj_b
+
+        ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(GGML_TYPE_F32); // memory_k
+        ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(GGML_TYPE_F32); // memory_v
+        */
+
+        ctx_size += (6 + 12*n_layer)*512; // object overhead
+
+        printf("%s: ggml ctx size = %6.2f MB\n", __func__, ctx_size/(1024.0*1024.0));
+    }
+
+    // create the ggml context
+    {
+        struct ggml_init_params params = {
+            /*.mem_size   =*/ ctx_size,
+            /*.mem_buffer =*/ NULL,
+            /*.no_alloc   =*/ true,
+        };
+
+        model.ctx = ggml_init(params);
+        if (!model.ctx) {
+            fprintf(stderr, "%s: ggml_init() failed\n", __func__);
+            return false;
+        }
+    }
+
+    // prepare memory for the weights
+    {
+        const auto & hparams = model.hparams;
+
+        const int n_embd  = hparams.n_embd;
+        const int n_layer = hparams.n_layer;
+        const int n_ctx   = hparams.n_ctx;
+        const int n_vocab = hparams.n_vocab;
+
+        const int head_dim = n_embd / hparams.n_head;
+        const int kv_heads = hparams.n_head; // 1 if MQA else hparams.n_head
+        const int kv_dim   = kv_heads * head_dim;
+
+        model.layers.resize(n_layer);
+
+        model.ln_f_g = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
+        model.ln_f_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
+
+        model.wte     = ggml_new_tensor_2d(ctx, wtype,         n_embd, n_vocab);
+        model.wpe     = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_ctx);
+        model.lm_head = ggml_new_tensor_2d(ctx, wtype,         n_embd, n_vocab);
+
+        // map by name
+        model.tensors["model/ln_f/g"] = model.ln_f_g;
+        model.tensors["model/ln_f/b"] = model.ln_f_b;
+
+        model.tensors["model/wte"]     = model.wte;
+        model.tensors["model/wpe"]     = model.wpe;
+        model.tensors["model/lm_head"] = model.lm_head;
+
+        for (int i = 0; i < n_layer; ++i) {
+            auto & layer = model.layers[i];
+
+            layer.ln_1_g        = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+            layer.ln_1_b        = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+
+            layer.ln_2_g        = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+            layer.ln_2_b        = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+
+            layer.c_attn_attn_w = ggml_new_tensor_2d(ctx, wtype,           n_embd, n_embd + 2*kv_dim);
+            layer.c_attn_attn_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd + 2*kv_dim);
+
+            layer.c_attn_proj_w = ggml_new_tensor_2d(ctx, wtype,           n_embd, n_embd);
+            layer.c_attn_proj_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+
+            layer.c_mlp_fc_w    = ggml_new_tensor_2d(ctx, wtype,           n_embd, 4*n_embd); //TODO: 4*n_embd = config.n_inner
+            layer.c_mlp_fc_b    = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 4*n_embd);
+
+            layer.c_mlp_proj_w  = ggml_new_tensor_2d(ctx, wtype,         4*n_embd, n_embd);
+            layer.c_mlp_proj_b  = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+
+            // map by name
+            model.tensors["model/h" + std::to_string(i) + "/ln_1/g"]        = layer.ln_1_g;
+            model.tensors["model/h" + std::to_string(i) + "/ln_1/b"]        = layer.ln_1_b;
+
+            model.tensors["model/h" + std::to_string(i) + "/ln_2/g"]        = layer.ln_2_g;
+            model.tensors["model/h" + std::to_string(i) + "/ln_2/b"]        = layer.ln_2_b;
+
+            model.tensors["model/h" + std::to_string(i) + "/attn/c_attn/w"] = layer.c_attn_attn_w;
+            model.tensors["model/h" + std::to_string(i) + "/attn/c_attn/b"] = layer.c_attn_attn_b;
+
+            model.tensors["model/h" + std::to_string(i) + "/attn/c_proj/w"] = layer.c_attn_proj_w;
+            model.tensors["model/h" + std::to_string(i) + "/attn/c_proj/b"] = layer.c_attn_proj_b;
+
+            model.tensors["model/h" + std::to_string(i) + "/mlp/c_fc/w"]    = layer.c_mlp_fc_w;
+            model.tensors["model/h" + std::to_string(i) + "/mlp/c_fc/b"]    = layer.c_mlp_fc_b;
+
+            model.tensors["model/h" + std::to_string(i) + "/mlp/c_proj/w"]  = layer.c_mlp_proj_w;
+            model.tensors["model/h" + std::to_string(i) + "/mlp/c_proj/b"]  = layer.c_mlp_proj_b;
+        }
+    }
+
+    // key + value memory
+    {
+        const auto & hparams = model.hparams;
+
+        const int n_embd  = hparams.n_embd;
+        const int n_layer = hparams.n_layer;
+        const int n_ctx   = hparams.n_ctx;
+
+        const int n_mem      = n_layer*n_ctx;
+        const int n_elements = n_embd*n_mem;
+
+        model.cache.buf.resize(2u*n_elements*ggml_type_size(GGML_TYPE_F16) + 2u*1024*1024);
+
+        struct ggml_init_params c_params;
+        c_params.mem_size   = model.cache.buf.size;
+        c_params.mem_buffer = model.cache.buf.addr;
+        c_params.no_alloc   = false;
+
+        model.cache.ctx = ggml_init(c_params);
+
+        if (!model.cache.ctx) {
+            fprintf(stderr, "%s: failed to allocate memory for kv cache\n", __func__);
+            return false;
+        }
+
+        model.cache.k = ggml_new_tensor_1d(model.cache.ctx, GGML_TYPE_F16, n_elements);
+        model.cache.v = ggml_new_tensor_1d(model.cache.ctx, GGML_TYPE_F16, n_elements);
+
+        const size_t memory_size = ggml_nbytes(model.cache.k) + ggml_nbytes(model.cache.v);
+
+        printf("%s: kv_cache memory size = %8.2f MB, n_mem = %d\n", __func__, memory_size/1024.0/1024.0, n_mem);
+    }
+
+    // load weights
+    {
+        size_t total_size = 0;
+
+        bool has_lm_head = false;
+
+        while (true) {
+            int32_t n_dims;
+            int32_t length;
+            int32_t ttype;
+
+            fin.read(reinterpret_cast<char *>(&n_dims), sizeof(n_dims));
+            fin.read(reinterpret_cast<char *>(&length), sizeof(length));
+            fin.read(reinterpret_cast<char *>(&ttype),  sizeof(ttype));
+
+            if (fin.eof()) {
+                break;
+            }
+
+            int32_t nelements = 1;
+            int32_t ne[2] = { 1, 1 };
+            for (int i = 0; i < n_dims; ++i) {
+                fin.read(reinterpret_cast<char *>(&ne[i]), sizeof(ne[i]));
+                nelements *= ne[i];
+            }
+
+            std::string name(length, 0);
+            fin.read(&name[0], length);
+
+            if (model.tensors.find(name.data()) == model.tensors.end()) {
+                fprintf(stderr, "%s: unknown tensor '%s' in model file\n", __func__, name.data());
+                return false;
+            }
+
+            auto tensor = model.tensors[name.data()];
+
+            if (tensor->ne[0] != ne[0] || tensor->ne[1] != ne[1]) {
+                fprintf(stderr, "%s: tensor '%s' has wrong shape in model file: got [%d, %d], expected [%d, %d]\n",
+                        __func__, name.data(), (int) tensor->ne[0], (int) tensor->ne[1], ne[0], ne[1]);
+                return false;
+            }
+            if (ggml_nelements(tensor) != nelements) {
+                fprintf(stderr, "%s: tensor '%s' has wrong size in model file. got %d, expected %d\n",
+                        __func__, name.data(), (int) ggml_nelements(tensor), nelements);
+                return false;
+            }
+
+            // for debugging
+            if (0) {
+                printf("%24s - [%5d, %5d], type = %6s, %6.2f MB, %9zu bytes\n", name.data(), ne[0], ne[1], ggml_type_name(ggml_type(ttype)), ggml_nbytes(tensor)/1024.0/1024.0, ggml_nbytes(tensor));
+            }
+
+            const size_t bpe = ggml_type_size(ggml_type(ttype));
+
+            if ((nelements*bpe)/ggml_blck_size(tensor->type) != ggml_nbytes(tensor)) {
+                fprintf(stderr, "%s: tensor '%s' has wrong size in model file: got %zu, expected %zu\n",
+                        __func__, name.data(), ggml_nbytes(tensor), nelements*bpe);
+                return false;
+            }
+
+            // mmap
+            size_t offset = fin.tellg();
+            size_t tensor_data_size = ggml_nbytes(tensor);
+            //offset = (offset + 31) & -32;
+            tensor->data = mm_addr + offset;
+            fin.seekg(offset + tensor_data_size);
+            total_size += tensor_data_size;
+
+            // GPT-2 models share the WTE tensor as the LM head
+            if (name == "model/wte" && has_lm_head == false) {
+                // Dont know if this is required, test models have an lm_head
+                model.lm_head->data = tensor->data;
+            }
+
+            if (name == "model/lm_head") {
+                has_lm_head = true;
+            }
+        }
+
+        printf("%s: model size  = %8.2f MB\n", __func__, total_size/1024.0/1024.0);
+    }
+
+    fin.close();
+
+#ifdef GGML_USE_CUBLAS
+    {
+        const auto & hparams = model.hparams;
+        const int n_gpu = std::min(n_gpu_layers, int(hparams.n_layer));
+
+        fprintf(stderr, "%s: [cublas] offloading %d layers to GPU\n", __func__, n_gpu);
+
+        size_t vram_total = 0;
+
+        for (int i = 0; i < n_gpu; ++i) {
+            const auto & layer = model.layers[i];
+
+            layer.c_attn_attn_w->backend = GGML_BACKEND_GPU;
+            ggml_cuda_transform_tensor((uint8_t *)layer.c_attn_attn_w->data, layer.c_attn_attn_w); vram_total += ggml_nbytes(layer.c_attn_attn_w);
+
+            layer.c_attn_proj_w->backend = GGML_BACKEND_GPU;
+            ggml_cuda_transform_tensor((uint8_t *)layer.c_attn_proj_w->data, layer.c_attn_proj_w); vram_total += ggml_nbytes(layer.c_attn_proj_w);
+
+            layer.c_mlp_fc_w->backend = GGML_BACKEND_GPU;
+            ggml_cuda_transform_tensor((uint8_t *)layer.c_mlp_fc_w->data, layer.c_mlp_fc_w); vram_total += ggml_nbytes(layer.c_mlp_fc_w);
+
+            layer.c_mlp_proj_w->backend = GGML_BACKEND_GPU;
+            ggml_cuda_transform_tensor((uint8_t *)layer.c_mlp_proj_w->data, layer.c_mlp_proj_w); vram_total += ggml_nbytes(layer.c_mlp_proj_w);
+        }
+
+        ggml_cuda_set_scratch_size(0); // disable scratch
+
+        //if (n_gpu_layers > (int) hparams.n_layer) {
+        //    fprintf(stderr, "%s: [cublas] offloading output layer to GPU\n", __func__);
+        //    ggml_cuda_transform_tensor(model.output); vram_total += ggml_nbytes(model.output);
+        //}
+
+        fprintf(stderr, "%s: [cublas] total VRAM used: %zu MB\n", __func__, vram_total / 1024 / 1024);
+    }
+#elif defined(GGML_USE_CLBLAST)
+    //From koboldcpp
+    {
+        const auto & hparams = model.hparams;
+        size_t vram_total = 0;
+        const int n_gpu = std::min(n_gpu_layers, int(hparams.n_layer));
+        fprintf(stderr, "%s: [opencl] offloading %d layers to GPU\n", __func__, n_gpu);
+        for (int i = 0; i < n_gpu; ++i) {
+            const auto & layer = model.layers[i];
+            layer.c_attn_attn_w->backend = GGML_BACKEND_GPU;
+            layer.c_attn_proj_w->backend = GGML_BACKEND_GPU;
+            layer.c_mlp_fc_w->backend = GGML_BACKEND_GPU;
+            layer.c_mlp_proj_w->backend = GGML_BACKEND_GPU;
+            ggml_cl_transform_tensor(layer.c_attn_attn_w->data,layer.c_attn_attn_w); vram_total += ggml_nbytes(layer.c_attn_attn_w);
+            ggml_cl_transform_tensor(layer.c_attn_proj_w->data,layer.c_attn_proj_w); vram_total += ggml_nbytes(layer.c_attn_proj_w);
+            ggml_cl_transform_tensor(layer.c_mlp_fc_w->data,layer.c_mlp_fc_w); vram_total += ggml_nbytes(layer.c_mlp_fc_w);
+            ggml_cl_transform_tensor(layer.c_mlp_proj_w->data,layer.c_mlp_proj_w); vram_total += ggml_nbytes(layer.c_mlp_proj_w);
+        }
+        fprintf(stderr, "%s: [opencl] total VRAM used: %zu MB\n", __func__, vram_total / 1024 / 1024);
+    }
+    #endif
+
+    return true;
+}
+
+// evaluate the transformer
+//
+//   - model:     the model
+//   - n_threads: number of threads to use
+//   - n_past:    the context size so far
+//   - embd_inp:  the embeddings of the tokens in the context
+//   - embd_w:    the predicted logits for the next token
+//
+bool starcoder_eval(
+        const starcoder_model & model,
+        const int n_threads,
+        const int n_past,
+        const std::vector<gpt_vocab::id> & embd_inp,
+              std::vector<float>         & embd_w,
+              size_t                     & mem_per_token) {
+
+    const int N = int(embd_inp.size());
+
+    const auto & hparams = model.hparams;
+
+    auto & cache = model.cache;
+
+    const int n_embd  = hparams.n_embd;
+    const int n_layer = hparams.n_layer;
+    const int n_ctx   = hparams.n_ctx;
+    const int n_head  = hparams.n_head;
+    const int n_vocab = hparams.n_vocab;
+
+    // Scratch is too small for large n_batch (256)
+    //static size_t buf_size = 256u*1024*1024;
+    static size_t buf_size = 256u*1024*1024*2;
+    static void * buf = malloc(buf_size);
+
+    // use 2 scratch buffers
+    // TODO: very hacky solution - reimplement in a more elegant way
+    static size_t scratch0_size = 256u*1024*1024*2;
+    static void * scratch0 = malloc(scratch0_size);
+
+    static size_t scratch1_size = 256u*1024*1024*2;
+    static void * scratch1 = malloc(scratch1_size);
+
+    if (mem_per_token > 0 && mem_per_token*N > buf_size) {
+        const size_t buf_size_new = size_t(1.1*(mem_per_token*N)); // add 10% to account for ggml object overhead
+        printf("\n%s: reallocating buffer from %zu to %zu bytes\n", __func__, buf_size, buf_size_new);
+
+        // reallocate
+        buf_size = buf_size_new;
+        buf = realloc(buf, buf_size);
+        if (buf == nullptr) {
+            fprintf(stderr, "%s: failed to allocate %zu bytes\n", __func__, buf_size);
+            return false;
+        }
+    }
+
+    struct ggml_init_params params = {
+        /*.mem_size   =*/ buf_size,
+        /*.mem_buffer =*/ buf,
+        /*.no_alloc   =*/ false,
+    };
+
+    struct ggml_context * ctx0 = ggml_init(params);
+    struct ggml_cgraph gf = {};
+
+    struct ggml_tensor * embd = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
+
+
+    memcpy(embd->data, embd_inp.data(), N*ggml_element_size(embd));
+
+    struct ggml_tensor * position = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
+    for (int i = 0; i < N; ++i) {
+        ((int32_t *) position->data)[i] = n_past + i;
+    }
+
+    // wte + wpe
+    struct ggml_tensor * inpL =
+        ggml_add(ctx0,
+                ggml_get_rows(ctx0, model.wte, embd),
+                ggml_get_rows(ctx0, model.wpe, position));
+
+    for (int il = 0; il < n_layer; ++il) {
+        struct ggml_tensor * cur;
+
+        ggml_set_scratch(ctx0, { 0, scratch0_size, scratch0, });
+
+        // norm
+        {
+            // [ 768, N]
+            cur = ggml_norm(ctx0, inpL, hparams.eps);
+
+            // cur = ln_1_g*cur + ln_1_b
+            // [ 768, N]
+            cur = ggml_add(ctx0,
+                    ggml_mul(ctx0,
+                        ggml_repeat(ctx0, model.layers[il].ln_1_g, cur),
+                        cur),
+                    ggml_repeat(ctx0, model.layers[il].ln_1_b, cur));
+        }
+
+        // attn
+        // [2304, 768] - model.layers[il].c_attn_attn_w
+        // [2304,   1] - model.layers[il].c_attn_attn_b
+        // [ 768,   N] - cur (in)
+        // [2304,   N] - cur (out)
+        //
+        // cur = attn_w*cur + attn_b
+        // [2304, N]
+        {
+            cur = ggml_mul_mat(ctx0,
+                    model.layers[il].c_attn_attn_w,
+                    cur);
+
+            cur = ggml_add(ctx0,
+                    ggml_repeat(ctx0, model.layers[il].c_attn_attn_b, cur),
+                    cur);
+        }
+
+        // self-attention
+        {
+            struct ggml_tensor * Qcur = ggml_view_2d(ctx0, cur, n_embd, N, cur->nb[1], 0*sizeof(float)*n_embd);
+            struct ggml_tensor * Kcur = ggml_view_2d(ctx0, cur, n_embd, N, cur->nb[1], 1*sizeof(float)*n_embd);
+            struct ggml_tensor * Vcur = ggml_view_2d(ctx0, cur, n_embd, N, cur->nb[1], 2*sizeof(float)*n_embd);
+
+            // store key and value to memory
+            if (N >= 1) {
+                struct ggml_tensor * k = ggml_view_1d(ctx0, cache.k, N*n_embd, (ggml_element_size(cache.k)*n_embd)*(il*n_ctx + n_past));
+                struct ggml_tensor * v = ggml_view_1d(ctx0, cache.v, N*n_embd, (ggml_element_size(cache.v)*n_embd)*(il*n_ctx + n_past));
+
+                ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Kcur, k));
+                ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Vcur, v));
+            }
+
+            // Q = Qcur.contiguous().view(n_embd/n_head, n_head, N).permute(0, 2, 1, 3)
+            // [64, N, 12]
+            struct ggml_tensor * Q =
+                ggml_permute(ctx0,
+                        ggml_cpy(ctx0,
+                            Qcur,
+                            ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_embd/n_head, n_head, N)),
+                        0, 2, 1, 3);
+
+            // K = Kmem.view(n_embd/n_head, n_head, n_past + N).permute(0, 2, 1, 3)
+            // [64, n_past + N, 12]
+            struct ggml_tensor * K =
+                ggml_permute(ctx0,
+                        ggml_reshape_3d(ctx0,
+                            ggml_view_1d(ctx0, cache.k, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(cache.k)*n_embd),
+                            n_embd/n_head, n_head, n_past + N),
+                        0, 2, 1, 3); //TODO: need to be tiled
+
+            // GG: flash attention
+            //struct ggml_tensor * V =
+            //    ggml_cpy(ctx0,
+            //            ggml_permute(ctx0,
+            //                ggml_reshape_3d(ctx0,
+            //                    ggml_view_1d(ctx0, model.memory_v, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(model.memory_v)*n_embd),
+            //                    n_embd/n_head, n_head, n_past + N),
+            //                1, 2, 0, 3),
+            //            ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_past + N, n_embd/n_head, n_head));
+
+            //struct ggml_tensor * KQV = ggml_flash_attn(ctx0, Q, K, V, true);
+
+            // K * Q
+            // [n_past + N, N, 12]
+            struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q); //TODO: check if it broadcasts
+
+            // KQ_scaled = KQ / sqrt(n_embd/n_head)
+            // [n_past + N, N, 12]
+            struct ggml_tensor * KQ_scaled =
+                ggml_scale_inplace(ctx0,
+                        KQ,
+                        ggml_new_f32(ctx0, 1.0f/sqrt(float(n_embd)/n_head))
+                        );
+
+            // KQ_masked = mask_past(KQ_scaled)
+            // [n_past + N, N, 12]
+            struct ggml_tensor * KQ_masked = ggml_diag_mask_inf_inplace(ctx0, KQ_scaled, n_past);
+
+            // KQ = soft_max(KQ_masked)
+            // [n_past + N, N, 12]
+            struct ggml_tensor * KQ_soft_max = ggml_soft_max_inplace(ctx0, KQ_masked);
+
+            // V_trans = Vmem.view(n_embd/n_head, n_head, n_past + N).permute(1, 2, 0, 3).contiguous()
+            // [n_past + N, 64, 12]
+            struct ggml_tensor * V_trans =
+                ggml_cpy(ctx0,
+                        ggml_permute(ctx0,
+                            ggml_reshape_3d(ctx0,
+                                ggml_view_1d(ctx0, cache.v, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(cache.v)*n_embd),
+                                n_embd/n_head, n_head, n_past + N),
+                            1, 2, 0, 3),
+                        ggml_new_tensor_3d(ctx0, cache.v->type, n_past + N, n_embd/n_head, n_head));
+
+            // KQV = transpose(V) * KQ_soft_max
+            // [64, N, 12]
+            struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V_trans, KQ_soft_max);
+
+            // KQV_merged = KQV.permute(0, 2, 1, 3)
+            // [64, 12, N]
+            struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
+
+            // cur = KQV_merged.contiguous().view(n_embd, N)
+            // [768, N]
+            cur = ggml_cpy(ctx0,
+                    KQV_merged,
+                    ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, N));
+        }
+
+        // projection
+        // [ 768, 768] - model.layers[il].c_attn_proj_w
+        // [ 768,   1] - model.layers[il].c_attn_proj_b
+        // [ 768,   N] - cur (in)
+        // [ 768,   N] - cur (out)
+        //
+        // cur = proj_w*cur + proj_b
+        // [768, N]
+        {
+            cur = ggml_mul_mat(ctx0,
+                    model.layers[il].c_attn_proj_w,
+                    cur);
+
+            cur = ggml_add(ctx0,
+                    ggml_repeat(ctx0, model.layers[il].c_attn_proj_b, cur),
+                    cur);
+        }
+
+        // add the input
+        cur = ggml_add(ctx0, cur, inpL);
+
+        struct ggml_tensor * inpFF = cur;
+
+        ggml_set_scratch(ctx0, { 0, scratch1_size, scratch1, });
+
+        // feed-forward network
+        {
+            // norm
+            {
+                cur = ggml_norm(ctx0, inpFF, hparams.eps);
+
+                // cur = ln_2_g*cur + ln_2_b
+                // [ 768, N]
+                cur = ggml_add(ctx0,
+                        ggml_mul(ctx0,
+                            ggml_repeat(ctx0, model.layers[il].ln_2_g, cur),
+                            cur),
+                        ggml_repeat(ctx0, model.layers[il].ln_2_b, cur));
+            }
+
+            // fully connected
+            // [3072, 768] - model.layers[il].c_mlp_fc_w
+            // [3072,   1] - model.layers[il].c_mlp_fc_b
+            // [ 768,   N] - cur (in)
+            // [3072,   N] - cur (out)
+            //
+            // cur = fc_w*cur + fc_b
+            // [3072, N]
+            cur = ggml_mul_mat(ctx0,
+                    model.layers[il].c_mlp_fc_w,
+                    cur);
+
+            cur = ggml_add(ctx0,
+                    ggml_repeat(ctx0, model.layers[il].c_mlp_fc_b, cur),
+                    cur);
+
+            // GELU activation
+            // [3072, N]
+            cur = ggml_gelu(ctx0, cur);
+
+            // projection
+            // [ 768, 3072] - model.layers[il].c_mlp_proj_w
+            // [ 768,    1] - model.layers[il].c_mlp_proj_b
+            // [3072,    N] - cur (in)
+            // [ 768,    N] - cur (out)
+            //
+            // cur = proj_w*cur + proj_b
+            // [768, N]
+            cur = ggml_mul_mat(ctx0,
+                    model.layers[il].c_mlp_proj_w,
+                    cur);
+
+            cur = ggml_add(ctx0,
+                    ggml_repeat(ctx0, model.layers[il].c_mlp_proj_b, cur),
+                    cur);
+        }
+
+        // input for next layer
+        inpL = ggml_add(ctx0, cur, inpFF);
+    }
+
+    ggml_set_scratch(ctx0, { 0, scratch0_size, scratch0, });
+
+    // norm
+    {
+        // [ 768, N]
+        inpL = ggml_norm(ctx0, inpL, hparams.eps);
+
+        // inpL = ln_f_g*inpL + ln_f_b
+        // [ 768, N]
+        inpL = ggml_add(ctx0,
+                ggml_mul(ctx0,
+                    ggml_repeat(ctx0, model.ln_f_g, inpL),
+                    inpL),
+                ggml_repeat(ctx0, model.ln_f_b, inpL));
+    }
+
+    ggml_set_scratch(ctx0, { 0, 0, nullptr, });
+
+    // inpL = WTE * inpL
+    // [ 768, 50257] - model.lm_head
+    // [ 768, N]     - inpL
+    inpL = ggml_mul_mat(ctx0, model.lm_head, inpL);
+
+    // logits -> probs
+    //inpL = ggml_soft_max_inplace(ctx0, inpL);
+
+    // run the computation
+    ggml_build_forward_expand(&gf, inpL);
+    ggml_graph_compute_with_ctx(ctx0, &gf, n_threads);
+
+    //if (n_past%100 == 0) {
+    //    ggml_graph_print   (&gf);
+    //    ggml_graph_dump_dot(&gf, NULL, "gpt-2.dot");
+    //}
+
+    //embd_w.resize(n_vocab*N);
+    //memcpy(embd_w.data(), ggml_get_data(inpL), sizeof(float)*n_vocab*N);
+
+    // return result just for the last token
+    embd_w.resize(n_vocab);
+    memcpy(embd_w.data(), (float *) ggml_get_data(inpL) + (n_vocab*(N-1)), sizeof(float)*n_vocab);
+
+    if (mem_per_token == 0) {
+        mem_per_token = ggml_used_mem(ctx0)/N;
+    }
+    //printf("used_mem = %zu MB\n", ggml_used_mem(ctx0)/(1024*1024));
+
+    ggml_free(ctx0);
+
+    return true;
+}
+
+
+int main(int argc, char ** argv) {
+    ggml_time_init();
+
+    const int64_t t_main_start_us = ggml_time_us();
+
+    gpt_params params;
+    params.model = "models/gpt-2-117M/ggml-model.bin";
+
+    if (gpt_params_parse(argc, argv, params) == false) {
+        return 1;
+    }
+
+    if (params.seed < 0) {
+        params.seed = int(time(NULL));
+    }
+
+    printf("%s: seed = %d\n", __func__, params.seed);
+
+    std::mt19937 rng(params.seed);
+    if (params.prompt.empty()) {
+        params.prompt = gpt_random_prompt(rng);
+    }
+
+    int64_t t_load_us = 0;
+
+    gpt_vocab vocab;
+    starcoder_model model;
+
+    // load the model
+    {
+        const int64_t t_start_us = ggml_time_us();
+
+        if (!starcoder_model_load(params.model, model, vocab, params.n_gpu_layers)) {
+            fprintf(stderr, "%s: failed to load model from '%s'\n", __func__, params.model.c_str());
+            return 1;
+        }
+
+        t_load_us = ggml_time_us() - t_start_us;
+
+        test_gpt_tokenizer(vocab, params.token_test);
+    }
+
+    int n_past = 0;
+
+    int64_t t_sample_us  = 0;
+    int64_t t_predict_us = 0;
+
+    std::vector<float> logits;
+
+    // tokenize the prompt
+    std::vector<gpt_vocab::id> embd_inp = ::gpt_tokenize(vocab, params.prompt);
+
+    params.n_predict = std::min(params.n_predict, model.hparams.n_ctx - (int) embd_inp.size());
+
+    printf("%s: prompt: '%s'\n", __func__, params.prompt.c_str());
+    printf("%s: number of tokens in prompt = %zu\n", __func__, embd_inp.size());
+    for (size_t i = 0; i < embd_inp.size(); i++) {
+        printf("%s: token[%zu] = %6d, %s\n", __func__, i, embd_inp[i], vocab.id_to_token.at(embd_inp[i]).c_str());
+    }
+    printf("\n\n");
+
+    // Handle StarChat "<|end|>" token.
+    gpt_vocab::id starchat_end_token = -1;
+    {
+        const auto it = vocab.token_to_id.find("<|end|>");
+        if (it != vocab.token_to_id.end()) {
+            starchat_end_token = it->second;
+        }
+    }
+
+    // submit the input prompt token-by-token
+    // this reduces the memory usage during inference, at the cost of a bit of speed at the beginning
+    std::vector<gpt_vocab::id> embd;
+
+    // determine the required inference memory per token:
+    size_t mem_per_token = 0;
+    printf("Calling starcoder_eval\n");
+    starcoder_eval(model, params.n_threads, 0, { 0, 1, 2, 3 }, logits, mem_per_token);
+
+    for (size_t i = embd.size(); i < embd_inp.size() + params.n_predict; i++) {
+        // predict
+        if (embd.size() > 0) {
+            const int64_t t_start_us = ggml_time_us();
+
+            if (!starcoder_eval(model, params.n_threads, n_past, embd, logits, mem_per_token)) {
+                printf("Failed to predict\n");
+                return 1;
+            }
+
+            // Should input processing count towards t_predict?
+            if (i > embd_inp.size()) {
+                t_predict_us += ggml_time_us() - t_start_us;
+            }
+        }
+
+        n_past += int(embd.size());
+        embd.clear();
+
+        if (i >= embd_inp.size()) {
+            // sample next token
+            const int   top_k = params.top_k;
+            const float top_p = params.top_p;
+            const float temp  = params.temp;
+
+            const int n_vocab = model.hparams.n_vocab;
+
+            gpt_vocab::id id = 0;
+
+            {
+                const int64_t t_start_sample_us = ggml_time_us();
+
+                id = gpt_sample_top_k_top_p(vocab, logits.data() + (logits.size() - n_vocab), top_k, top_p, temp, rng);
+
+                t_sample_us += ggml_time_us() - t_start_sample_us;
+            }
+
+            // add it to the context
+            embd.push_back(id);
+        } else {
+            // if here, it means we are still processing the input prompt
+            for (size_t k = i; k < embd_inp.size(); k++) {
+                embd.push_back(embd_inp[k]);
+                if (int32_t(embd.size()) >= params.n_batch) {
+                    break;
+                }
+            }
+            i += int(embd.size()) - 1;
+        }
+
+        // display text
+        for (auto id : embd) {
+            printf("%s", vocab.id_to_token[id].c_str());
+        }
+        fflush(stdout);
+
+        // check if model is santacoder
+        if (model.hparams.n_layer <= 30 && embd.back() == 49152) {
+            break;
+        }
+        // check if model is starcoder
+        else if (embd.back() == 0) { //TODO: this is only for starcoder
+            break;
+        }
+        // Handle StarChat "<|end|>" token.
+        else if (embd.back() == starchat_end_token) {
+            //break;
+        }
+    }
+
+    // report timing
+    {
+        const int64_t t_main_end_us = ggml_time_us();
+
+        printf("\n\n");
+        printf("%s: mem per token = %8zu bytes\n", __func__, mem_per_token);
+        printf("%s:     load time = %8.2f ms\n", __func__, t_load_us/1000.0f);
+        printf("%s:   sample time = %8.2f ms\n", __func__, t_sample_us/1000.0f);
+        //Shouldnt the input prompt be subracted?
+        printf("%s:  predict time = %8.2f ms / %.2f ms per token\n", __func__, t_predict_us/1000.0f, t_predict_us/1000.0f/(n_past - embd_inp.size()));
+        //printf("%s:  predict time = %8.2f ms / %.2f ms per token\n", __func__, t_predict_us/1000.0f, t_predict_us/1000.0f/n_past);
+
+        printf("%s:    total time = %8.2f ms\n", __func__, (t_main_end_us - t_main_start_us)/1000.0f);
+    }
+
+    ggml_free(model.ctx);
+
+    if (model.mm_addr) {
+	    munmap_file(model.mm_addr, model.mm_length);
+    }
+
+    return 0;
+}

ggml/examples/stb_image.h

@@ -0,0 +1,7987 @@
+/* stb_image - v2.28 - public domain image loader - http://nothings.org/stb
+                                  no warranty implied; use at your own risk
+
+   Do this:
+      #define STB_IMAGE_IMPLEMENTATION
+   before you include this file in *one* C or C++ file to create the implementation.
+
+   // i.e. it should look like this:
+   #include ...
+   #include ...
+   #include ...
+   #define STB_IMAGE_IMPLEMENTATION
+   #include "stb_image.h"
+
+   You can #define STBI_ASSERT(x) before the #include to avoid using assert.h.
+   And #define STBI_MALLOC, STBI_REALLOC, and STBI_FREE to avoid using malloc,realloc,free
+
+
+   QUICK NOTES:
+      Primarily of interest to game developers and other people who can
+          avoid problematic images and only need the trivial interface
+
+      JPEG baseline & progressive (12 bpc/arithmetic not supported, same as stock IJG lib)
+      PNG 1/2/4/8/16-bit-per-channel
+
+      TGA (not sure what subset, if a subset)
+      BMP non-1bpp, non-RLE
+      PSD (composited view only, no extra channels, 8/16 bit-per-channel)
+
+      GIF (*comp always reports as 4-channel)
+      HDR (radiance rgbE format)
+      PIC (Softimage PIC)
+      PNM (PPM and PGM binary only)
+
+      Animated GIF still needs a proper API, but here's one way to do it:
+          http://gist.github.com/urraka/685d9a6340b26b830d49
+
+      - decode from memory or through FILE (define STBI_NO_STDIO to remove code)
+      - decode from arbitrary I/O callbacks
+      - SIMD acceleration on x86/x64 (SSE2) and ARM (NEON)
+
+   Full documentation under "DOCUMENTATION" below.
+
+
+LICENSE
+
+  See end of file for license information.
+
+RECENT REVISION HISTORY:
+
+      2.28  (2023-01-29) many error fixes, security errors, just tons of stuff
+      2.27  (2021-07-11) document stbi_info better, 16-bit PNM support, bug fixes
+      2.26  (2020-07-13) many minor fixes
+      2.25  (2020-02-02) fix warnings
+      2.24  (2020-02-02) fix warnings; thread-local failure_reason and flip_vertically
+      2.23  (2019-08-11) fix clang static analysis warning
+      2.22  (2019-03-04) gif fixes, fix warnings
+      2.21  (2019-02-25) fix typo in comment
+      2.20  (2019-02-07) support utf8 filenames in Windows; fix warnings and platform ifdefs
+      2.19  (2018-02-11) fix warning
+      2.18  (2018-01-30) fix warnings
+      2.17  (2018-01-29) bugfix, 1-bit BMP, 16-bitness query, fix warnings
+      2.16  (2017-07-23) all functions have 16-bit variants; optimizations; bugfixes
+      2.15  (2017-03-18) fix png-1,2,4; all Imagenet JPGs; no runtime SSE detection on GCC
+      2.14  (2017-03-03) remove deprecated STBI_JPEG_OLD; fixes for Imagenet JPGs
+      2.13  (2016-12-04) experimental 16-bit API, only for PNG so far; fixes
+      2.12  (2016-04-02) fix typo in 2.11 PSD fix that caused crashes
+      2.11  (2016-04-02) 16-bit PNGS; enable SSE2 in non-gcc x64
+                         RGB-format JPEG; remove white matting in PSD;
+                         allocate large structures on the stack;
+                         correct channel count for PNG & BMP
+      2.10  (2016-01-22) avoid warning introduced in 2.09
+      2.09  (2016-01-16) 16-bit TGA; comments in PNM files; STBI_REALLOC_SIZED
+
+   See end of file for full revision history.
+
+
+ ============================    Contributors    =========================
+
+ Image formats                          Extensions, features
+    Sean Barrett (jpeg, png, bmp)          Jetro Lauha (stbi_info)
+    Nicolas Schulz (hdr, psd)              Martin "SpartanJ" Golini (stbi_info)
+    Jonathan Dummer (tga)                  James "moose2000" Brown (iPhone PNG)
+    Jean-Marc Lienher (gif)                Ben "Disch" Wenger (io callbacks)
+    Tom Seddon (pic)                       Omar Cornut (1/2/4-bit PNG)
+    Thatcher Ulrich (psd)                  Nicolas Guillemot (vertical flip)
+    Ken Miller (pgm, ppm)                  Richard Mitton (16-bit PSD)
+    github:urraka (animated gif)           Junggon Kim (PNM comments)
+    Christopher Forseth (animated gif)     Daniel Gibson (16-bit TGA)
+                                           socks-the-fox (16-bit PNG)
+                                           Jeremy Sawicki (handle all ImageNet JPGs)
+ Optimizations & bugfixes                  Mikhail Morozov (1-bit BMP)
+    Fabian "ryg" Giesen                    Anael Seghezzi (is-16-bit query)
+    Arseny Kapoulkine                      Simon Breuss (16-bit PNM)
+    John-Mark Allen
+    Carmelo J Fdez-Aguera
+
+ Bug & warning fixes
+    Marc LeBlanc            David Woo          Guillaume George     Martins Mozeiko
+    Christpher Lloyd        Jerry Jansson      Joseph Thomson       Blazej Dariusz Roszkowski
+    Phil Jordan                                Dave Moore           Roy Eltham
+    Hayaki Saito            Nathan Reed        Won Chun
+    Luke Graham             Johan Duparc       Nick Verigakis       the Horde3D community
+    Thomas Ruf              Ronny Chevalier                         github:rlyeh
+    Janez Zemva             John Bartholomew   Michal Cichon        github:romigrou
+    Jonathan Blow           Ken Hamada         Tero Hanninen        github:svdijk
+    Eugene Golushkov        Laurent Gomila     Cort Stratton        github:snagar
+    Aruelien Pocheville     Sergio Gonzalez    Thibault Reuille     github:Zelex
+    Cass Everitt            Ryamond Barbiero                        github:grim210
+    Paul Du Bois            Engin Manap        Aldo Culquicondor    github:sammyhw
+    Philipp Wiesemann       Dale Weiler        Oriol Ferrer Mesia   github:phprus
+    Josh Tobin              Neil Bickford      Matthew Gregan       github:poppolopoppo
+    Julian Raschke          Gregory Mullen     Christian Floisand   github:darealshinji
+    Baldur Karlsson         Kevin Schmidt      JR Smith             github:Michaelangel007
+                            Brad Weinberger    Matvey Cherevko      github:mosra
+    Luca Sas                Alexander Veselov  Zack Middleton       [reserved]
+    Ryan C. Gordon          [reserved]                              [reserved]
+                     DO NOT ADD YOUR NAME HERE
+
+                     Jacko Dirks
+
+  To add your name to the credits, pick a random blank space in the middle and fill it.
+  80% of merge conflicts on stb PRs are due to people adding their name at the end
+  of the credits.
+*/
+
+#ifndef STBI_INCLUDE_STB_IMAGE_H
+#define STBI_INCLUDE_STB_IMAGE_H
+
+// DOCUMENTATION
+//
+// Limitations:
+//    - no 12-bit-per-channel JPEG
+//    - no JPEGs with arithmetic coding
+//    - GIF always returns *comp=4
+//
+// Basic usage (see HDR discussion below for HDR usage):
+//    int x,y,n;
+//    unsigned char *data = stbi_load(filename, &x, &y, &n, 0);
+//    // ... process data if not NULL ...
+//    // ... x = width, y = height, n = # 8-bit components per pixel ...
+//    // ... replace '0' with '1'..'4' to force that many components per pixel
+//    // ... but 'n' will always be the number that it would have been if you said 0
+//    stbi_image_free(data);
+//
+// Standard parameters:
+//    int *x                 -- outputs image width in pixels
+//    int *y                 -- outputs image height in pixels
+//    int *channels_in_file  -- outputs # of image components in image file
+//    int desired_channels   -- if non-zero, # of image components requested in result
+//
+// The return value from an image loader is an 'unsigned char *' which points
+// to the pixel data, or NULL on an allocation failure or if the image is
+// corrupt or invalid. The pixel data consists of *y scanlines of *x pixels,
+// with each pixel consisting of N interleaved 8-bit components; the first
+// pixel pointed to is top-left-most in the image. There is no padding between
+// image scanlines or between pixels, regardless of format. The number of
+// components N is 'desired_channels' if desired_channels is non-zero, or
+// *channels_in_file otherwise. If desired_channels is non-zero,
+// *channels_in_file has the number of components that _would_ have been
+// output otherwise. E.g. if you set desired_channels to 4, you will always
+// get RGBA output, but you can check *channels_in_file to see if it's trivially
+// opaque because e.g. there were only 3 channels in the source image.
+//
+// An output image with N components has the following components interleaved
+// in this order in each pixel:
+//
+//     N=#comp     components
+//       1           grey
+//       2           grey, alpha
+//       3           red, green, blue
+//       4           red, green, blue, alpha
+//
+// If image loading fails for any reason, the return value will be NULL,
+// and *x, *y, *channels_in_file will be unchanged. The function
+// stbi_failure_reason() can be queried for an extremely brief, end-user
+// unfriendly explanation of why the load failed. Define STBI_NO_FAILURE_STRINGS
+// to avoid compiling these strings at all, and STBI_FAILURE_USERMSG to get slightly
+// more user-friendly ones.
+//
+// Paletted PNG, BMP, GIF, and PIC images are automatically depalettized.
+//
+// To query the width, height and component count of an image without having to
+// decode the full file, you can use the stbi_info family of functions:
+//
+//   int x,y,n,ok;
+//   ok = stbi_info(filename, &x, &y, &n);
+//   // returns ok=1 and sets x, y, n if image is a supported format,
+//   // 0 otherwise.
+//
+// Note that stb_image pervasively uses ints in its public API for sizes,
+// including sizes of memory buffers. This is now part of the API and thus
+// hard to change without causing breakage. As a result, the various image
+// loaders all have certain limits on image size; these differ somewhat
+// by format but generally boil down to either just under 2GB or just under
+// 1GB. When the decoded image would be larger than this, stb_image decoding
+// will fail.
+//
+// Additionally, stb_image will reject image files that have any of their
+// dimensions set to a larger value than the configurable STBI_MAX_DIMENSIONS,
+// which defaults to 2**24 = 16777216 pixels. Due to the above memory limit,
+// the only way to have an image with such dimensions load correctly
+// is for it to have a rather extreme aspect ratio. Either way, the
+// assumption here is that such larger images are likely to be malformed
+// or malicious. If you do need to load an image with individual dimensions
+// larger than that, and it still fits in the overall size limit, you can
+// #define STBI_MAX_DIMENSIONS on your own to be something larger.
+//
+// ===========================================================================
+//
+// UNICODE:
+//
+//   If compiling for Windows and you wish to use Unicode filenames, compile
+//   with
+//       #define STBI_WINDOWS_UTF8
+//   and pass utf8-encoded filenames. Call stbi_convert_wchar_to_utf8 to convert
+//   Windows wchar_t filenames to utf8.
+//
+// ===========================================================================
+//
+// Philosophy
+//
+// stb libraries are designed with the following priorities:
+//
+//    1. easy to use
+//    2. easy to maintain
+//    3. good performance
+//
+// Sometimes I let "good performance" creep up in priority over "easy to maintain",
+// and for best performance I may provide less-easy-to-use APIs that give higher
+// performance, in addition to the easy-to-use ones. Nevertheless, it's important
+// to keep in mind that from the standpoint of you, a client of this library,
+// all you care about is #1 and #3, and stb libraries DO NOT emphasize #3 above all.
+//
+// Some secondary priorities arise directly from the first two, some of which
+// provide more explicit reasons why performance can't be emphasized.
+//
+//    - Portable ("ease of use")
+//    - Small source code footprint ("easy to maintain")
+//    - No dependencies ("ease of use")
+//
+// ===========================================================================
+//
+// I/O callbacks
+//
+// I/O callbacks allow you to read from arbitrary sources, like packaged
+// files or some other source. Data read from callbacks are processed
+// through a small internal buffer (currently 128 bytes) to try to reduce
+// overhead.
+//
+// The three functions you must define are "read" (reads some bytes of data),
+// "skip" (skips some bytes of data), "eof" (reports if the stream is at the end).
+//
+// ===========================================================================
+//
+// SIMD support
+//
+// The JPEG decoder will try to automatically use SIMD kernels on x86 when
+// supported by the compiler. For ARM Neon support, you must explicitly
+// request it.
+//
+// (The old do-it-yourself SIMD API is no longer supported in the current
+// code.)
+//
+// On x86, SSE2 will automatically be used when available based on a run-time
+// test; if not, the generic C versions are used as a fall-back. On ARM targets,
+// the typical path is to have separate builds for NEON and non-NEON devices
+// (at least this is true for iOS and Android). Therefore, the NEON support is
+// toggled by a build flag: define STBI_NEON to get NEON loops.
+//
+// If for some reason you do not want to use any of SIMD code, or if
+// you have issues compiling it, you can disable it entirely by
+// defining STBI_NO_SIMD.
+//
+// ===========================================================================
+//
+// HDR image support   (disable by defining STBI_NO_HDR)
+//
+// stb_image supports loading HDR images in general, and currently the Radiance
+// .HDR file format specifically. You can still load any file through the existing
+// interface; if you attempt to load an HDR file, it will be automatically remapped
+// to LDR, assuming gamma 2.2 and an arbitrary scale factor defaulting to 1;
+// both of these constants can be reconfigured through this interface:
+//
+//     stbi_hdr_to_ldr_gamma(2.2f);
+//     stbi_hdr_to_ldr_scale(1.0f);
+//
+// (note, do not use _inverse_ constants; stbi_image will invert them
+// appropriately).
+//
+// Additionally, there is a new, parallel interface for loading files as
+// (linear) floats to preserve the full dynamic range:
+//
+//    float *data = stbi_loadf(filename, &x, &y, &n, 0);
+//
+// If you load LDR images through this interface, those images will
+// be promoted to floating point values, run through the inverse of
+// constants corresponding to the above:
+//
+//     stbi_ldr_to_hdr_scale(1.0f);
+//     stbi_ldr_to_hdr_gamma(2.2f);
+//
+// Finally, given a filename (or an open file or memory block--see header
+// file for details) containing image data, you can query for the "most
+// appropriate" interface to use (that is, whether the image is HDR or
+// not), using:
+//
+//     stbi_is_hdr(char *filename);
+//
+// ===========================================================================
+//
+// iPhone PNG support:
+//
+// We optionally support converting iPhone-formatted PNGs (which store
+// premultiplied BGRA) back to RGB, even though they're internally encoded
+// differently. To enable this conversion, call
+// stbi_convert_iphone_png_to_rgb(1).
+//
+// Call stbi_set_unpremultiply_on_load(1) as well to force a divide per
+// pixel to remove any premultiplied alpha *only* if the image file explicitly
+// says there's premultiplied data (currently only happens in iPhone images,
+// and only if iPhone convert-to-rgb processing is on).
+//
+// ===========================================================================
+//
+// ADDITIONAL CONFIGURATION
+//
+//  - You can suppress implementation of any of the decoders to reduce
+//    your code footprint by #defining one or more of the following
+//    symbols before creating the implementation.
+//
+//        STBI_NO_JPEG
+//        STBI_NO_PNG
+//        STBI_NO_BMP
+//        STBI_NO_PSD
+//        STBI_NO_TGA
+//        STBI_NO_GIF
+//        STBI_NO_HDR
+//        STBI_NO_PIC
+//        STBI_NO_PNM   (.ppm and .pgm)
+//
+//  - You can request *only* certain decoders and suppress all other ones
+//    (this will be more forward-compatible, as addition of new decoders
+//    doesn't require you to disable them explicitly):
+//
+//        STBI_ONLY_JPEG
+//        STBI_ONLY_PNG
+//        STBI_ONLY_BMP
+//        STBI_ONLY_PSD
+//        STBI_ONLY_TGA
+//        STBI_ONLY_GIF
+//        STBI_ONLY_HDR
+//        STBI_ONLY_PIC
+//        STBI_ONLY_PNM   (.ppm and .pgm)
+//
+//   - If you use STBI_NO_PNG (or _ONLY_ without PNG), and you still
+//     want the zlib decoder to be available, #define STBI_SUPPORT_ZLIB
+//
+//  - If you define STBI_MAX_DIMENSIONS, stb_image will reject images greater
+//    than that size (in either width or height) without further processing.
+//    This is to let programs in the wild set an upper bound to prevent
+//    denial-of-service attacks on untrusted data, as one could generate a
+//    valid image of gigantic dimensions and force stb_image to allocate a
+//    huge block of memory and spend disproportionate time decoding it. By
+//    default this is set to (1 << 24), which is 16777216, but that's still
+//    very big.
+
+#ifndef STBI_NO_STDIO
+#include <stdio.h>
+#endif // STBI_NO_STDIO
+
+#define STBI_VERSION 1
+
+enum
+{
+   STBI_default = 0, // only used for desired_channels
+
+   STBI_grey       = 1,
+   STBI_grey_alpha = 2,
+   STBI_rgb        = 3,
+   STBI_rgb_alpha  = 4
+};
+
+#include <stdlib.h>
+typedef unsigned char stbi_uc;
+typedef unsigned short stbi_us;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#ifndef STBIDEF
+#ifdef STB_IMAGE_STATIC
+#define STBIDEF static
+#else
+#define STBIDEF extern
+#endif
+#endif
+
+//////////////////////////////////////////////////////////////////////////////
+//
+// PRIMARY API - works on images of any type
+//
+
+//
+// load image by filename, open file, or memory buffer
+//
+
+typedef struct
+{
+   int      (*read)  (void *user,char *data,int size);   // fill 'data' with 'size' bytes.  return number of bytes actually read
+   void     (*skip)  (void *user,int n);                 // skip the next 'n' bytes, or 'unget' the last -n bytes if negative
+   int      (*eof)   (void *user);                       // returns nonzero if we are at end of file/data
+} stbi_io_callbacks;
+
+////////////////////////////////////
+//
+// 8-bits-per-channel interface
+//
+
+STBIDEF stbi_uc *stbi_load_from_memory   (stbi_uc           const *buffer, int len   , int *x, int *y, int *channels_in_file, int desired_channels);
+STBIDEF stbi_uc *stbi_load_from_callbacks(stbi_io_callbacks const *clbk  , void *user, int *x, int *y, int *channels_in_file, int desired_channels);
+
+#ifndef STBI_NO_STDIO
+STBIDEF stbi_uc *stbi_load            (char const *filename, int *x, int *y, int *channels_in_file, int desired_channels);
+STBIDEF stbi_uc *stbi_load_from_file  (FILE *f, int *x, int *y, int *channels_in_file, int desired_channels);
+// for stbi_load_from_file, file pointer is left pointing immediately after image
+#endif
+
+#ifndef STBI_NO_GIF
+STBIDEF stbi_uc *stbi_load_gif_from_memory(stbi_uc const *buffer, int len, int **delays, int *x, int *y, int *z, int *comp, int req_comp);
+#endif
+
+#ifdef STBI_WINDOWS_UTF8
+STBIDEF int stbi_convert_wchar_to_utf8(char *buffer, size_t bufferlen, const wchar_t* input);
+#endif
+
+////////////////////////////////////
+//
+// 16-bits-per-channel interface
+//
+
+STBIDEF stbi_us *stbi_load_16_from_memory   (stbi_uc const *buffer, int len, int *x, int *y, int *channels_in_file, int desired_channels);
+STBIDEF stbi_us *stbi_load_16_from_callbacks(stbi_io_callbacks const *clbk, void *user, int *x, int *y, int *channels_in_file, int desired_channels);
+
+#ifndef STBI_NO_STDIO
+STBIDEF stbi_us *stbi_load_16          (char const *filename, int *x, int *y, int *channels_in_file, int desired_channels);
+STBIDEF stbi_us *stbi_load_from_file_16(FILE *f, int *x, int *y, int *channels_in_file, int desired_channels);
+#endif
+
+////////////////////////////////////
+//
+// float-per-channel interface
+//
+#ifndef STBI_NO_LINEAR
+   STBIDEF float *stbi_loadf_from_memory     (stbi_uc const *buffer, int len, int *x, int *y, int *channels_in_file, int desired_channels);
+   STBIDEF float *stbi_loadf_from_callbacks  (stbi_io_callbacks const *clbk, void *user, int *x, int *y,  int *channels_in_file, int desired_channels);
+
+   #ifndef STBI_NO_STDIO
+   STBIDEF float *stbi_loadf            (char const *filename, int *x, int *y, int *channels_in_file, int desired_channels);
+   STBIDEF float *stbi_loadf_from_file  (FILE *f, int *x, int *y, int *channels_in_file, int desired_channels);
+   #endif
+#endif
+
+#ifndef STBI_NO_HDR
+   STBIDEF void   stbi_hdr_to_ldr_gamma(float gamma);
+   STBIDEF void   stbi_hdr_to_ldr_scale(float scale);
+#endif // STBI_NO_HDR
+
+#ifndef STBI_NO_LINEAR
+   STBIDEF void   stbi_ldr_to_hdr_gamma(float gamma);
+   STBIDEF void   stbi_ldr_to_hdr_scale(float scale);
+#endif // STBI_NO_LINEAR
+
+// stbi_is_hdr is always defined, but always returns false if STBI_NO_HDR
+STBIDEF int    stbi_is_hdr_from_callbacks(stbi_io_callbacks const *clbk, void *user);
+STBIDEF int    stbi_is_hdr_from_memory(stbi_uc const *buffer, int len);
+#ifndef STBI_NO_STDIO
+STBIDEF int      stbi_is_hdr          (char const *filename);
+STBIDEF int      stbi_is_hdr_from_file(FILE *f);
+#endif // STBI_NO_STDIO
+
+
+// get a VERY brief reason for failure
+// on most compilers (and ALL modern mainstream compilers) this is threadsafe
+STBIDEF const char *stbi_failure_reason  (void);
+
+// free the loaded image -- this is just free()
+STBIDEF void     stbi_image_free      (void *retval_from_stbi_load);
+
+// get image dimensions & components without fully decoding
+STBIDEF int      stbi_info_from_memory(stbi_uc const *buffer, int len, int *x, int *y, int *comp);
+STBIDEF int      stbi_info_from_callbacks(stbi_io_callbacks const *clbk, void *user, int *x, int *y, int *comp);
+STBIDEF int      stbi_is_16_bit_from_memory(stbi_uc const *buffer, int len);
+STBIDEF int      stbi_is_16_bit_from_callbacks(stbi_io_callbacks const *clbk, void *user);
+
+#ifndef STBI_NO_STDIO
+STBIDEF int      stbi_info               (char const *filename,     int *x, int *y, int *comp);
+STBIDEF int      stbi_info_from_file     (FILE *f,                  int *x, int *y, int *comp);
+STBIDEF int      stbi_is_16_bit          (char const *filename);
+STBIDEF int      stbi_is_16_bit_from_file(FILE *f);
+#endif
+
+
+
+// for image formats that explicitly notate that they have premultiplied alpha,
+// we just return the colors as stored in the file. set this flag to force
+// unpremultiplication. results are undefined if the unpremultiply overflow.
+STBIDEF void stbi_set_unpremultiply_on_load(int flag_true_if_should_unpremultiply);
+
+// indicate whether we should process iphone images back to canonical format,
+// or just pass them through "as-is"
+STBIDEF void stbi_convert_iphone_png_to_rgb(int flag_true_if_should_convert);
+
+// flip the image vertically, so the first pixel in the output array is the bottom left
+STBIDEF void stbi_set_flip_vertically_on_load(int flag_true_if_should_flip);
+
+// as above, but only applies to images loaded on the thread that calls the function
+// this function is only available if your compiler supports thread-local variables;
+// calling it will fail to link if your compiler doesn't
+STBIDEF void stbi_set_unpremultiply_on_load_thread(int flag_true_if_should_unpremultiply);
+STBIDEF void stbi_convert_iphone_png_to_rgb_thread(int flag_true_if_should_convert);
+STBIDEF void stbi_set_flip_vertically_on_load_thread(int flag_true_if_should_flip);
+
+// ZLIB client - used by PNG, available for other purposes
+
+STBIDEF char *stbi_zlib_decode_malloc_guesssize(const char *buffer, int len, int initial_size, int *outlen);
+STBIDEF char *stbi_zlib_decode_malloc_guesssize_headerflag(const char *buffer, int len, int initial_size, int *outlen, int parse_header);
+STBIDEF char *stbi_zlib_decode_malloc(const char *buffer, int len, int *outlen);
+STBIDEF int   stbi_zlib_decode_buffer(char *obuffer, int olen, const char *ibuffer, int ilen);
+
+STBIDEF char *stbi_zlib_decode_noheader_malloc(const char *buffer, int len, int *outlen);
+STBIDEF int   stbi_zlib_decode_noheader_buffer(char *obuffer, int olen, const char *ibuffer, int ilen);
+
+
+#ifdef __cplusplus
+}
+#endif
+
+//
+//
+////   end header file   /////////////////////////////////////////////////////
+#endif // STBI_INCLUDE_STB_IMAGE_H
+
+#ifdef STB_IMAGE_IMPLEMENTATION
+
+#if defined(STBI_ONLY_JPEG) || defined(STBI_ONLY_PNG) || defined(STBI_ONLY_BMP) \
+  || defined(STBI_ONLY_TGA) || defined(STBI_ONLY_GIF) || defined(STBI_ONLY_PSD) \
+  || defined(STBI_ONLY_HDR) || defined(STBI_ONLY_PIC) || defined(STBI_ONLY_PNM) \
+  || defined(STBI_ONLY_ZLIB)
+   #ifndef STBI_ONLY_JPEG
+   #define STBI_NO_JPEG
+   #endif
+   #ifndef STBI_ONLY_PNG
+   #define STBI_NO_PNG
+   #endif
+   #ifndef STBI_ONLY_BMP
+   #define STBI_NO_BMP
+   #endif
+   #ifndef STBI_ONLY_PSD
+   #define STBI_NO_PSD
+   #endif
+   #ifndef STBI_ONLY_TGA
+   #define STBI_NO_TGA
+   #endif
+   #ifndef STBI_ONLY_GIF
+   #define STBI_NO_GIF
+   #endif
+   #ifndef STBI_ONLY_HDR
+   #define STBI_NO_HDR
+   #endif
+   #ifndef STBI_ONLY_PIC
+   #define STBI_NO_PIC
+   #endif
+   #ifndef STBI_ONLY_PNM
+   #define STBI_NO_PNM
+   #endif
+#endif
+
+#if defined(STBI_NO_PNG) && !defined(STBI_SUPPORT_ZLIB) && !defined(STBI_NO_ZLIB)
+#define STBI_NO_ZLIB
+#endif
+
+
+#include <stdarg.h>
+#include <stddef.h> // ptrdiff_t on osx
+#include <stdlib.h>
+#include <string.h>
+#include <limits.h>
+
+#if !defined(STBI_NO_LINEAR) || !defined(STBI_NO_HDR)
+#include <math.h>  // ldexp, pow
+#endif
+
+#ifndef STBI_NO_STDIO
+#include <stdio.h>
+#endif
+
+#ifndef STBI_ASSERT
+#include <assert.h>
+#define STBI_ASSERT(x) assert(x)
+#endif
+
+#ifdef __cplusplus
+#define STBI_EXTERN extern "C"
+#else
+#define STBI_EXTERN extern
+#endif
+
+
+#ifndef _MSC_VER
+   #ifdef __cplusplus
+   #define stbi_inline inline
+   #else
+   #define stbi_inline
+   #endif
+#else
+   #define stbi_inline __forceinline
+#endif
+
+#ifndef STBI_NO_THREAD_LOCALS
+   #if defined(__cplusplus) &&  __cplusplus >= 201103L
+      #define STBI_THREAD_LOCAL       thread_local
+   #elif defined(__GNUC__) && __GNUC__ < 5
+      #define STBI_THREAD_LOCAL       __thread
+   #elif defined(_MSC_VER)
+      #define STBI_THREAD_LOCAL       __declspec(thread)
+   #elif defined (__STDC_VERSION__) && __STDC_VERSION__ >= 201112L && !defined(__STDC_NO_THREADS__)
+      #define STBI_THREAD_LOCAL       _Thread_local
+   #endif
+
+   #ifndef STBI_THREAD_LOCAL
+      #if defined(__GNUC__)
+        #define STBI_THREAD_LOCAL       __thread
+      #endif
+   #endif
+#endif
+
+#if defined(_MSC_VER) || defined(__SYMBIAN32__)
+typedef unsigned short stbi__uint16;
+typedef   signed short stbi__int16;
+typedef unsigned int   stbi__uint32;
+typedef   signed int   stbi__int32;
+#else
+#include <stdint.h>
+typedef uint16_t stbi__uint16;
+typedef int16_t  stbi__int16;
+typedef uint32_t stbi__uint32;
+typedef int32_t  stbi__int32;
+#endif
+
+// should produce compiler error if size is wrong
+typedef unsigned char validate_uint32[sizeof(stbi__uint32)==4 ? 1 : -1];
+
+#ifdef _MSC_VER
+#define STBI_NOTUSED(v)  (void)(v)
+#else
+#define STBI_NOTUSED(v)  (void)sizeof(v)
+#endif
+
+#ifdef _MSC_VER
+#define STBI_HAS_LROTL
+#endif
+
+#ifdef STBI_HAS_LROTL
+   #define stbi_lrot(x,y)  _lrotl(x,y)
+#else
+   #define stbi_lrot(x,y)  (((x) << (y)) | ((x) >> (-(y) & 31)))
+#endif
+
+#if defined(STBI_MALLOC) && defined(STBI_FREE) && (defined(STBI_REALLOC) || defined(STBI_REALLOC_SIZED))
+// ok
+#elif !defined(STBI_MALLOC) && !defined(STBI_FREE) && !defined(STBI_REALLOC) && !defined(STBI_REALLOC_SIZED)
+// ok
+#else
+#error "Must define all or none of STBI_MALLOC, STBI_FREE, and STBI_REALLOC (or STBI_REALLOC_SIZED)."
+#endif
+
+#ifndef STBI_MALLOC
+#define STBI_MALLOC(sz)           malloc(sz)
+#define STBI_REALLOC(p,newsz)     realloc(p,newsz)
+#define STBI_FREE(p)              free(p)
+#endif
+
+#ifndef STBI_REALLOC_SIZED
+#define STBI_REALLOC_SIZED(p,oldsz,newsz) STBI_REALLOC(p,newsz)
+#endif
+
+// x86/x64 detection
+#if defined(__x86_64__) || defined(_M_X64)
+#define STBI__X64_TARGET
+#elif defined(__i386) || defined(_M_IX86)
+#define STBI__X86_TARGET
+#endif
+
+#if defined(__GNUC__) && defined(STBI__X86_TARGET) && !defined(__SSE2__) && !defined(STBI_NO_SIMD)
+// gcc doesn't support sse2 intrinsics unless you compile with -msse2,
+// which in turn means it gets to use SSE2 everywhere. This is unfortunate,
+// but previous attempts to provide the SSE2 functions with runtime
+// detection caused numerous issues. The way architecture extensions are
+// exposed in GCC/Clang is, sadly, not really suited for one-file libs.
+// New behavior: if compiled with -msse2, we use SSE2 without any
+// detection; if not, we don't use it at all.
+#define STBI_NO_SIMD
+#endif
+
+#if defined(__MINGW32__) && defined(STBI__X86_TARGET) && !defined(STBI_MINGW_ENABLE_SSE2) && !defined(STBI_NO_SIMD)
+// Note that __MINGW32__ doesn't actually mean 32-bit, so we have to avoid STBI__X64_TARGET
+//
+// 32-bit MinGW wants ESP to be 16-byte aligned, but this is not in the
+// Windows ABI and VC++ as well as Windows DLLs don't maintain that invariant.
+// As a result, enabling SSE2 on 32-bit MinGW is dangerous when not
+// simultaneously enabling "-mstackrealign".
+//
+// See https://github.com/nothings/stb/issues/81 for more information.
+//
+// So default to no SSE2 on 32-bit MinGW. If you've read this far and added
+// -mstackrealign to your build settings, feel free to #define STBI_MINGW_ENABLE_SSE2.
+#define STBI_NO_SIMD
+#endif
+
+#if !defined(STBI_NO_SIMD) && (defined(STBI__X86_TARGET) || defined(STBI__X64_TARGET))
+#define STBI_SSE2
+#include <emmintrin.h>
+
+#ifdef _MSC_VER
+
+#if _MSC_VER >= 1400  // not VC6
+#include <intrin.h> // __cpuid
+static int stbi__cpuid3(void)
+{
+   int info[4];
+   __cpuid(info,1);
+   return info[3];
+}
+#else
+static int stbi__cpuid3(void)
+{
+   int res;
+   __asm {
+      mov  eax,1
+      cpuid
+      mov  res,edx
+   }
+   return res;
+}
+#endif
+
+#define STBI_SIMD_ALIGN(type, name) __declspec(align(16)) type name
+
+#if !defined(STBI_NO_JPEG) && defined(STBI_SSE2)
+static int stbi__sse2_available(void)
+{
+   int info3 = stbi__cpuid3();
+   return ((info3 >> 26) & 1) != 0;
+}
+#endif
+
+#else // assume GCC-style if not VC++
+#define STBI_SIMD_ALIGN(type, name) type name __attribute__((aligned(16)))
+
+#if !defined(STBI_NO_JPEG) && defined(STBI_SSE2)
+static int stbi__sse2_available(void)
+{
+   // If we're even attempting to compile this on GCC/Clang, that means
+   // -msse2 is on, which means the compiler is allowed to use SSE2
+   // instructions at will, and so are we.
+   return 1;
+}
+#endif
+
+#endif
+#endif
+
+// ARM NEON
+#if defined(STBI_NO_SIMD) && defined(STBI_NEON)
+#undef STBI_NEON
+#endif
+
+#ifdef STBI_NEON
+#include <arm_neon.h>
+#ifdef _MSC_VER
+#define STBI_SIMD_ALIGN(type, name) __declspec(align(16)) type name
+#else
+#define STBI_SIMD_ALIGN(type, name) type name __attribute__((aligned(16)))
+#endif
+#endif
+
+#ifndef STBI_SIMD_ALIGN
+#define STBI_SIMD_ALIGN(type, name) type name
+#endif
+
+#ifndef STBI_MAX_DIMENSIONS
+#define STBI_MAX_DIMENSIONS (1 << 24)
+#endif
+
+///////////////////////////////////////////////
+//
+//  stbi__context struct and start_xxx functions
+
+// stbi__context structure is our basic context used by all images, so it
+// contains all the IO context, plus some basic image information
+typedef struct
+{
+   stbi__uint32 img_x, img_y;
+   int img_n, img_out_n;
+
+   stbi_io_callbacks io;
+   void *io_user_data;
+
+   int read_from_callbacks;
+   int buflen;
+   stbi_uc buffer_start[128];
+   int callback_already_read;
+
+   stbi_uc *img_buffer, *img_buffer_end;
+   stbi_uc *img_buffer_original, *img_buffer_original_end;
+} stbi__context;
+
+
+static void stbi__refill_buffer(stbi__context *s);
+
+// initialize a memory-decode context
+static void stbi__start_mem(stbi__context *s, stbi_uc const *buffer, int len)
+{
+   s->io.read = NULL;
+   s->read_from_callbacks = 0;
+   s->callback_already_read = 0;
+   s->img_buffer = s->img_buffer_original = (stbi_uc *) buffer;
+   s->img_buffer_end = s->img_buffer_original_end = (stbi_uc *) buffer+len;
+}
+
+// initialize a callback-based context
+static void stbi__start_callbacks(stbi__context *s, stbi_io_callbacks *c, void *user)
+{
+   s->io = *c;
+   s->io_user_data = user;
+   s->buflen = sizeof(s->buffer_start);
+   s->read_from_callbacks = 1;
+   s->callback_already_read = 0;
+   s->img_buffer = s->img_buffer_original = s->buffer_start;
+   stbi__refill_buffer(s);
+   s->img_buffer_original_end = s->img_buffer_end;
+}
+
+#ifndef STBI_NO_STDIO
+
+static int stbi__stdio_read(void *user, char *data, int size)
+{
+   return (int) fread(data,1,size,(FILE*) user);
+}
+
+static void stbi__stdio_skip(void *user, int n)
+{
+   int ch;
+   fseek((FILE*) user, n, SEEK_CUR);
+   ch = fgetc((FILE*) user);  /* have to read a byte to reset feof()'s flag */
+   if (ch != EOF) {
+      ungetc(ch, (FILE *) user);  /* push byte back onto stream if valid. */
+   }
+}
+
+static int stbi__stdio_eof(void *user)
+{
+   return feof((FILE*) user) || ferror((FILE *) user);
+}
+
+static stbi_io_callbacks stbi__stdio_callbacks =
+{
+   stbi__stdio_read,
+   stbi__stdio_skip,
+   stbi__stdio_eof,
+};
+
+static void stbi__start_file(stbi__context *s, FILE *f)
+{
+   stbi__start_callbacks(s, &stbi__stdio_callbacks, (void *) f);
+}
+
+//static void stop_file(stbi__context *s) { }
+
+#endif // !STBI_NO_STDIO
+
+static void stbi__rewind(stbi__context *s)
+{
+   // conceptually rewind SHOULD rewind to the beginning of the stream,
+   // but we just rewind to the beginning of the initial buffer, because
+   // we only use it after doing 'test', which only ever looks at at most 92 bytes
+   s->img_buffer = s->img_buffer_original;
+   s->img_buffer_end = s->img_buffer_original_end;
+}
+
+enum
+{
+   STBI_ORDER_RGB,
+   STBI_ORDER_BGR
+};
+
+typedef struct
+{
+   int bits_per_channel;
+   int num_channels;
+   int channel_order;
+} stbi__result_info;
+
+#ifndef STBI_NO_JPEG
+static int      stbi__jpeg_test(stbi__context *s);
+static void    *stbi__jpeg_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri);
+static int      stbi__jpeg_info(stbi__context *s, int *x, int *y, int *comp);
+#endif
+
+#ifndef STBI_NO_PNG
+static int      stbi__png_test(stbi__context *s);
+static void    *stbi__png_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri);
+static int      stbi__png_info(stbi__context *s, int *x, int *y, int *comp);
+static int      stbi__png_is16(stbi__context *s);
+#endif
+
+#ifndef STBI_NO_BMP
+static int      stbi__bmp_test(stbi__context *s);
+static void    *stbi__bmp_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri);
+static int      stbi__bmp_info(stbi__context *s, int *x, int *y, int *comp);
+#endif
+
+#ifndef STBI_NO_TGA
+static int      stbi__tga_test(stbi__context *s);
+static void    *stbi__tga_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri);
+static int      stbi__tga_info(stbi__context *s, int *x, int *y, int *comp);
+#endif
+
+#ifndef STBI_NO_PSD
+static int      stbi__psd_test(stbi__context *s);
+static void    *stbi__psd_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri, int bpc);
+static int      stbi__psd_info(stbi__context *s, int *x, int *y, int *comp);
+static int      stbi__psd_is16(stbi__context *s);
+#endif
+
+#ifndef STBI_NO_HDR
+static int      stbi__hdr_test(stbi__context *s);
+static float   *stbi__hdr_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri);
+static int      stbi__hdr_info(stbi__context *s, int *x, int *y, int *comp);
+#endif
+
+#ifndef STBI_NO_PIC
+static int      stbi__pic_test(stbi__context *s);
+static void    *stbi__pic_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri);
+static int      stbi__pic_info(stbi__context *s, int *x, int *y, int *comp);
+#endif
+
+#ifndef STBI_NO_GIF
+static int      stbi__gif_test(stbi__context *s);
+static void    *stbi__gif_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri);
+static void    *stbi__load_gif_main(stbi__context *s, int **delays, int *x, int *y, int *z, int *comp, int req_comp);
+static int      stbi__gif_info(stbi__context *s, int *x, int *y, int *comp);
+#endif
+
+#ifndef STBI_NO_PNM
+static int      stbi__pnm_test(stbi__context *s);
+static void    *stbi__pnm_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri);
+static int      stbi__pnm_info(stbi__context *s, int *x, int *y, int *comp);
+static int      stbi__pnm_is16(stbi__context *s);
+#endif
+
+static
+#ifdef STBI_THREAD_LOCAL
+STBI_THREAD_LOCAL
+#endif
+const char *stbi__g_failure_reason;
+
+STBIDEF const char *stbi_failure_reason(void)
+{
+   return stbi__g_failure_reason;
+}
+
+#ifndef STBI_NO_FAILURE_STRINGS
+static int stbi__err(const char *str)
+{
+   stbi__g_failure_reason = str;
+   return 0;
+}
+#endif
+
+static void *stbi__malloc(size_t size)
+{
+    return STBI_MALLOC(size);
+}
+
+// stb_image uses ints pervasively, including for offset calculations.
+// therefore the largest decoded image size we can support with the
+// current code, even on 64-bit targets, is INT_MAX. this is not a
+// significant limitation for the intended use case.
+//
+// we do, however, need to make sure our size calculations don't
+// overflow. hence a few helper functions for size calculations that
+// multiply integers together, making sure that they're non-negative
+// and no overflow occurs.
+
+// return 1 if the sum is valid, 0 on overflow.
+// negative terms are considered invalid.
+static int stbi__addsizes_valid(int a, int b)
+{
+   if (b < 0) return 0;
+   // now 0 <= b <= INT_MAX, hence also
+   // 0 <= INT_MAX - b <= INTMAX.
+   // And "a + b <= INT_MAX" (which might overflow) is the
+   // same as a <= INT_MAX - b (no overflow)
+   return a <= INT_MAX - b;
+}
+
+// returns 1 if the product is valid, 0 on overflow.
+// negative factors are considered invalid.
+static int stbi__mul2sizes_valid(int a, int b)
+{
+   if (a < 0 || b < 0) return 0;
+   if (b == 0) return 1; // mul-by-0 is always safe
+   // portable way to check for no overflows in a*b
+   return a <= INT_MAX/b;
+}
+
+#if !defined(STBI_NO_JPEG) || !defined(STBI_NO_PNG) || !defined(STBI_NO_TGA) || !defined(STBI_NO_HDR)
+// returns 1 if "a*b + add" has no negative terms/factors and doesn't overflow
+static int stbi__mad2sizes_valid(int a, int b, int add)
+{
+   return stbi__mul2sizes_valid(a, b) && stbi__addsizes_valid(a*b, add);
+}
+#endif
+
+// returns 1 if "a*b*c + add" has no negative terms/factors and doesn't overflow
+static int stbi__mad3sizes_valid(int a, int b, int c, int add)
+{
+   return stbi__mul2sizes_valid(a, b) && stbi__mul2sizes_valid(a*b, c) &&
+      stbi__addsizes_valid(a*b*c, add);
+}
+
+// returns 1 if "a*b*c*d + add" has no negative terms/factors and doesn't overflow
+#if !defined(STBI_NO_LINEAR) || !defined(STBI_NO_HDR) || !defined(STBI_NO_PNM)
+static int stbi__mad4sizes_valid(int a, int b, int c, int d, int add)
+{
+   return stbi__mul2sizes_valid(a, b) && stbi__mul2sizes_valid(a*b, c) &&
+      stbi__mul2sizes_valid(a*b*c, d) && stbi__addsizes_valid(a*b*c*d, add);
+}
+#endif
+
+#if !defined(STBI_NO_JPEG) || !defined(STBI_NO_PNG) || !defined(STBI_NO_TGA) || !defined(STBI_NO_HDR)
+// mallocs with size overflow checking
+static void *stbi__malloc_mad2(int a, int b, int add)
+{
+   if (!stbi__mad2sizes_valid(a, b, add)) return NULL;
+   return stbi__malloc(a*b + add);
+}
+#endif
+
+static void *stbi__malloc_mad3(int a, int b, int c, int add)
+{
+   if (!stbi__mad3sizes_valid(a, b, c, add)) return NULL;
+   return stbi__malloc(a*b*c + add);
+}
+
+#if !defined(STBI_NO_LINEAR) || !defined(STBI_NO_HDR) || !defined(STBI_NO_PNM)
+static void *stbi__malloc_mad4(int a, int b, int c, int d, int add)
+{
+   if (!stbi__mad4sizes_valid(a, b, c, d, add)) return NULL;
+   return stbi__malloc(a*b*c*d + add);
+}
+#endif
+
+// returns 1 if the sum of two signed ints is valid (between -2^31 and 2^31-1 inclusive), 0 on overflow.
+static int stbi__addints_valid(int a, int b)
+{
+   if ((a >= 0) != (b >= 0)) return 1; // a and b have different signs, so no overflow
+   if (a < 0 && b < 0) return a >= INT_MIN - b; // same as a + b >= INT_MIN; INT_MIN - b cannot overflow since b < 0.
+   return a <= INT_MAX - b;
+}
+
+// returns 1 if the product of two signed shorts is valid, 0 on overflow.
+static int stbi__mul2shorts_valid(short a, short b)
+{
+   if (b == 0 || b == -1) return 1; // multiplication by 0 is always 0; check for -1 so SHRT_MIN/b doesn't overflow
+   if ((a >= 0) == (b >= 0)) return a <= SHRT_MAX/b; // product is positive, so similar to mul2sizes_valid
+   if (b < 0) return a <= SHRT_MIN / b; // same as a * b >= SHRT_MIN
+   return a >= SHRT_MIN / b;
+}
+
+// stbi__err - error
+// stbi__errpf - error returning pointer to float
+// stbi__errpuc - error returning pointer to unsigned char
+
+#ifdef STBI_NO_FAILURE_STRINGS
+   #define stbi__err(x,y)  0
+#elif defined(STBI_FAILURE_USERMSG)
+   #define stbi__err(x,y)  stbi__err(y)
+#else
+   #define stbi__err(x,y)  stbi__err(x)
+#endif
+
+#define stbi__errpf(x,y)   ((float *)(size_t) (stbi__err(x,y)?NULL:NULL))
+#define stbi__errpuc(x,y)  ((unsigned char *)(size_t) (stbi__err(x,y)?NULL:NULL))
+
+STBIDEF void stbi_image_free(void *retval_from_stbi_load)
+{
+   STBI_FREE(retval_from_stbi_load);
+}
+
+#ifndef STBI_NO_LINEAR
+static float   *stbi__ldr_to_hdr(stbi_uc *data, int x, int y, int comp);
+#endif
+
+#ifndef STBI_NO_HDR
+static stbi_uc *stbi__hdr_to_ldr(float   *data, int x, int y, int comp);
+#endif
+
+static int stbi__vertically_flip_on_load_global = 0;
+
+STBIDEF void stbi_set_flip_vertically_on_load(int flag_true_if_should_flip)
+{
+   stbi__vertically_flip_on_load_global = flag_true_if_should_flip;
+}
+
+#ifndef STBI_THREAD_LOCAL
+#define stbi__vertically_flip_on_load  stbi__vertically_flip_on_load_global
+#else
+static STBI_THREAD_LOCAL int stbi__vertically_flip_on_load_local, stbi__vertically_flip_on_load_set;
+
+STBIDEF void stbi_set_flip_vertically_on_load_thread(int flag_true_if_should_flip)
+{
+   stbi__vertically_flip_on_load_local = flag_true_if_should_flip;
+   stbi__vertically_flip_on_load_set = 1;
+}
+
+#define stbi__vertically_flip_on_load  (stbi__vertically_flip_on_load_set       \
+                                         ? stbi__vertically_flip_on_load_local  \
+                                         : stbi__vertically_flip_on_load_global)
+#endif // STBI_THREAD_LOCAL
+
+static void *stbi__load_main(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri, int bpc)
+{
+   memset(ri, 0, sizeof(*ri)); // make sure it's initialized if we add new fields
+   ri->bits_per_channel = 8; // default is 8 so most paths don't have to be changed
+   ri->channel_order = STBI_ORDER_RGB; // all current input & output are this, but this is here so we can add BGR order
+   ri->num_channels = 0;
+
+   // test the formats with a very explicit header first (at least a FOURCC
+   // or distinctive magic number first)
+   #ifndef STBI_NO_PNG
+   if (stbi__png_test(s))  return stbi__png_load(s,x,y,comp,req_comp, ri);
+   #endif
+   #ifndef STBI_NO_BMP
+   if (stbi__bmp_test(s))  return stbi__bmp_load(s,x,y,comp,req_comp, ri);
+   #endif
+   #ifndef STBI_NO_GIF
+   if (stbi__gif_test(s))  return stbi__gif_load(s,x,y,comp,req_comp, ri);
+   #endif
+   #ifndef STBI_NO_PSD
+   if (stbi__psd_test(s))  return stbi__psd_load(s,x,y,comp,req_comp, ri, bpc);
+   #else
+   STBI_NOTUSED(bpc);
+   #endif
+   #ifndef STBI_NO_PIC
+   if (stbi__pic_test(s))  return stbi__pic_load(s,x,y,comp,req_comp, ri);
+   #endif
+
+   // then the formats that can end up attempting to load with just 1 or 2
+   // bytes matching expectations; these are prone to false positives, so
+   // try them later
+   #ifndef STBI_NO_JPEG
+   if (stbi__jpeg_test(s)) return stbi__jpeg_load(s,x,y,comp,req_comp, ri);
+   #endif
+   #ifndef STBI_NO_PNM
+   if (stbi__pnm_test(s))  return stbi__pnm_load(s,x,y,comp,req_comp, ri);
+   #endif
+
+   #ifndef STBI_NO_HDR
+   if (stbi__hdr_test(s)) {
+      float *hdr = stbi__hdr_load(s, x,y,comp,req_comp, ri);
+      return stbi__hdr_to_ldr(hdr, *x, *y, req_comp ? req_comp : *comp);
+   }
+   #endif
+
+   #ifndef STBI_NO_TGA
+   // test tga last because it's a crappy test!
+   if (stbi__tga_test(s))
+      return stbi__tga_load(s,x,y,comp,req_comp, ri);
+   #endif
+
+   return stbi__errpuc("unknown image type", "Image not of any known type, or corrupt");
+}
+
+static stbi_uc *stbi__convert_16_to_8(stbi__uint16 *orig, int w, int h, int channels)
+{
+   int i;
+   int img_len = w * h * channels;
+   stbi_uc *reduced;
+
+   reduced = (stbi_uc *) stbi__malloc(img_len);
+   if (reduced == NULL) return stbi__errpuc("outofmem", "Out of memory");
+
+   for (i = 0; i < img_len; ++i)
+      reduced[i] = (stbi_uc)((orig[i] >> 8) & 0xFF); // top half of each byte is sufficient approx of 16->8 bit scaling
+
+   STBI_FREE(orig);
+   return reduced;
+}
+
+static stbi__uint16 *stbi__convert_8_to_16(stbi_uc *orig, int w, int h, int channels)
+{
+   int i;
+   int img_len = w * h * channels;
+   stbi__uint16 *enlarged;
+
+   enlarged = (stbi__uint16 *) stbi__malloc(img_len*2);
+   if (enlarged == NULL) return (stbi__uint16 *) stbi__errpuc("outofmem", "Out of memory");
+
+   for (i = 0; i < img_len; ++i)
+      enlarged[i] = (stbi__uint16)((orig[i] << 8) + orig[i]); // replicate to high and low byte, maps 0->0, 255->0xffff
+
+   STBI_FREE(orig);
+   return enlarged;
+}
+
+static void stbi__vertical_flip(void *image, int w, int h, int bytes_per_pixel)
+{
+   int row;
+   size_t bytes_per_row = (size_t)w * bytes_per_pixel;
+   stbi_uc temp[2048];
+   stbi_uc *bytes = (stbi_uc *)image;
+
+   for (row = 0; row < (h>>1); row++) {
+      stbi_uc *row0 = bytes + row*bytes_per_row;
+      stbi_uc *row1 = bytes + (h - row - 1)*bytes_per_row;
+      // swap row0 with row1
+      size_t bytes_left = bytes_per_row;
+      while (bytes_left) {
+         size_t bytes_copy = (bytes_left < sizeof(temp)) ? bytes_left : sizeof(temp);
+         memcpy(temp, row0, bytes_copy);
+         memcpy(row0, row1, bytes_copy);
+         memcpy(row1, temp, bytes_copy);
+         row0 += bytes_copy;
+         row1 += bytes_copy;
+         bytes_left -= bytes_copy;
+      }
+   }
+}
+
+#ifndef STBI_NO_GIF
+static void stbi__vertical_flip_slices(void *image, int w, int h, int z, int bytes_per_pixel)
+{
+   int slice;
+   int slice_size = w * h * bytes_per_pixel;
+
+   stbi_uc *bytes = (stbi_uc *)image;
+   for (slice = 0; slice < z; ++slice) {
+      stbi__vertical_flip(bytes, w, h, bytes_per_pixel);
+      bytes += slice_size;
+   }
+}
+#endif
+
+static unsigned char *stbi__load_and_postprocess_8bit(stbi__context *s, int *x, int *y, int *comp, int req_comp)
+{
+   stbi__result_info ri;
+   void *result = stbi__load_main(s, x, y, comp, req_comp, &ri, 8);
+
+   if (result == NULL)
+      return NULL;
+
+   // it is the responsibility of the loaders to make sure we get either 8 or 16 bit.
+   STBI_ASSERT(ri.bits_per_channel == 8 || ri.bits_per_channel == 16);
+
+   if (ri.bits_per_channel != 8) {
+      result = stbi__convert_16_to_8((stbi__uint16 *) result, *x, *y, req_comp == 0 ? *comp : req_comp);
+      ri.bits_per_channel = 8;
+   }
+
+   // @TODO: move stbi__convert_format to here
+
+   if (stbi__vertically_flip_on_load) {
+      int channels = req_comp ? req_comp : *comp;
+      stbi__vertical_flip(result, *x, *y, channels * sizeof(stbi_uc));
+   }
+
+   return (unsigned char *) result;
+}
+
+static stbi__uint16 *stbi__load_and_postprocess_16bit(stbi__context *s, int *x, int *y, int *comp, int req_comp)
+{
+   stbi__result_info ri;
+   void *result = stbi__load_main(s, x, y, comp, req_comp, &ri, 16);
+
+   if (result == NULL)
+      return NULL;
+
+   // it is the responsibility of the loaders to make sure we get either 8 or 16 bit.
+   STBI_ASSERT(ri.bits_per_channel == 8 || ri.bits_per_channel == 16);
+
+   if (ri.bits_per_channel != 16) {
+      result = stbi__convert_8_to_16((stbi_uc *) result, *x, *y, req_comp == 0 ? *comp : req_comp);
+      ri.bits_per_channel = 16;
+   }
+
+   // @TODO: move stbi__convert_format16 to here
+   // @TODO: special case RGB-to-Y (and RGBA-to-YA) for 8-bit-to-16-bit case to keep more precision
+
+   if (stbi__vertically_flip_on_load) {
+      int channels = req_comp ? req_comp : *comp;
+      stbi__vertical_flip(result, *x, *y, channels * sizeof(stbi__uint16));
+   }
+
+   return (stbi__uint16 *) result;
+}
+
+#if !defined(STBI_NO_HDR) && !defined(STBI_NO_LINEAR)
+static void stbi__float_postprocess(float *result, int *x, int *y, int *comp, int req_comp)
+{
+   if (stbi__vertically_flip_on_load && result != NULL) {
+      int channels = req_comp ? req_comp : *comp;
+      stbi__vertical_flip(result, *x, *y, channels * sizeof(float));
+   }
+}
+#endif
+
+#ifndef STBI_NO_STDIO
+
+#if defined(_WIN32) && defined(STBI_WINDOWS_UTF8)
+STBI_EXTERN __declspec(dllimport) int __stdcall MultiByteToWideChar(unsigned int cp, unsigned long flags, const char *str, int cbmb, wchar_t *widestr, int cchwide);
+STBI_EXTERN __declspec(dllimport) int __stdcall WideCharToMultiByte(unsigned int cp, unsigned long flags, const wchar_t *widestr, int cchwide, char *str, int cbmb, const char *defchar, int *used_default);
+#endif
+
+#if defined(_WIN32) && defined(STBI_WINDOWS_UTF8)
+STBIDEF int stbi_convert_wchar_to_utf8(char *buffer, size_t bufferlen, const wchar_t* input)
+{
+	return WideCharToMultiByte(65001 /* UTF8 */, 0, input, -1, buffer, (int) bufferlen, NULL, NULL);
+}
+#endif
+
+static FILE *stbi__fopen(char const *filename, char const *mode)
+{
+   FILE *f;
+#if defined(_WIN32) && defined(STBI_WINDOWS_UTF8)
+   wchar_t wMode[64];
+   wchar_t wFilename[1024];
+	if (0 == MultiByteToWideChar(65001 /* UTF8 */, 0, filename, -1, wFilename, sizeof(wFilename)/sizeof(*wFilename)))
+      return 0;
+
+	if (0 == MultiByteToWideChar(65001 /* UTF8 */, 0, mode, -1, wMode, sizeof(wMode)/sizeof(*wMode)))
+      return 0;
+
+#if defined(_MSC_VER) && _MSC_VER >= 1400
+	if (0 != _wfopen_s(&f, wFilename, wMode))
+		f = 0;
+#else
+   f = _wfopen(wFilename, wMode);
+#endif
+
+#elif defined(_MSC_VER) && _MSC_VER >= 1400
+   if (0 != fopen_s(&f, filename, mode))
+      f=0;
+#else
+   f = fopen(filename, mode);
+#endif
+   return f;
+}
+
+
+STBIDEF stbi_uc *stbi_load(char const *filename, int *x, int *y, int *comp, int req_comp)
+{
+   FILE *f = stbi__fopen(filename, "rb");
+   unsigned char *result;
+   if (!f) return stbi__errpuc("can't fopen", "Unable to open file");
+   result = stbi_load_from_file(f,x,y,comp,req_comp);
+   fclose(f);
+   return result;
+}
+
+STBIDEF stbi_uc *stbi_load_from_file(FILE *f, int *x, int *y, int *comp, int req_comp)
+{
+   unsigned char *result;
+   stbi__context s;
+   stbi__start_file(&s,f);
+   result = stbi__load_and_postprocess_8bit(&s,x,y,comp,req_comp);
+   if (result) {
+      // need to 'unget' all the characters in the IO buffer
+      fseek(f, - (int) (s.img_buffer_end - s.img_buffer), SEEK_CUR);
+   }
+   return result;
+}
+
+STBIDEF stbi__uint16 *stbi_load_from_file_16(FILE *f, int *x, int *y, int *comp, int req_comp)
+{
+   stbi__uint16 *result;
+   stbi__context s;
+   stbi__start_file(&s,f);
+   result = stbi__load_and_postprocess_16bit(&s,x,y,comp,req_comp);
+   if (result) {
+      // need to 'unget' all the characters in the IO buffer
+      fseek(f, - (int) (s.img_buffer_end - s.img_buffer), SEEK_CUR);
+   }
+   return result;
+}
+
+STBIDEF stbi_us *stbi_load_16(char const *filename, int *x, int *y, int *comp, int req_comp)
+{
+   FILE *f = stbi__fopen(filename, "rb");
+   stbi__uint16 *result;
+   if (!f) return (stbi_us *) stbi__errpuc("can't fopen", "Unable to open file");
+   result = stbi_load_from_file_16(f,x,y,comp,req_comp);
+   fclose(f);
+   return result;
+}
+
+
+#endif //!STBI_NO_STDIO
+
+STBIDEF stbi_us *stbi_load_16_from_memory(stbi_uc const *buffer, int len, int *x, int *y, int *channels_in_file, int desired_channels)
+{
+   stbi__context s;
+   stbi__start_mem(&s,buffer,len);
+   return stbi__load_and_postprocess_16bit(&s,x,y,channels_in_file,desired_channels);
+}
+
+STBIDEF stbi_us *stbi_load_16_from_callbacks(stbi_io_callbacks const *clbk, void *user, int *x, int *y, int *channels_in_file, int desired_channels)
+{
+   stbi__context s;
+   stbi__start_callbacks(&s, (stbi_io_callbacks *)clbk, user);
+   return stbi__load_and_postprocess_16bit(&s,x,y,channels_in_file,desired_channels);
+}
+
+STBIDEF stbi_uc *stbi_load_from_memory(stbi_uc const *buffer, int len, int *x, int *y, int *comp, int req_comp)
+{
+   stbi__context s;
+   stbi__start_mem(&s,buffer,len);
+   return stbi__load_and_postprocess_8bit(&s,x,y,comp,req_comp);
+}
+
+STBIDEF stbi_uc *stbi_load_from_callbacks(stbi_io_callbacks const *clbk, void *user, int *x, int *y, int *comp, int req_comp)
+{
+   stbi__context s;
+   stbi__start_callbacks(&s, (stbi_io_callbacks *) clbk, user);
+   return stbi__load_and_postprocess_8bit(&s,x,y,comp,req_comp);
+}
+
+#ifndef STBI_NO_GIF
+STBIDEF stbi_uc *stbi_load_gif_from_memory(stbi_uc const *buffer, int len, int **delays, int *x, int *y, int *z, int *comp, int req_comp)
+{
+   unsigned char *result;
+   stbi__context s;
+   stbi__start_mem(&s,buffer,len);
+
+   result = (unsigned char*) stbi__load_gif_main(&s, delays, x, y, z, comp, req_comp);
+   if (stbi__vertically_flip_on_load) {
+      stbi__vertical_flip_slices( result, *x, *y, *z, *comp );
+   }
+
+   return result;
+}
+#endif
+
+#ifndef STBI_NO_LINEAR
+static float *stbi__loadf_main(stbi__context *s, int *x, int *y, int *comp, int req_comp)
+{
+   unsigned char *data;
+   #ifndef STBI_NO_HDR
+   if (stbi__hdr_test(s)) {
+      stbi__result_info ri;
+      float *hdr_data = stbi__hdr_load(s,x,y,comp,req_comp, &ri);
+      if (hdr_data)
+         stbi__float_postprocess(hdr_data,x,y,comp,req_comp);
+      return hdr_data;
+   }
+   #endif
+   data = stbi__load_and_postprocess_8bit(s, x, y, comp, req_comp);
+   if (data)
+      return stbi__ldr_to_hdr(data, *x, *y, req_comp ? req_comp : *comp);
+   return stbi__errpf("unknown image type", "Image not of any known type, or corrupt");
+}
+
+STBIDEF float *stbi_loadf_from_memory(stbi_uc const *buffer, int len, int *x, int *y, int *comp, int req_comp)
+{
+   stbi__context s;
+   stbi__start_mem(&s,buffer,len);
+   return stbi__loadf_main(&s,x,y,comp,req_comp);
+}
+
+STBIDEF float *stbi_loadf_from_callbacks(stbi_io_callbacks const *clbk, void *user, int *x, int *y, int *comp, int req_comp)
+{
+   stbi__context s;
+   stbi__start_callbacks(&s, (stbi_io_callbacks *) clbk, user);
+   return stbi__loadf_main(&s,x,y,comp,req_comp);
+}
+
+#ifndef STBI_NO_STDIO
+STBIDEF float *stbi_loadf(char const *filename, int *x, int *y, int *comp, int req_comp)
+{
+   float *result;
+   FILE *f = stbi__fopen(filename, "rb");
+   if (!f) return stbi__errpf("can't fopen", "Unable to open file");
+   result = stbi_loadf_from_file(f,x,y,comp,req_comp);
+   fclose(f);
+   return result;
+}
+
+STBIDEF float *stbi_loadf_from_file(FILE *f, int *x, int *y, int *comp, int req_comp)
+{
+   stbi__context s;
+   stbi__start_file(&s,f);
+   return stbi__loadf_main(&s,x,y,comp,req_comp);
+}
+#endif // !STBI_NO_STDIO
+
+#endif // !STBI_NO_LINEAR
+
+// these is-hdr-or-not is defined independent of whether STBI_NO_LINEAR is
+// defined, for API simplicity; if STBI_NO_LINEAR is defined, it always
+// reports false!
+
+STBIDEF int stbi_is_hdr_from_memory(stbi_uc const *buffer, int len)
+{
+   #ifndef STBI_NO_HDR
+   stbi__context s;
+   stbi__start_mem(&s,buffer,len);
+   return stbi__hdr_test(&s);
+   #else
+   STBI_NOTUSED(buffer);
+   STBI_NOTUSED(len);
+   return 0;
+   #endif
+}
+
+#ifndef STBI_NO_STDIO
+STBIDEF int      stbi_is_hdr          (char const *filename)
+{
+   FILE *f = stbi__fopen(filename, "rb");
+   int result=0;
+   if (f) {
+      result = stbi_is_hdr_from_file(f);
+      fclose(f);
+   }
+   return result;
+}
+
+STBIDEF int stbi_is_hdr_from_file(FILE *f)
+{
+   #ifndef STBI_NO_HDR
+   long pos = ftell(f);
+   int res;
+   stbi__context s;
+   stbi__start_file(&s,f);
+   res = stbi__hdr_test(&s);
+   fseek(f, pos, SEEK_SET);
+   return res;
+   #else
+   STBI_NOTUSED(f);
+   return 0;
+   #endif
+}
+#endif // !STBI_NO_STDIO
+
+STBIDEF int      stbi_is_hdr_from_callbacks(stbi_io_callbacks const *clbk, void *user)
+{
+   #ifndef STBI_NO_HDR
+   stbi__context s;
+   stbi__start_callbacks(&s, (stbi_io_callbacks *) clbk, user);
+   return stbi__hdr_test(&s);
+   #else
+   STBI_NOTUSED(clbk);
+   STBI_NOTUSED(user);
+   return 0;
+   #endif
+}
+
+#ifndef STBI_NO_LINEAR
+static float stbi__l2h_gamma=2.2f, stbi__l2h_scale=1.0f;
+
+STBIDEF void   stbi_ldr_to_hdr_gamma(float gamma) { stbi__l2h_gamma = gamma; }
+STBIDEF void   stbi_ldr_to_hdr_scale(float scale) { stbi__l2h_scale = scale; }
+#endif
+
+static float stbi__h2l_gamma_i=1.0f/2.2f, stbi__h2l_scale_i=1.0f;
+
+STBIDEF void   stbi_hdr_to_ldr_gamma(float gamma) { stbi__h2l_gamma_i = 1/gamma; }
+STBIDEF void   stbi_hdr_to_ldr_scale(float scale) { stbi__h2l_scale_i = 1/scale; }
+
+
+//////////////////////////////////////////////////////////////////////////////
+//
+// Common code used by all image loaders
+//
+
+enum
+{
+   STBI__SCAN_load=0,
+   STBI__SCAN_type,
+   STBI__SCAN_header
+};
+
+static void stbi__refill_buffer(stbi__context *s)
+{
+   int n = (s->io.read)(s->io_user_data,(char*)s->buffer_start,s->buflen);
+   s->callback_already_read += (int) (s->img_buffer - s->img_buffer_original);
+   if (n == 0) {
+      // at end of file, treat same as if from memory, but need to handle case
+      // where s->img_buffer isn't pointing to safe memory, e.g. 0-byte file
+      s->read_from_callbacks = 0;
+      s->img_buffer = s->buffer_start;
+      s->img_buffer_end = s->buffer_start+1;
+      *s->img_buffer = 0;
+   } else {
+      s->img_buffer = s->buffer_start;
+      s->img_buffer_end = s->buffer_start + n;
+   }
+}
+
+stbi_inline static stbi_uc stbi__get8(stbi__context *s)
+{
+   if (s->img_buffer < s->img_buffer_end)
+      return *s->img_buffer++;
+   if (s->read_from_callbacks) {
+      stbi__refill_buffer(s);
+      return *s->img_buffer++;
+   }
+   return 0;
+}
+
+#if defined(STBI_NO_JPEG) && defined(STBI_NO_HDR) && defined(STBI_NO_PIC) && defined(STBI_NO_PNM)
+// nothing
+#else
+stbi_inline static int stbi__at_eof(stbi__context *s)
+{
+   if (s->io.read) {
+      if (!(s->io.eof)(s->io_user_data)) return 0;
+      // if feof() is true, check if buffer = end
+      // special case: we've only got the special 0 character at the end
+      if (s->read_from_callbacks == 0) return 1;
+   }
+
+   return s->img_buffer >= s->img_buffer_end;
+}
+#endif
+
+#if defined(STBI_NO_JPEG) && defined(STBI_NO_PNG) && defined(STBI_NO_BMP) && defined(STBI_NO_PSD) && defined(STBI_NO_TGA) && defined(STBI_NO_GIF) && defined(STBI_NO_PIC)
+// nothing
+#else
+static void stbi__skip(stbi__context *s, int n)
+{
+   if (n == 0) return;  // already there!
+   if (n < 0) {
+      s->img_buffer = s->img_buffer_end;
+      return;
+   }
+   if (s->io.read) {
+      int blen = (int) (s->img_buffer_end - s->img_buffer);
+      if (blen < n) {
+         s->img_buffer = s->img_buffer_end;
+         (s->io.skip)(s->io_user_data, n - blen);
+         return;
+      }
+   }
+   s->img_buffer += n;
+}
+#endif
+
+#if defined(STBI_NO_PNG) && defined(STBI_NO_TGA) && defined(STBI_NO_HDR) && defined(STBI_NO_PNM)
+// nothing
+#else
+static int stbi__getn(stbi__context *s, stbi_uc *buffer, int n)
+{
+   if (s->io.read) {
+      int blen = (int) (s->img_buffer_end - s->img_buffer);
+      if (blen < n) {
+         int res, count;
+
+         memcpy(buffer, s->img_buffer, blen);
+
+         count = (s->io.read)(s->io_user_data, (char*) buffer + blen, n - blen);
+         res = (count == (n-blen));
+         s->img_buffer = s->img_buffer_end;
+         return res;
+      }
+   }
+
+   if (s->img_buffer+n <= s->img_buffer_end) {
+      memcpy(buffer, s->img_buffer, n);
+      s->img_buffer += n;
+      return 1;
+   } else
+      return 0;
+}
+#endif
+
+#if defined(STBI_NO_JPEG) && defined(STBI_NO_PNG) && defined(STBI_NO_PSD) && defined(STBI_NO_PIC)
+// nothing
+#else
+static int stbi__get16be(stbi__context *s)
+{
+   int z = stbi__get8(s);
+   return (z << 8) + stbi__get8(s);
+}
+#endif
+
+#if defined(STBI_NO_PNG) && defined(STBI_NO_PSD) && defined(STBI_NO_PIC)
+// nothing
+#else
+static stbi__uint32 stbi__get32be(stbi__context *s)
+{
+   stbi__uint32 z = stbi__get16be(s);
+   return (z << 16) + stbi__get16be(s);
+}
+#endif
+
+#if defined(STBI_NO_BMP) && defined(STBI_NO_TGA) && defined(STBI_NO_GIF)
+// nothing
+#else
+static int stbi__get16le(stbi__context *s)
+{
+   int z = stbi__get8(s);
+   return z + (stbi__get8(s) << 8);
+}
+#endif
+
+#ifndef STBI_NO_BMP
+static stbi__uint32 stbi__get32le(stbi__context *s)
+{
+   stbi__uint32 z = stbi__get16le(s);
+   z += (stbi__uint32)stbi__get16le(s) << 16;
+   return z;
+}
+#endif
+
+#define STBI__BYTECAST(x)  ((stbi_uc) ((x) & 255))  // truncate int to byte without warnings
+
+#if defined(STBI_NO_JPEG) && defined(STBI_NO_PNG) && defined(STBI_NO_BMP) && defined(STBI_NO_PSD) && defined(STBI_NO_TGA) && defined(STBI_NO_GIF) && defined(STBI_NO_PIC) && defined(STBI_NO_PNM)
+// nothing
+#else
+//////////////////////////////////////////////////////////////////////////////
+//
+//  generic converter from built-in img_n to req_comp
+//    individual types do this automatically as much as possible (e.g. jpeg
+//    does all cases internally since it needs to colorspace convert anyway,
+//    and it never has alpha, so very few cases ). png can automatically
+//    interleave an alpha=255 channel, but falls back to this for other cases
+//
+//  assume data buffer is malloced, so malloc a new one and free that one
+//  only failure mode is malloc failing
+
+static stbi_uc stbi__compute_y(int r, int g, int b)
+{
+   return (stbi_uc) (((r*77) + (g*150) +  (29*b)) >> 8);
+}
+#endif
+
+#if defined(STBI_NO_PNG) && defined(STBI_NO_BMP) && defined(STBI_NO_PSD) && defined(STBI_NO_TGA) && defined(STBI_NO_GIF) && defined(STBI_NO_PIC) && defined(STBI_NO_PNM)
+// nothing
+#else
+static unsigned char *stbi__convert_format(unsigned char *data, int img_n, int req_comp, unsigned int x, unsigned int y)
+{
+   int i,j;
+   unsigned char *good;
+
+   if (req_comp == img_n) return data;
+   STBI_ASSERT(req_comp >= 1 && req_comp <= 4);
+
+   good = (unsigned char *) stbi__malloc_mad3(req_comp, x, y, 0);
+   if (good == NULL) {
+      STBI_FREE(data);
+      return stbi__errpuc("outofmem", "Out of memory");
+   }
+
+   for (j=0; j < (int) y; ++j) {
+      unsigned char *src  = data + j * x * img_n   ;
+      unsigned char *dest = good + j * x * req_comp;
+
+      #define STBI__COMBO(a,b)  ((a)*8+(b))
+      #define STBI__CASE(a,b)   case STBI__COMBO(a,b): for(i=x-1; i >= 0; --i, src += a, dest += b)
+      // convert source image with img_n components to one with req_comp components;
+      // avoid switch per pixel, so use switch per scanline and massive macros
+      switch (STBI__COMBO(img_n, req_comp)) {
+         STBI__CASE(1,2) { dest[0]=src[0]; dest[1]=255;                                     } break;
+         STBI__CASE(1,3) { dest[0]=dest[1]=dest[2]=src[0];                                  } break;
+         STBI__CASE(1,4) { dest[0]=dest[1]=dest[2]=src[0]; dest[3]=255;                     } break;
+         STBI__CASE(2,1) { dest[0]=src[0];                                                  } break;
+         STBI__CASE(2,3) { dest[0]=dest[1]=dest[2]=src[0];                                  } break;
+         STBI__CASE(2,4) { dest[0]=dest[1]=dest[2]=src[0]; dest[3]=src[1];                  } break;
+         STBI__CASE(3,4) { dest[0]=src[0];dest[1]=src[1];dest[2]=src[2];dest[3]=255;        } break;
+         STBI__CASE(3,1) { dest[0]=stbi__compute_y(src[0],src[1],src[2]);                   } break;
+         STBI__CASE(3,2) { dest[0]=stbi__compute_y(src[0],src[1],src[2]); dest[1] = 255;    } break;
+         STBI__CASE(4,1) { dest[0]=stbi__compute_y(src[0],src[1],src[2]);                   } break;
+         STBI__CASE(4,2) { dest[0]=stbi__compute_y(src[0],src[1],src[2]); dest[1] = src[3]; } break;
+         STBI__CASE(4,3) { dest[0]=src[0];dest[1]=src[1];dest[2]=src[2];                    } break;
+         default: STBI_ASSERT(0); STBI_FREE(data); STBI_FREE(good); return stbi__errpuc("unsupported", "Unsupported format conversion");
+      }
+      #undef STBI__CASE
+   }
+
+   STBI_FREE(data);
+   return good;
+}
+#endif
+
+#if defined(STBI_NO_PNG) && defined(STBI_NO_PSD)
+// nothing
+#else
+static stbi__uint16 stbi__compute_y_16(int r, int g, int b)
+{
+   return (stbi__uint16) (((r*77) + (g*150) +  (29*b)) >> 8);
+}
+#endif
+
+#if defined(STBI_NO_PNG) && defined(STBI_NO_PSD)
+// nothing
+#else
+static stbi__uint16 *stbi__convert_format16(stbi__uint16 *data, int img_n, int req_comp, unsigned int x, unsigned int y)
+{
+   int i,j;
+   stbi__uint16 *good;
+
+   if (req_comp == img_n) return data;
+   STBI_ASSERT(req_comp >= 1 && req_comp <= 4);
+
+   good = (stbi__uint16 *) stbi__malloc(req_comp * x * y * 2);
+   if (good == NULL) {
+      STBI_FREE(data);
+      return (stbi__uint16 *) stbi__errpuc("outofmem", "Out of memory");
+   }
+
+   for (j=0; j < (int) y; ++j) {
+      stbi__uint16 *src  = data + j * x * img_n   ;
+      stbi__uint16 *dest = good + j * x * req_comp;
+
+      #define STBI__COMBO(a,b)  ((a)*8+(b))
+      #define STBI__CASE(a,b)   case STBI__COMBO(a,b): for(i=x-1; i >= 0; --i, src += a, dest += b)
+      // convert source image with img_n components to one with req_comp components;
+      // avoid switch per pixel, so use switch per scanline and massive macros
+      switch (STBI__COMBO(img_n, req_comp)) {
+         STBI__CASE(1,2) { dest[0]=src[0]; dest[1]=0xffff;                                     } break;
+         STBI__CASE(1,3) { dest[0]=dest[1]=dest[2]=src[0];                                     } break;
+         STBI__CASE(1,4) { dest[0]=dest[1]=dest[2]=src[0]; dest[3]=0xffff;                     } break;
+         STBI__CASE(2,1) { dest[0]=src[0];                                                     } break;
+         STBI__CASE(2,3) { dest[0]=dest[1]=dest[2]=src[0];                                     } break;
+         STBI__CASE(2,4) { dest[0]=dest[1]=dest[2]=src[0]; dest[3]=src[1];                     } break;
+         STBI__CASE(3,4) { dest[0]=src[0];dest[1]=src[1];dest[2]=src[2];dest[3]=0xffff;        } break;
+         STBI__CASE(3,1) { dest[0]=stbi__compute_y_16(src[0],src[1],src[2]);                   } break;
+         STBI__CASE(3,2) { dest[0]=stbi__compute_y_16(src[0],src[1],src[2]); dest[1] = 0xffff; } break;
+         STBI__CASE(4,1) { dest[0]=stbi__compute_y_16(src[0],src[1],src[2]);                   } break;
+         STBI__CASE(4,2) { dest[0]=stbi__compute_y_16(src[0],src[1],src[2]); dest[1] = src[3]; } break;
+         STBI__CASE(4,3) { dest[0]=src[0];dest[1]=src[1];dest[2]=src[2];                       } break;
+         default: STBI_ASSERT(0); STBI_FREE(data); STBI_FREE(good); return (stbi__uint16*) stbi__errpuc("unsupported", "Unsupported format conversion");
+      }
+      #undef STBI__CASE
+   }
+
+   STBI_FREE(data);
+   return good;
+}
+#endif
+
+#ifndef STBI_NO_LINEAR
+static float   *stbi__ldr_to_hdr(stbi_uc *data, int x, int y, int comp)
+{
+   int i,k,n;
+   float *output;
+   if (!data) return NULL;
+   output = (float *) stbi__malloc_mad4(x, y, comp, sizeof(float), 0);
+   if (output == NULL) { STBI_FREE(data); return stbi__errpf("outofmem", "Out of memory"); }
+   // compute number of non-alpha components
+   if (comp & 1) n = comp; else n = comp-1;
+   for (i=0; i < x*y; ++i) {
+      for (k=0; k < n; ++k) {
+         output[i*comp + k] = (float) (pow(data[i*comp+k]/255.0f, stbi__l2h_gamma) * stbi__l2h_scale);
+      }
+   }
+   if (n < comp) {
+      for (i=0; i < x*y; ++i) {
+         output[i*comp + n] = data[i*comp + n]/255.0f;
+      }
+   }
+   STBI_FREE(data);
+   return output;
+}
+#endif
+
+#ifndef STBI_NO_HDR
+#define stbi__float2int(x)   ((int) (x))
+static stbi_uc *stbi__hdr_to_ldr(float   *data, int x, int y, int comp)
+{
+   int i,k,n;
+   stbi_uc *output;
+   if (!data) return NULL;
+   output = (stbi_uc *) stbi__malloc_mad3(x, y, comp, 0);
+   if (output == NULL) { STBI_FREE(data); return stbi__errpuc("outofmem", "Out of memory"); }
+   // compute number of non-alpha components
+   if (comp & 1) n = comp; else n = comp-1;
+   for (i=0; i < x*y; ++i) {
+      for (k=0; k < n; ++k) {
+         float z = (float) pow(data[i*comp+k]*stbi__h2l_scale_i, stbi__h2l_gamma_i) * 255 + 0.5f;
+         if (z < 0) z = 0;
+         if (z > 255) z = 255;
+         output[i*comp + k] = (stbi_uc) stbi__float2int(z);
+      }
+      if (k < comp) {
+         float z = data[i*comp+k] * 255 + 0.5f;
+         if (z < 0) z = 0;
+         if (z > 255) z = 255;
+         output[i*comp + k] = (stbi_uc) stbi__float2int(z);
+      }
+   }
+   STBI_FREE(data);
+   return output;
+}
+#endif
+
+//////////////////////////////////////////////////////////////////////////////
+//
+//  "baseline" JPEG/JFIF decoder
+//
+//    simple implementation
+//      - doesn't support delayed output of y-dimension
+//      - simple interface (only one output format: 8-bit interleaved RGB)
+//      - doesn't try to recover corrupt jpegs
+//      - doesn't allow partial loading, loading multiple at once
+//      - still fast on x86 (copying globals into locals doesn't help x86)
+//      - allocates lots of intermediate memory (full size of all components)
+//        - non-interleaved case requires this anyway
+//        - allows good upsampling (see next)
+//    high-quality
+//      - upsampled channels are bilinearly interpolated, even across blocks
+//      - quality integer IDCT derived from IJG's 'slow'
+//    performance
+//      - fast huffman; reasonable integer IDCT
+//      - some SIMD kernels for common paths on targets with SSE2/NEON
+//      - uses a lot of intermediate memory, could cache poorly
+
+#ifndef STBI_NO_JPEG
+
+// huffman decoding acceleration
+#define FAST_BITS   9  // larger handles more cases; smaller stomps less cache
+
+typedef struct
+{
+   stbi_uc  fast[1 << FAST_BITS];
+   // weirdly, repacking this into AoS is a 10% speed loss, instead of a win
+   stbi__uint16 code[256];
+   stbi_uc  values[256];
+   stbi_uc  size[257];
+   unsigned int maxcode[18];
+   int    delta[17];   // old 'firstsymbol' - old 'firstcode'
+} stbi__huffman;
+
+typedef struct
+{
+   stbi__context *s;
+   stbi__huffman huff_dc[4];
+   stbi__huffman huff_ac[4];
+   stbi__uint16 dequant[4][64];
+   stbi__int16 fast_ac[4][1 << FAST_BITS];
+
+// sizes for components, interleaved MCUs
+   int img_h_max, img_v_max;
+   int img_mcu_x, img_mcu_y;
+   int img_mcu_w, img_mcu_h;
+
+// definition of jpeg image component
+   struct
+   {
+      int id;
+      int h,v;
+      int tq;
+      int hd,ha;
+      int dc_pred;
+
+      int x,y,w2,h2;
+      stbi_uc *data;
+      void *raw_data, *raw_coeff;
+      stbi_uc *linebuf;
+      short   *coeff;   // progressive only
+      int      coeff_w, coeff_h; // number of 8x8 coefficient blocks
+   } img_comp[4];
+
+   stbi__uint32   code_buffer; // jpeg entropy-coded buffer
+   int            code_bits;   // number of valid bits
+   unsigned char  marker;      // marker seen while filling entropy buffer
+   int            nomore;      // flag if we saw a marker so must stop
+
+   int            progressive;
+   int            spec_start;
+   int            spec_end;
+   int            succ_high;
+   int            succ_low;
+   int            eob_run;
+   int            jfif;
+   int            app14_color_transform; // Adobe APP14 tag
+   int            rgb;
+
+   int scan_n, order[4];
+   int restart_interval, todo;
+
+// kernels
+   void (*idct_block_kernel)(stbi_uc *out, int out_stride, short data[64]);
+   void (*YCbCr_to_RGB_kernel)(stbi_uc *out, const stbi_uc *y, const stbi_uc *pcb, const stbi_uc *pcr, int count, int step);
+   stbi_uc *(*resample_row_hv_2_kernel)(stbi_uc *out, stbi_uc *in_near, stbi_uc *in_far, int w, int hs);
+} stbi__jpeg;
+
+static int stbi__build_huffman(stbi__huffman *h, int *count)
+{
+   int i,j,k=0;
+   unsigned int code;
+   // build size list for each symbol (from JPEG spec)
+   for (i=0; i < 16; ++i) {
+      for (j=0; j < count[i]; ++j) {
+         h->size[k++] = (stbi_uc) (i+1);
+         if(k >= 257) return stbi__err("bad size list","Corrupt JPEG");
+      }
+   }
+   h->size[k] = 0;
+
+   // compute actual symbols (from jpeg spec)
+   code = 0;
+   k = 0;
+   for(j=1; j <= 16; ++j) {
+      // compute delta to add to code to compute symbol id
+      h->delta[j] = k - code;
+      if (h->size[k] == j) {
+         while (h->size[k] == j)
+            h->code[k++] = (stbi__uint16) (code++);
+         if (code-1 >= (1u << j)) return stbi__err("bad code lengths","Corrupt JPEG");
+      }
+      // compute largest code + 1 for this size, preshifted as needed later
+      h->maxcode[j] = code << (16-j);
+      code <<= 1;
+   }
+   h->maxcode[j] = 0xffffffff;
+
+   // build non-spec acceleration table; 255 is flag for not-accelerated
+   memset(h->fast, 255, 1 << FAST_BITS);
+   for (i=0; i < k; ++i) {
+      int s = h->size[i];
+      if (s <= FAST_BITS) {
+         int c = h->code[i] << (FAST_BITS-s);
+         int m = 1 << (FAST_BITS-s);
+         for (j=0; j < m; ++j) {
+            h->fast[c+j] = (stbi_uc) i;
+         }
+      }
+   }
+   return 1;
+}
+
+// build a table that decodes both magnitude and value of small ACs in
+// one go.
+static void stbi__build_fast_ac(stbi__int16 *fast_ac, stbi__huffman *h)
+{
+   int i;
+   for (i=0; i < (1 << FAST_BITS); ++i) {
+      stbi_uc fast = h->fast[i];
+      fast_ac[i] = 0;
+      if (fast < 255) {
+         int rs = h->values[fast];
+         int run = (rs >> 4) & 15;
+         int magbits = rs & 15;
+         int len = h->size[fast];
+
+         if (magbits && len + magbits <= FAST_BITS) {
+            // magnitude code followed by receive_extend code
+            int k = ((i << len) & ((1 << FAST_BITS) - 1)) >> (FAST_BITS - magbits);
+            int m = 1 << (magbits - 1);
+            if (k < m) k += (~0U << magbits) + 1;
+            // if the result is small enough, we can fit it in fast_ac table
+            if (k >= -128 && k <= 127)
+               fast_ac[i] = (stbi__int16) ((k * 256) + (run * 16) + (len + magbits));
+         }
+      }
+   }
+}
+
+static void stbi__grow_buffer_unsafe(stbi__jpeg *j)
+{
+   do {
+      unsigned int b = j->nomore ? 0 : stbi__get8(j->s);
+      if (b == 0xff) {
+         int c = stbi__get8(j->s);
+         while (c == 0xff) c = stbi__get8(j->s); // consume fill bytes
+         if (c != 0) {
+            j->marker = (unsigned char) c;
+            j->nomore = 1;
+            return;
+         }
+      }
+      j->code_buffer |= b << (24 - j->code_bits);
+      j->code_bits += 8;
+   } while (j->code_bits <= 24);
+}
+
+// (1 << n) - 1
+static const stbi__uint32 stbi__bmask[17]={0,1,3,7,15,31,63,127,255,511,1023,2047,4095,8191,16383,32767,65535};
+
+// decode a jpeg huffman value from the bitstream
+stbi_inline static int stbi__jpeg_huff_decode(stbi__jpeg *j, stbi__huffman *h)
+{
+   unsigned int temp;
+   int c,k;
+
+   if (j->code_bits < 16) stbi__grow_buffer_unsafe(j);
+
+   // look at the top FAST_BITS and determine what symbol ID it is,
+   // if the code is <= FAST_BITS
+   c = (j->code_buffer >> (32 - FAST_BITS)) & ((1 << FAST_BITS)-1);
+   k = h->fast[c];
+   if (k < 255) {
+      int s = h->size[k];
+      if (s > j->code_bits)
+         return -1;
+      j->code_buffer <<= s;
+      j->code_bits -= s;
+      return h->values[k];
+   }
+
+   // naive test is to shift the code_buffer down so k bits are
+   // valid, then test against maxcode. To speed this up, we've
+   // preshifted maxcode left so that it has (16-k) 0s at the
+   // end; in other words, regardless of the number of bits, it
+   // wants to be compared against something shifted to have 16;
+   // that way we don't need to shift inside the loop.
+   temp = j->code_buffer >> 16;
+   for (k=FAST_BITS+1 ; ; ++k)
+      if (temp < h->maxcode[k])
+         break;
+   if (k == 17) {
+      // error! code not found
+      j->code_bits -= 16;
+      return -1;
+   }
+
+   if (k > j->code_bits)
+      return -1;
+
+   // convert the huffman code to the symbol id
+   c = ((j->code_buffer >> (32 - k)) & stbi__bmask[k]) + h->delta[k];
+   if(c < 0 || c >= 256) // symbol id out of bounds!
+       return -1;
+   STBI_ASSERT((((j->code_buffer) >> (32 - h->size[c])) & stbi__bmask[h->size[c]]) == h->code[c]);
+
+   // convert the id to a symbol
+   j->code_bits -= k;
+   j->code_buffer <<= k;
+   return h->values[c];
+}
+
+// bias[n] = (-1<<n) + 1
+static const int stbi__jbias[16] = {0,-1,-3,-7,-15,-31,-63,-127,-255,-511,-1023,-2047,-4095,-8191,-16383,-32767};
+
+// combined JPEG 'receive' and JPEG 'extend', since baseline
+// always extends everything it receives.
+stbi_inline static int stbi__extend_receive(stbi__jpeg *j, int n)
+{
+   unsigned int k;
+   int sgn;
+   if (j->code_bits < n) stbi__grow_buffer_unsafe(j);
+   if (j->code_bits < n) return 0; // ran out of bits from stream, return 0s intead of continuing
+
+   sgn = j->code_buffer >> 31; // sign bit always in MSB; 0 if MSB clear (positive), 1 if MSB set (negative)
+   k = stbi_lrot(j->code_buffer, n);
+   j->code_buffer = k & ~stbi__bmask[n];
+   k &= stbi__bmask[n];
+   j->code_bits -= n;
+   return k + (stbi__jbias[n] & (sgn - 1));
+}
+
+// get some unsigned bits
+stbi_inline static int stbi__jpeg_get_bits(stbi__jpeg *j, int n)
+{
+   unsigned int k;
+   if (j->code_bits < n) stbi__grow_buffer_unsafe(j);
+   if (j->code_bits < n) return 0; // ran out of bits from stream, return 0s intead of continuing
+   k = stbi_lrot(j->code_buffer, n);
+   j->code_buffer = k & ~stbi__bmask[n];
+   k &= stbi__bmask[n];
+   j->code_bits -= n;
+   return k;
+}
+
+stbi_inline static int stbi__jpeg_get_bit(stbi__jpeg *j)
+{
+   unsigned int k;
+   if (j->code_bits < 1) stbi__grow_buffer_unsafe(j);
+   if (j->code_bits < 1) return 0; // ran out of bits from stream, return 0s intead of continuing
+   k = j->code_buffer;
+   j->code_buffer <<= 1;
+   --j->code_bits;
+   return k & 0x80000000;
+}
+
+// given a value that's at position X in the zigzag stream,
+// where does it appear in the 8x8 matrix coded as row-major?
+static const stbi_uc stbi__jpeg_dezigzag[64+15] =
+{
+    0,  1,  8, 16,  9,  2,  3, 10,
+   17, 24, 32, 25, 18, 11,  4,  5,
+   12, 19, 26, 33, 40, 48, 41, 34,
+   27, 20, 13,  6,  7, 14, 21, 28,
+   35, 42, 49, 56, 57, 50, 43, 36,
+   29, 22, 15, 23, 30, 37, 44, 51,
+   58, 59, 52, 45, 38, 31, 39, 46,
+   53, 60, 61, 54, 47, 55, 62, 63,
+   // let corrupt input sample past end
+   63, 63, 63, 63, 63, 63, 63, 63,
+   63, 63, 63, 63, 63, 63, 63
+};
+
+// decode one 64-entry block--
+static int stbi__jpeg_decode_block(stbi__jpeg *j, short data[64], stbi__huffman *hdc, stbi__huffman *hac, stbi__int16 *fac, int b, stbi__uint16 *dequant)
+{
+   int diff,dc,k;
+   int t;
+
+   if (j->code_bits < 16) stbi__grow_buffer_unsafe(j);
+   t = stbi__jpeg_huff_decode(j, hdc);
+   if (t < 0 || t > 15) return stbi__err("bad huffman code","Corrupt JPEG");
+
+   // 0 all the ac values now so we can do it 32-bits at a time
+   memset(data,0,64*sizeof(data[0]));
+
+   diff = t ? stbi__extend_receive(j, t) : 0;
+   if (!stbi__addints_valid(j->img_comp[b].dc_pred, diff)) return stbi__err("bad delta","Corrupt JPEG");
+   dc = j->img_comp[b].dc_pred + diff;
+   j->img_comp[b].dc_pred = dc;
+   if (!stbi__mul2shorts_valid(dc, dequant[0])) return stbi__err("can't merge dc and ac", "Corrupt JPEG");
+   data[0] = (short) (dc * dequant[0]);
+
+   // decode AC components, see JPEG spec
+   k = 1;
+   do {
+      unsigned int zig;
+      int c,r,s;
+      if (j->code_bits < 16) stbi__grow_buffer_unsafe(j);
+      c = (j->code_buffer >> (32 - FAST_BITS)) & ((1 << FAST_BITS)-1);
+      r = fac[c];
+      if (r) { // fast-AC path
+         k += (r >> 4) & 15; // run
+         s = r & 15; // combined length
+         if (s > j->code_bits) return stbi__err("bad huffman code", "Combined length longer than code bits available");
+         j->code_buffer <<= s;
+         j->code_bits -= s;
+         // decode into unzigzag'd location
+         zig = stbi__jpeg_dezigzag[k++];
+         data[zig] = (short) ((r >> 8) * dequant[zig]);
+      } else {
+         int rs = stbi__jpeg_huff_decode(j, hac);
+         if (rs < 0) return stbi__err("bad huffman code","Corrupt JPEG");
+         s = rs & 15;
+         r = rs >> 4;
+         if (s == 0) {
+            if (rs != 0xf0) break; // end block
+            k += 16;
+         } else {
+            k += r;
+            // decode into unzigzag'd location
+            zig = stbi__jpeg_dezigzag[k++];
+            data[zig] = (short) (stbi__extend_receive(j,s) * dequant[zig]);
+         }
+      }
+   } while (k < 64);
+   return 1;
+}
+
+static int stbi__jpeg_decode_block_prog_dc(stbi__jpeg *j, short data[64], stbi__huffman *hdc, int b)
+{
+   int diff,dc;
+   int t;
+   if (j->spec_end != 0) return stbi__err("can't merge dc and ac", "Corrupt JPEG");
+
+   if (j->code_bits < 16) stbi__grow_buffer_unsafe(j);
+
+   if (j->succ_high == 0) {
+      // first scan for DC coefficient, must be first
+      memset(data,0,64*sizeof(data[0])); // 0 all the ac values now
+      t = stbi__jpeg_huff_decode(j, hdc);
+      if (t < 0 || t > 15) return stbi__err("can't merge dc and ac", "Corrupt JPEG");
+      diff = t ? stbi__extend_receive(j, t) : 0;
+
+      if (!stbi__addints_valid(j->img_comp[b].dc_pred, diff)) return stbi__err("bad delta", "Corrupt JPEG");
+      dc = j->img_comp[b].dc_pred + diff;
+      j->img_comp[b].dc_pred = dc;
+      if (!stbi__mul2shorts_valid(dc, 1 << j->succ_low)) return stbi__err("can't merge dc and ac", "Corrupt JPEG");
+      data[0] = (short) (dc * (1 << j->succ_low));
+   } else {
+      // refinement scan for DC coefficient
+      if (stbi__jpeg_get_bit(j))
+         data[0] += (short) (1 << j->succ_low);
+   }
+   return 1;
+}
+
+// @OPTIMIZE: store non-zigzagged during the decode passes,
+// and only de-zigzag when dequantizing
+static int stbi__jpeg_decode_block_prog_ac(stbi__jpeg *j, short data[64], stbi__huffman *hac, stbi__int16 *fac)
+{
+   int k;
+   if (j->spec_start == 0) return stbi__err("can't merge dc and ac", "Corrupt JPEG");
+
+   if (j->succ_high == 0) {
+      int shift = j->succ_low;
+
+      if (j->eob_run) {
+         --j->eob_run;
+         return 1;
+      }
+
+      k = j->spec_start;
+      do {
+         unsigned int zig;
+         int c,r,s;
+         if (j->code_bits < 16) stbi__grow_buffer_unsafe(j);
+         c = (j->code_buffer >> (32 - FAST_BITS)) & ((1 << FAST_BITS)-1);
+         r = fac[c];
+         if (r) { // fast-AC path
+            k += (r >> 4) & 15; // run
+            s = r & 15; // combined length
+            if (s > j->code_bits) return stbi__err("bad huffman code", "Combined length longer than code bits available");
+            j->code_buffer <<= s;
+            j->code_bits -= s;
+            zig = stbi__jpeg_dezigzag[k++];
+            data[zig] = (short) ((r >> 8) * (1 << shift));
+         } else {
+            int rs = stbi__jpeg_huff_decode(j, hac);
+            if (rs < 0) return stbi__err("bad huffman code","Corrupt JPEG");
+            s = rs & 15;
+            r = rs >> 4;
+            if (s == 0) {
+               if (r < 15) {
+                  j->eob_run = (1 << r);
+                  if (r)
+                     j->eob_run += stbi__jpeg_get_bits(j, r);
+                  --j->eob_run;
+                  break;
+               }
+               k += 16;
+            } else {
+               k += r;
+               zig = stbi__jpeg_dezigzag[k++];
+               data[zig] = (short) (stbi__extend_receive(j,s) * (1 << shift));
+            }
+         }
+      } while (k <= j->spec_end);
+   } else {
+      // refinement scan for these AC coefficients
+
+      short bit = (short) (1 << j->succ_low);
+
+      if (j->eob_run) {
+         --j->eob_run;
+         for (k = j->spec_start; k <= j->spec_end; ++k) {
+            short *p = &data[stbi__jpeg_dezigzag[k]];
+            if (*p != 0)
+               if (stbi__jpeg_get_bit(j))
+                  if ((*p & bit)==0) {
+                     if (*p > 0)
+                        *p += bit;
+                     else
+                        *p -= bit;
+                  }
+         }
+      } else {
+         k = j->spec_start;
+         do {
+            int r,s;
+            int rs = stbi__jpeg_huff_decode(j, hac); // @OPTIMIZE see if we can use the fast path here, advance-by-r is so slow, eh
+            if (rs < 0) return stbi__err("bad huffman code","Corrupt JPEG");
+            s = rs & 15;
+            r = rs >> 4;
+            if (s == 0) {
+               if (r < 15) {
+                  j->eob_run = (1 << r) - 1;
+                  if (r)
+                     j->eob_run += stbi__jpeg_get_bits(j, r);
+                  r = 64; // force end of block
+               } else {
+                  // r=15 s=0 should write 16 0s, so we just do
+                  // a run of 15 0s and then write s (which is 0),
+                  // so we don't have to do anything special here
+               }
+            } else {
+               if (s != 1) return stbi__err("bad huffman code", "Corrupt JPEG");
+               // sign bit
+               if (stbi__jpeg_get_bit(j))
+                  s = bit;
+               else
+                  s = -bit;
+            }
+
+            // advance by r
+            while (k <= j->spec_end) {
+               short *p = &data[stbi__jpeg_dezigzag[k++]];
+               if (*p != 0) {
+                  if (stbi__jpeg_get_bit(j))
+                     if ((*p & bit)==0) {
+                        if (*p > 0)
+                           *p += bit;
+                        else
+                           *p -= bit;
+                     }
+               } else {
+                  if (r == 0) {
+                     *p = (short) s;
+                     break;
+                  }
+                  --r;
+               }
+            }
+         } while (k <= j->spec_end);
+      }
+   }
+   return 1;
+}
+
+// take a -128..127 value and stbi__clamp it and convert to 0..255
+stbi_inline static stbi_uc stbi__clamp(int x)
+{
+   // trick to use a single test to catch both cases
+   if ((unsigned int) x > 255) {
+      if (x < 0) return 0;
+      if (x > 255) return 255;
+   }
+   return (stbi_uc) x;
+}
+
+#define stbi__f2f(x)  ((int) (((x) * 4096 + 0.5)))
+#define stbi__fsh(x)  ((x) * 4096)
+
+// derived from jidctint -- DCT_ISLOW
+#define STBI__IDCT_1D(s0,s1,s2,s3,s4,s5,s6,s7) \
+   int t0,t1,t2,t3,p1,p2,p3,p4,p5,x0,x1,x2,x3; \
+   p2 = s2;                                    \
+   p3 = s6;                                    \
+   p1 = (p2+p3) * stbi__f2f(0.5411961f);       \
+   t2 = p1 + p3*stbi__f2f(-1.847759065f);      \
+   t3 = p1 + p2*stbi__f2f( 0.765366865f);      \
+   p2 = s0;                                    \
+   p3 = s4;                                    \
+   t0 = stbi__fsh(p2+p3);                      \
+   t1 = stbi__fsh(p2-p3);                      \
+   x0 = t0+t3;                                 \
+   x3 = t0-t3;                                 \
+   x1 = t1+t2;                                 \
+   x2 = t1-t2;                                 \
+   t0 = s7;                                    \
+   t1 = s5;                                    \
+   t2 = s3;                                    \
+   t3 = s1;                                    \
+   p3 = t0+t2;                                 \
+   p4 = t1+t3;                                 \
+   p1 = t0+t3;                                 \
+   p2 = t1+t2;                                 \
+   p5 = (p3+p4)*stbi__f2f( 1.175875602f);      \
+   t0 = t0*stbi__f2f( 0.298631336f);           \
+   t1 = t1*stbi__f2f( 2.053119869f);           \
+   t2 = t2*stbi__f2f( 3.072711026f);           \
+   t3 = t3*stbi__f2f( 1.501321110f);           \
+   p1 = p5 + p1*stbi__f2f(-0.899976223f);      \
+   p2 = p5 + p2*stbi__f2f(-2.562915447f);      \
+   p3 = p3*stbi__f2f(-1.961570560f);           \
+   p4 = p4*stbi__f2f(-0.390180644f);           \
+   t3 += p1+p4;                                \
+   t2 += p2+p3;                                \
+   t1 += p2+p4;                                \
+   t0 += p1+p3;
+
+static void stbi__idct_block(stbi_uc *out, int out_stride, short data[64])
+{
+   int i,val[64],*v=val;
+   stbi_uc *o;
+   short *d = data;
+
+   // columns
+   for (i=0; i < 8; ++i,++d, ++v) {
+      // if all zeroes, shortcut -- this avoids dequantizing 0s and IDCTing
+      if (d[ 8]==0 && d[16]==0 && d[24]==0 && d[32]==0
+           && d[40]==0 && d[48]==0 && d[56]==0) {
+         //    no shortcut                 0     seconds
+         //    (1|2|3|4|5|6|7)==0          0     seconds
+         //    all separate               -0.047 seconds
+         //    1 && 2|3 && 4|5 && 6|7:    -0.047 seconds
+         int dcterm = d[0]*4;
+         v[0] = v[8] = v[16] = v[24] = v[32] = v[40] = v[48] = v[56] = dcterm;
+      } else {
+         STBI__IDCT_1D(d[ 0],d[ 8],d[16],d[24],d[32],d[40],d[48],d[56])
+         // constants scaled things up by 1<<12; let's bring them back
+         // down, but keep 2 extra bits of precision
+         x0 += 512; x1 += 512; x2 += 512; x3 += 512;
+         v[ 0] = (x0+t3) >> 10;
+         v[56] = (x0-t3) >> 10;
+         v[ 8] = (x1+t2) >> 10;
+         v[48] = (x1-t2) >> 10;
+         v[16] = (x2+t1) >> 10;
+         v[40] = (x2-t1) >> 10;
+         v[24] = (x3+t0) >> 10;
+         v[32] = (x3-t0) >> 10;
+      }
+   }
+
+   for (i=0, v=val, o=out; i < 8; ++i,v+=8,o+=out_stride) {
+      // no fast case since the first 1D IDCT spread components out
+      STBI__IDCT_1D(v[0],v[1],v[2],v[3],v[4],v[5],v[6],v[7])
+      // constants scaled things up by 1<<12, plus we had 1<<2 from first
+      // loop, plus horizontal and vertical each scale by sqrt(8) so together
+      // we've got an extra 1<<3, so 1<<17 total we need to remove.
+      // so we want to round that, which means adding 0.5 * 1<<17,
+      // aka 65536. Also, we'll end up with -128 to 127 that we want
+      // to encode as 0..255 by adding 128, so we'll add that before the shift
+      x0 += 65536 + (128<<17);
+      x1 += 65536 + (128<<17);
+      x2 += 65536 + (128<<17);
+      x3 += 65536 + (128<<17);
+      // tried computing the shifts into temps, or'ing the temps to see
+      // if any were out of range, but that was slower
+      o[0] = stbi__clamp((x0+t3) >> 17);
+      o[7] = stbi__clamp((x0-t3) >> 17);
+      o[1] = stbi__clamp((x1+t2) >> 17);
+      o[6] = stbi__clamp((x1-t2) >> 17);
+      o[2] = stbi__clamp((x2+t1) >> 17);
+      o[5] = stbi__clamp((x2-t1) >> 17);
+      o[3] = stbi__clamp((x3+t0) >> 17);
+      o[4] = stbi__clamp((x3-t0) >> 17);
+   }
+}
+
+#ifdef STBI_SSE2
+// sse2 integer IDCT. not the fastest possible implementation but it
+// produces bit-identical results to the generic C version so it's
+// fully "transparent".
+static void stbi__idct_simd(stbi_uc *out, int out_stride, short data[64])
+{
+   // This is constructed to match our regular (generic) integer IDCT exactly.
+   __m128i row0, row1, row2, row3, row4, row5, row6, row7;
+   __m128i tmp;
+
+   // dot product constant: even elems=x, odd elems=y
+   #define dct_const(x,y)  _mm_setr_epi16((x),(y),(x),(y),(x),(y),(x),(y))
+
+   // out(0) = c0[even]*x + c0[odd]*y   (c0, x, y 16-bit, out 32-bit)
+   // out(1) = c1[even]*x + c1[odd]*y
+   #define dct_rot(out0,out1, x,y,c0,c1) \
+      __m128i c0##lo = _mm_unpacklo_epi16((x),(y)); \
+      __m128i c0##hi = _mm_unpackhi_epi16((x),(y)); \
+      __m128i out0##_l = _mm_madd_epi16(c0##lo, c0); \
+      __m128i out0##_h = _mm_madd_epi16(c0##hi, c0); \
+      __m128i out1##_l = _mm_madd_epi16(c0##lo, c1); \
+      __m128i out1##_h = _mm_madd_epi16(c0##hi, c1)
+
+   // out = in << 12  (in 16-bit, out 32-bit)
+   #define dct_widen(out, in) \
+      __m128i out##_l = _mm_srai_epi32(_mm_unpacklo_epi16(_mm_setzero_si128(), (in)), 4); \
+      __m128i out##_h = _mm_srai_epi32(_mm_unpackhi_epi16(_mm_setzero_si128(), (in)), 4)
+
+   // wide add
+   #define dct_wadd(out, a, b) \
+      __m128i out##_l = _mm_add_epi32(a##_l, b##_l); \
+      __m128i out##_h = _mm_add_epi32(a##_h, b##_h)
+
+   // wide sub
+   #define dct_wsub(out, a, b) \
+      __m128i out##_l = _mm_sub_epi32(a##_l, b##_l); \
+      __m128i out##_h = _mm_sub_epi32(a##_h, b##_h)
+
+   // butterfly a/b, add bias, then shift by "s" and pack
+   #define dct_bfly32o(out0, out1, a,b,bias,s) \
+      { \
+         __m128i abiased_l = _mm_add_epi32(a##_l, bias); \
+         __m128i abiased_h = _mm_add_epi32(a##_h, bias); \
+         dct_wadd(sum, abiased, b); \
+         dct_wsub(dif, abiased, b); \
+         out0 = _mm_packs_epi32(_mm_srai_epi32(sum_l, s), _mm_srai_epi32(sum_h, s)); \
+         out1 = _mm_packs_epi32(_mm_srai_epi32(dif_l, s), _mm_srai_epi32(dif_h, s)); \
+      }
+
+   // 8-bit interleave step (for transposes)
+   #define dct_interleave8(a, b) \
+      tmp = a; \
+      a = _mm_unpacklo_epi8(a, b); \
+      b = _mm_unpackhi_epi8(tmp, b)
+
+   // 16-bit interleave step (for transposes)
+   #define dct_interleave16(a, b) \
+      tmp = a; \
+      a = _mm_unpacklo_epi16(a, b); \
+      b = _mm_unpackhi_epi16(tmp, b)
+
+   #define dct_pass(bias,shift) \
+      { \
+         /* even part */ \
+         dct_rot(t2e,t3e, row2,row6, rot0_0,rot0_1); \
+         __m128i sum04 = _mm_add_epi16(row0, row4); \
+         __m128i dif04 = _mm_sub_epi16(row0, row4); \
+         dct_widen(t0e, sum04); \
+         dct_widen(t1e, dif04); \
+         dct_wadd(x0, t0e, t3e); \
+         dct_wsub(x3, t0e, t3e); \
+         dct_wadd(x1, t1e, t2e); \
+         dct_wsub(x2, t1e, t2e); \
+         /* odd part */ \
+         dct_rot(y0o,y2o, row7,row3, rot2_0,rot2_1); \
+         dct_rot(y1o,y3o, row5,row1, rot3_0,rot3_1); \
+         __m128i sum17 = _mm_add_epi16(row1, row7); \
+         __m128i sum35 = _mm_add_epi16(row3, row5); \
+         dct_rot(y4o,y5o, sum17,sum35, rot1_0,rot1_1); \
+         dct_wadd(x4, y0o, y4o); \
+         dct_wadd(x5, y1o, y5o); \
+         dct_wadd(x6, y2o, y5o); \
+         dct_wadd(x7, y3o, y4o); \
+         dct_bfly32o(row0,row7, x0,x7,bias,shift); \
+         dct_bfly32o(row1,row6, x1,x6,bias,shift); \
+         dct_bfly32o(row2,row5, x2,x5,bias,shift); \
+         dct_bfly32o(row3,row4, x3,x4,bias,shift); \
+      }
+
+   __m128i rot0_0 = dct_const(stbi__f2f(0.5411961f), stbi__f2f(0.5411961f) + stbi__f2f(-1.847759065f));
+   __m128i rot0_1 = dct_const(stbi__f2f(0.5411961f) + stbi__f2f( 0.765366865f), stbi__f2f(0.5411961f));
+   __m128i rot1_0 = dct_const(stbi__f2f(1.175875602f) + stbi__f2f(-0.899976223f), stbi__f2f(1.175875602f));
+   __m128i rot1_1 = dct_const(stbi__f2f(1.175875602f), stbi__f2f(1.175875602f) + stbi__f2f(-2.562915447f));
+   __m128i rot2_0 = dct_const(stbi__f2f(-1.961570560f) + stbi__f2f( 0.298631336f), stbi__f2f(-1.961570560f));
+   __m128i rot2_1 = dct_const(stbi__f2f(-1.961570560f), stbi__f2f(-1.961570560f) + stbi__f2f( 3.072711026f));
+   __m128i rot3_0 = dct_const(stbi__f2f(-0.390180644f) + stbi__f2f( 2.053119869f), stbi__f2f(-0.390180644f));
+   __m128i rot3_1 = dct_const(stbi__f2f(-0.390180644f), stbi__f2f(-0.390180644f) + stbi__f2f( 1.501321110f));
+
+   // rounding biases in column/row passes, see stbi__idct_block for explanation.
+   __m128i bias_0 = _mm_set1_epi32(512);
+   __m128i bias_1 = _mm_set1_epi32(65536 + (128<<17));
+
+   // load
+   row0 = _mm_load_si128((const __m128i *) (data + 0*8));
+   row1 = _mm_load_si128((const __m128i *) (data + 1*8));
+   row2 = _mm_load_si128((const __m128i *) (data + 2*8));
+   row3 = _mm_load_si128((const __m128i *) (data + 3*8));
+   row4 = _mm_load_si128((const __m128i *) (data + 4*8));
+   row5 = _mm_load_si128((const __m128i *) (data + 5*8));
+   row6 = _mm_load_si128((const __m128i *) (data + 6*8));
+   row7 = _mm_load_si128((const __m128i *) (data + 7*8));
+
+   // column pass
+   dct_pass(bias_0, 10);
+
+   {
+      // 16bit 8x8 transpose pass 1
+      dct_interleave16(row0, row4);
+      dct_interleave16(row1, row5);
+      dct_interleave16(row2, row6);
+      dct_interleave16(row3, row7);
+
+      // transpose pass 2
+      dct_interleave16(row0, row2);
+      dct_interleave16(row1, row3);
+      dct_interleave16(row4, row6);
+      dct_interleave16(row5, row7);
+
+      // transpose pass 3
+      dct_interleave16(row0, row1);
+      dct_interleave16(row2, row3);
+      dct_interleave16(row4, row5);
+      dct_interleave16(row6, row7);
+   }
+
+   // row pass
+   dct_pass(bias_1, 17);
+
+   {
+      // pack
+      __m128i p0 = _mm_packus_epi16(row0, row1); // a0a1a2a3...a7b0b1b2b3...b7
+      __m128i p1 = _mm_packus_epi16(row2, row3);
+      __m128i p2 = _mm_packus_epi16(row4, row5);
+      __m128i p3 = _mm_packus_epi16(row6, row7);
+
+      // 8bit 8x8 transpose pass 1
+      dct_interleave8(p0, p2); // a0e0a1e1...
+      dct_interleave8(p1, p3); // c0g0c1g1...
+
+      // transpose pass 2
+      dct_interleave8(p0, p1); // a0c0e0g0...
+      dct_interleave8(p2, p3); // b0d0f0h0...
+
+      // transpose pass 3
+      dct_interleave8(p0, p2); // a0b0c0d0...
+      dct_interleave8(p1, p3); // a4b4c4d4...
+
+      // store
+      _mm_storel_epi64((__m128i *) out, p0); out += out_stride;
+      _mm_storel_epi64((__m128i *) out, _mm_shuffle_epi32(p0, 0x4e)); out += out_stride;
+      _mm_storel_epi64((__m128i *) out, p2); out += out_stride;
+      _mm_storel_epi64((__m128i *) out, _mm_shuffle_epi32(p2, 0x4e)); out += out_stride;
+      _mm_storel_epi64((__m128i *) out, p1); out += out_stride;
+      _mm_storel_epi64((__m128i *) out, _mm_shuffle_epi32(p1, 0x4e)); out += out_stride;
+      _mm_storel_epi64((__m128i *) out, p3); out += out_stride;
+      _mm_storel_epi64((__m128i *) out, _mm_shuffle_epi32(p3, 0x4e));
+   }
+
+#undef dct_const
+#undef dct_rot
+#undef dct_widen
+#undef dct_wadd
+#undef dct_wsub
+#undef dct_bfly32o
+#undef dct_interleave8
+#undef dct_interleave16
+#undef dct_pass
+}
+
+#endif // STBI_SSE2
+
+#ifdef STBI_NEON
+
+// NEON integer IDCT. should produce bit-identical
+// results to the generic C version.
+static void stbi__idct_simd(stbi_uc *out, int out_stride, short data[64])
+{
+   int16x8_t row0, row1, row2, row3, row4, row5, row6, row7;
+
+   int16x4_t rot0_0 = vdup_n_s16(stbi__f2f(0.5411961f));
+   int16x4_t rot0_1 = vdup_n_s16(stbi__f2f(-1.847759065f));
+   int16x4_t rot0_2 = vdup_n_s16(stbi__f2f( 0.765366865f));
+   int16x4_t rot1_0 = vdup_n_s16(stbi__f2f( 1.175875602f));
+   int16x4_t rot1_1 = vdup_n_s16(stbi__f2f(-0.899976223f));
+   int16x4_t rot1_2 = vdup_n_s16(stbi__f2f(-2.562915447f));
+   int16x4_t rot2_0 = vdup_n_s16(stbi__f2f(-1.961570560f));
+   int16x4_t rot2_1 = vdup_n_s16(stbi__f2f(-0.390180644f));
+   int16x4_t rot3_0 = vdup_n_s16(stbi__f2f( 0.298631336f));
+   int16x4_t rot3_1 = vdup_n_s16(stbi__f2f( 2.053119869f));
+   int16x4_t rot3_2 = vdup_n_s16(stbi__f2f( 3.072711026f));
+   int16x4_t rot3_3 = vdup_n_s16(stbi__f2f( 1.501321110f));
+
+#define dct_long_mul(out, inq, coeff) \
+   int32x4_t out##_l = vmull_s16(vget_low_s16(inq), coeff); \
+   int32x4_t out##_h = vmull_s16(vget_high_s16(inq), coeff)
+
+#define dct_long_mac(out, acc, inq, coeff) \
+   int32x4_t out##_l = vmlal_s16(acc##_l, vget_low_s16(inq), coeff); \
+   int32x4_t out##_h = vmlal_s16(acc##_h, vget_high_s16(inq), coeff)
+
+#define dct_widen(out, inq) \
+   int32x4_t out##_l = vshll_n_s16(vget_low_s16(inq), 12); \
+   int32x4_t out##_h = vshll_n_s16(vget_high_s16(inq), 12)
+
+// wide add
+#define dct_wadd(out, a, b) \
+   int32x4_t out##_l = vaddq_s32(a##_l, b##_l); \
+   int32x4_t out##_h = vaddq_s32(a##_h, b##_h)
+
+// wide sub
+#define dct_wsub(out, a, b) \
+   int32x4_t out##_l = vsubq_s32(a##_l, b##_l); \
+   int32x4_t out##_h = vsubq_s32(a##_h, b##_h)
+
+// butterfly a/b, then shift using "shiftop" by "s" and pack
+#define dct_bfly32o(out0,out1, a,b,shiftop,s) \
+   { \
+      dct_wadd(sum, a, b); \
+      dct_wsub(dif, a, b); \
+      out0 = vcombine_s16(shiftop(sum_l, s), shiftop(sum_h, s)); \
+      out1 = vcombine_s16(shiftop(dif_l, s), shiftop(dif_h, s)); \
+   }
+
+#define dct_pass(shiftop, shift) \
+   { \
+      /* even part */ \
+      int16x8_t sum26 = vaddq_s16(row2, row6); \
+      dct_long_mul(p1e, sum26, rot0_0); \
+      dct_long_mac(t2e, p1e, row6, rot0_1); \
+      dct_long_mac(t3e, p1e, row2, rot0_2); \
+      int16x8_t sum04 = vaddq_s16(row0, row4); \
+      int16x8_t dif04 = vsubq_s16(row0, row4); \
+      dct_widen(t0e, sum04); \
+      dct_widen(t1e, dif04); \
+      dct_wadd(x0, t0e, t3e); \
+      dct_wsub(x3, t0e, t3e); \
+      dct_wadd(x1, t1e, t2e); \
+      dct_wsub(x2, t1e, t2e); \
+      /* odd part */ \
+      int16x8_t sum15 = vaddq_s16(row1, row5); \
+      int16x8_t sum17 = vaddq_s16(row1, row7); \
+      int16x8_t sum35 = vaddq_s16(row3, row5); \
+      int16x8_t sum37 = vaddq_s16(row3, row7); \
+      int16x8_t sumodd = vaddq_s16(sum17, sum35); \
+      dct_long_mul(p5o, sumodd, rot1_0); \
+      dct_long_mac(p1o, p5o, sum17, rot1_1); \
+      dct_long_mac(p2o, p5o, sum35, rot1_2); \
+      dct_long_mul(p3o, sum37, rot2_0); \
+      dct_long_mul(p4o, sum15, rot2_1); \
+      dct_wadd(sump13o, p1o, p3o); \
+      dct_wadd(sump24o, p2o, p4o); \
+      dct_wadd(sump23o, p2o, p3o); \
+      dct_wadd(sump14o, p1o, p4o); \
+      dct_long_mac(x4, sump13o, row7, rot3_0); \
+      dct_long_mac(x5, sump24o, row5, rot3_1); \
+      dct_long_mac(x6, sump23o, row3, rot3_2); \
+      dct_long_mac(x7, sump14o, row1, rot3_3); \
+      dct_bfly32o(row0,row7, x0,x7,shiftop,shift); \
+      dct_bfly32o(row1,row6, x1,x6,shiftop,shift); \
+      dct_bfly32o(row2,row5, x2,x5,shiftop,shift); \
+      dct_bfly32o(row3,row4, x3,x4,shiftop,shift); \
+   }
+
+   // load
+   row0 = vld1q_s16(data + 0*8);
+   row1 = vld1q_s16(data + 1*8);
+   row2 = vld1q_s16(data + 2*8);
+   row3 = vld1q_s16(data + 3*8);
+   row4 = vld1q_s16(data + 4*8);
+   row5 = vld1q_s16(data + 5*8);
+   row6 = vld1q_s16(data + 6*8);
+   row7 = vld1q_s16(data + 7*8);
+
+   // add DC bias
+   row0 = vaddq_s16(row0, vsetq_lane_s16(1024, vdupq_n_s16(0), 0));
+
+   // column pass
+   dct_pass(vrshrn_n_s32, 10);
+
+   // 16bit 8x8 transpose
+   {
+// these three map to a single VTRN.16, VTRN.32, and VSWP, respectively.
+// whether compilers actually get this is another story, sadly.
+#define dct_trn16(x, y) { int16x8x2_t t = vtrnq_s16(x, y); x = t.val[0]; y = t.val[1]; }
+#define dct_trn32(x, y) { int32x4x2_t t = vtrnq_s32(vreinterpretq_s32_s16(x), vreinterpretq_s32_s16(y)); x = vreinterpretq_s16_s32(t.val[0]); y = vreinterpretq_s16_s32(t.val[1]); }
+#define dct_trn64(x, y) { int16x8_t x0 = x; int16x8_t y0 = y; x = vcombine_s16(vget_low_s16(x0), vget_low_s16(y0)); y = vcombine_s16(vget_high_s16(x0), vget_high_s16(y0)); }
+
+      // pass 1
+      dct_trn16(row0, row1); // a0b0a2b2a4b4a6b6
+      dct_trn16(row2, row3);
+      dct_trn16(row4, row5);
+      dct_trn16(row6, row7);
+
+      // pass 2
+      dct_trn32(row0, row2); // a0b0c0d0a4b4c4d4
+      dct_trn32(row1, row3);
+      dct_trn32(row4, row6);
+      dct_trn32(row5, row7);
+
+      // pass 3
+      dct_trn64(row0, row4); // a0b0c0d0e0f0g0h0
+      dct_trn64(row1, row5);
+      dct_trn64(row2, row6);
+      dct_trn64(row3, row7);
+
+#undef dct_trn16
+#undef dct_trn32
+#undef dct_trn64
+   }
+
+   // row pass
+   // vrshrn_n_s32 only supports shifts up to 16, we need
+   // 17. so do a non-rounding shift of 16 first then follow
+   // up with a rounding shift by 1.
+   dct_pass(vshrn_n_s32, 16);
+
+   {
+      // pack and round
+      uint8x8_t p0 = vqrshrun_n_s16(row0, 1);
+      uint8x8_t p1 = vqrshrun_n_s16(row1, 1);
+      uint8x8_t p2 = vqrshrun_n_s16(row2, 1);
+      uint8x8_t p3 = vqrshrun_n_s16(row3, 1);
+      uint8x8_t p4 = vqrshrun_n_s16(row4, 1);
+      uint8x8_t p5 = vqrshrun_n_s16(row5, 1);
+      uint8x8_t p6 = vqrshrun_n_s16(row6, 1);
+      uint8x8_t p7 = vqrshrun_n_s16(row7, 1);
+
+      // again, these can translate into one instruction, but often don't.
+#define dct_trn8_8(x, y) { uint8x8x2_t t = vtrn_u8(x, y); x = t.val[0]; y = t.val[1]; }
+#define dct_trn8_16(x, y) { uint16x4x2_t t = vtrn_u16(vreinterpret_u16_u8(x), vreinterpret_u16_u8(y)); x = vreinterpret_u8_u16(t.val[0]); y = vreinterpret_u8_u16(t.val[1]); }
+#define dct_trn8_32(x, y) { uint32x2x2_t t = vtrn_u32(vreinterpret_u32_u8(x), vreinterpret_u32_u8(y)); x = vreinterpret_u8_u32(t.val[0]); y = vreinterpret_u8_u32(t.val[1]); }
+
+      // sadly can't use interleaved stores here since we only write
+      // 8 bytes to each scan line!
+
+      // 8x8 8-bit transpose pass 1
+      dct_trn8_8(p0, p1);
+      dct_trn8_8(p2, p3);
+      dct_trn8_8(p4, p5);
+      dct_trn8_8(p6, p7);
+
+      // pass 2
+      dct_trn8_16(p0, p2);
+      dct_trn8_16(p1, p3);
+      dct_trn8_16(p4, p6);
+      dct_trn8_16(p5, p7);
+
+      // pass 3
+      dct_trn8_32(p0, p4);
+      dct_trn8_32(p1, p5);
+      dct_trn8_32(p2, p6);
+      dct_trn8_32(p3, p7);
+
+      // store
+      vst1_u8(out, p0); out += out_stride;
+      vst1_u8(out, p1); out += out_stride;
+      vst1_u8(out, p2); out += out_stride;
+      vst1_u8(out, p3); out += out_stride;
+      vst1_u8(out, p4); out += out_stride;
+      vst1_u8(out, p5); out += out_stride;
+      vst1_u8(out, p6); out += out_stride;
+      vst1_u8(out, p7);
+
+#undef dct_trn8_8
+#undef dct_trn8_16
+#undef dct_trn8_32
+   }
+
+#undef dct_long_mul
+#undef dct_long_mac
+#undef dct_widen
+#undef dct_wadd
+#undef dct_wsub
+#undef dct_bfly32o
+#undef dct_pass
+}
+
+#endif // STBI_NEON
+
+#define STBI__MARKER_none  0xff
+// if there's a pending marker from the entropy stream, return that
+// otherwise, fetch from the stream and get a marker. if there's no
+// marker, return 0xff, which is never a valid marker value
+static stbi_uc stbi__get_marker(stbi__jpeg *j)
+{
+   stbi_uc x;
+   if (j->marker != STBI__MARKER_none) { x = j->marker; j->marker = STBI__MARKER_none; return x; }
+   x = stbi__get8(j->s);
+   if (x != 0xff) return STBI__MARKER_none;
+   while (x == 0xff)
+      x = stbi__get8(j->s); // consume repeated 0xff fill bytes
+   return x;
+}
+
+// in each scan, we'll have scan_n components, and the order
+// of the components is specified by order[]
+#define STBI__RESTART(x)     ((x) >= 0xd0 && (x) <= 0xd7)
+
+// after a restart interval, stbi__jpeg_reset the entropy decoder and
+// the dc prediction
+static void stbi__jpeg_reset(stbi__jpeg *j)
+{
+   j->code_bits = 0;
+   j->code_buffer = 0;
+   j->nomore = 0;
+   j->img_comp[0].dc_pred = j->img_comp[1].dc_pred = j->img_comp[2].dc_pred = j->img_comp[3].dc_pred = 0;
+   j->marker = STBI__MARKER_none;
+   j->todo = j->restart_interval ? j->restart_interval : 0x7fffffff;
+   j->eob_run = 0;
+   // no more than 1<<31 MCUs if no restart_interal? that's plenty safe,
+   // since we don't even allow 1<<30 pixels
+}
+
+static int stbi__parse_entropy_coded_data(stbi__jpeg *z)
+{
+   stbi__jpeg_reset(z);
+   if (!z->progressive) {
+      if (z->scan_n == 1) {
+         int i,j;
+         STBI_SIMD_ALIGN(short, data[64]);
+         int n = z->order[0];
+         // non-interleaved data, we just need to process one block at a time,
+         // in trivial scanline order
+         // number of blocks to do just depends on how many actual "pixels" this
+         // component has, independent of interleaved MCU blocking and such
+         int w = (z->img_comp[n].x+7) >> 3;
+         int h = (z->img_comp[n].y+7) >> 3;
+         for (j=0; j < h; ++j) {
+            for (i=0; i < w; ++i) {
+               int ha = z->img_comp[n].ha;
+               if (!stbi__jpeg_decode_block(z, data, z->huff_dc+z->img_comp[n].hd, z->huff_ac+ha, z->fast_ac[ha], n, z->dequant[z->img_comp[n].tq])) return 0;
+               z->idct_block_kernel(z->img_comp[n].data+z->img_comp[n].w2*j*8+i*8, z->img_comp[n].w2, data);
+               // every data block is an MCU, so countdown the restart interval
+               if (--z->todo <= 0) {
+                  if (z->code_bits < 24) stbi__grow_buffer_unsafe(z);
+                  // if it's NOT a restart, then just bail, so we get corrupt data
+                  // rather than no data
+                  if (!STBI__RESTART(z->marker)) return 1;
+                  stbi__jpeg_reset(z);
+               }
+            }
+         }
+         return 1;
+      } else { // interleaved
+         int i,j,k,x,y;
+         STBI_SIMD_ALIGN(short, data[64]);
+         for (j=0; j < z->img_mcu_y; ++j) {
+            for (i=0; i < z->img_mcu_x; ++i) {
+               // scan an interleaved mcu... process scan_n components in order
+               for (k=0; k < z->scan_n; ++k) {
+                  int n = z->order[k];
+                  // scan out an mcu's worth of this component; that's just determined
+                  // by the basic H and V specified for the component
+                  for (y=0; y < z->img_comp[n].v; ++y) {
+                     for (x=0; x < z->img_comp[n].h; ++x) {
+                        int x2 = (i*z->img_comp[n].h + x)*8;
+                        int y2 = (j*z->img_comp[n].v + y)*8;
+                        int ha = z->img_comp[n].ha;
+                        if (!stbi__jpeg_decode_block(z, data, z->huff_dc+z->img_comp[n].hd, z->huff_ac+ha, z->fast_ac[ha], n, z->dequant[z->img_comp[n].tq])) return 0;
+                        z->idct_block_kernel(z->img_comp[n].data+z->img_comp[n].w2*y2+x2, z->img_comp[n].w2, data);
+                     }
+                  }
+               }
+               // after all interleaved components, that's an interleaved MCU,
+               // so now count down the restart interval
+               if (--z->todo <= 0) {
+                  if (z->code_bits < 24) stbi__grow_buffer_unsafe(z);
+                  if (!STBI__RESTART(z->marker)) return 1;
+                  stbi__jpeg_reset(z);
+               }
+            }
+         }
+         return 1;
+      }
+   } else {
+      if (z->scan_n == 1) {
+         int i,j;
+         int n = z->order[0];
+         // non-interleaved data, we just need to process one block at a time,
+         // in trivial scanline order
+         // number of blocks to do just depends on how many actual "pixels" this
+         // component has, independent of interleaved MCU blocking and such
+         int w = (z->img_comp[n].x+7) >> 3;
+         int h = (z->img_comp[n].y+7) >> 3;
+         for (j=0; j < h; ++j) {
+            for (i=0; i < w; ++i) {
+               short *data = z->img_comp[n].coeff + 64 * (i + j * z->img_comp[n].coeff_w);
+               if (z->spec_start == 0) {
+                  if (!stbi__jpeg_decode_block_prog_dc(z, data, &z->huff_dc[z->img_comp[n].hd], n))
+                     return 0;
+               } else {
+                  int ha = z->img_comp[n].ha;
+                  if (!stbi__jpeg_decode_block_prog_ac(z, data, &z->huff_ac[ha], z->fast_ac[ha]))
+                     return 0;
+               }
+               // every data block is an MCU, so countdown the restart interval
+               if (--z->todo <= 0) {
+                  if (z->code_bits < 24) stbi__grow_buffer_unsafe(z);
+                  if (!STBI__RESTART(z->marker)) return 1;
+                  stbi__jpeg_reset(z);
+               }
+            }
+         }
+         return 1;
+      } else { // interleaved
+         int i,j,k,x,y;
+         for (j=0; j < z->img_mcu_y; ++j) {
+            for (i=0; i < z->img_mcu_x; ++i) {
+               // scan an interleaved mcu... process scan_n components in order
+               for (k=0; k < z->scan_n; ++k) {
+                  int n = z->order[k];
+                  // scan out an mcu's worth of this component; that's just determined
+                  // by the basic H and V specified for the component
+                  for (y=0; y < z->img_comp[n].v; ++y) {
+                     for (x=0; x < z->img_comp[n].h; ++x) {
+                        int x2 = (i*z->img_comp[n].h + x);
+                        int y2 = (j*z->img_comp[n].v + y);
+                        short *data = z->img_comp[n].coeff + 64 * (x2 + y2 * z->img_comp[n].coeff_w);
+                        if (!stbi__jpeg_decode_block_prog_dc(z, data, &z->huff_dc[z->img_comp[n].hd], n))
+                           return 0;
+                     }
+                  }
+               }
+               // after all interleaved components, that's an interleaved MCU,
+               // so now count down the restart interval
+               if (--z->todo <= 0) {
+                  if (z->code_bits < 24) stbi__grow_buffer_unsafe(z);
+                  if (!STBI__RESTART(z->marker)) return 1;
+                  stbi__jpeg_reset(z);
+               }
+            }
+         }
+         return 1;
+      }
+   }
+}
+
+static void stbi__jpeg_dequantize(short *data, stbi__uint16 *dequant)
+{
+   int i;
+   for (i=0; i < 64; ++i)
+      data[i] *= dequant[i];
+}
+
+static void stbi__jpeg_finish(stbi__jpeg *z)
+{
+   if (z->progressive) {
+      // dequantize and idct the data
+      int i,j,n;
+      for (n=0; n < z->s->img_n; ++n) {
+         int w = (z->img_comp[n].x+7) >> 3;
+         int h = (z->img_comp[n].y+7) >> 3;
+         for (j=0; j < h; ++j) {
+            for (i=0; i < w; ++i) {
+               short *data = z->img_comp[n].coeff + 64 * (i + j * z->img_comp[n].coeff_w);
+               stbi__jpeg_dequantize(data, z->dequant[z->img_comp[n].tq]);
+               z->idct_block_kernel(z->img_comp[n].data+z->img_comp[n].w2*j*8+i*8, z->img_comp[n].w2, data);
+            }
+         }
+      }
+   }
+}
+
+static int stbi__process_marker(stbi__jpeg *z, int m)
+{
+   int L;
+   switch (m) {
+      case STBI__MARKER_none: // no marker found
+         return stbi__err("expected marker","Corrupt JPEG");
+
+      case 0xDD: // DRI - specify restart interval
+         if (stbi__get16be(z->s) != 4) return stbi__err("bad DRI len","Corrupt JPEG");
+         z->restart_interval = stbi__get16be(z->s);
+         return 1;
+
+      case 0xDB: // DQT - define quantization table
+         L = stbi__get16be(z->s)-2;
+         while (L > 0) {
+            int q = stbi__get8(z->s);
+            int p = q >> 4, sixteen = (p != 0);
+            int t = q & 15,i;
+            if (p != 0 && p != 1) return stbi__err("bad DQT type","Corrupt JPEG");
+            if (t > 3) return stbi__err("bad DQT table","Corrupt JPEG");
+
+            for (i=0; i < 64; ++i)
+               z->dequant[t][stbi__jpeg_dezigzag[i]] = (stbi__uint16)(sixteen ? stbi__get16be(z->s) : stbi__get8(z->s));
+            L -= (sixteen ? 129 : 65);
+         }
+         return L==0;
+
+      case 0xC4: // DHT - define huffman table
+         L = stbi__get16be(z->s)-2;
+         while (L > 0) {
+            stbi_uc *v;
+            int sizes[16],i,n=0;
+            int q = stbi__get8(z->s);
+            int tc = q >> 4;
+            int th = q & 15;
+            if (tc > 1 || th > 3) return stbi__err("bad DHT header","Corrupt JPEG");
+            for (i=0; i < 16; ++i) {
+               sizes[i] = stbi__get8(z->s);
+               n += sizes[i];
+            }
+            if(n > 256) return stbi__err("bad DHT header","Corrupt JPEG"); // Loop over i < n would write past end of values!
+            L -= 17;
+            if (tc == 0) {
+               if (!stbi__build_huffman(z->huff_dc+th, sizes)) return 0;
+               v = z->huff_dc[th].values;
+            } else {
+               if (!stbi__build_huffman(z->huff_ac+th, sizes)) return 0;
+               v = z->huff_ac[th].values;
+            }
+            for (i=0; i < n; ++i)
+               v[i] = stbi__get8(z->s);
+            if (tc != 0)
+               stbi__build_fast_ac(z->fast_ac[th], z->huff_ac + th);
+            L -= n;
+         }
+         return L==0;
+   }
+
+   // check for comment block or APP blocks
+   if ((m >= 0xE0 && m <= 0xEF) || m == 0xFE) {
+      L = stbi__get16be(z->s);
+      if (L < 2) {
+         if (m == 0xFE)
+            return stbi__err("bad COM len","Corrupt JPEG");
+         else
+            return stbi__err("bad APP len","Corrupt JPEG");
+      }
+      L -= 2;
+
+      if (m == 0xE0 && L >= 5) { // JFIF APP0 segment
+         static const unsigned char tag[5] = {'J','F','I','F','\0'};
+         int ok = 1;
+         int i;
+         for (i=0; i < 5; ++i)
+            if (stbi__get8(z->s) != tag[i])
+               ok = 0;
+         L -= 5;
+         if (ok)
+            z->jfif = 1;
+      } else if (m == 0xEE && L >= 12) { // Adobe APP14 segment
+         static const unsigned char tag[6] = {'A','d','o','b','e','\0'};
+         int ok = 1;
+         int i;
+         for (i=0; i < 6; ++i)
+            if (stbi__get8(z->s) != tag[i])
+               ok = 0;
+         L -= 6;
+         if (ok) {
+            stbi__get8(z->s); // version
+            stbi__get16be(z->s); // flags0
+            stbi__get16be(z->s); // flags1
+            z->app14_color_transform = stbi__get8(z->s); // color transform
+            L -= 6;
+         }
+      }
+
+      stbi__skip(z->s, L);
+      return 1;
+   }
+
+   return stbi__err("unknown marker","Corrupt JPEG");
+}
+
+// after we see SOS
+static int stbi__process_scan_header(stbi__jpeg *z)
+{
+   int i;
+   int Ls = stbi__get16be(z->s);
+   z->scan_n = stbi__get8(z->s);
+   if (z->scan_n < 1 || z->scan_n > 4 || z->scan_n > (int) z->s->img_n) return stbi__err("bad SOS component count","Corrupt JPEG");
+   if (Ls != 6+2*z->scan_n) return stbi__err("bad SOS len","Corrupt JPEG");
+   for (i=0; i < z->scan_n; ++i) {
+      int id = stbi__get8(z->s), which;
+      int q = stbi__get8(z->s);
+      for (which = 0; which < z->s->img_n; ++which)
+         if (z->img_comp[which].id == id)
+            break;
+      if (which == z->s->img_n) return 0; // no match
+      z->img_comp[which].hd = q >> 4;   if (z->img_comp[which].hd > 3) return stbi__err("bad DC huff","Corrupt JPEG");
+      z->img_comp[which].ha = q & 15;   if (z->img_comp[which].ha > 3) return stbi__err("bad AC huff","Corrupt JPEG");
+      z->order[i] = which;
+   }
+
+   {
+      int aa;
+      z->spec_start = stbi__get8(z->s);
+      z->spec_end   = stbi__get8(z->s); // should be 63, but might be 0
+      aa = stbi__get8(z->s);
+      z->succ_high = (aa >> 4);
+      z->succ_low  = (aa & 15);
+      if (z->progressive) {
+         if (z->spec_start > 63 || z->spec_end > 63  || z->spec_start > z->spec_end || z->succ_high > 13 || z->succ_low > 13)
+            return stbi__err("bad SOS", "Corrupt JPEG");
+      } else {
+         if (z->spec_start != 0) return stbi__err("bad SOS","Corrupt JPEG");
+         if (z->succ_high != 0 || z->succ_low != 0) return stbi__err("bad SOS","Corrupt JPEG");
+         z->spec_end = 63;
+      }
+   }
+
+   return 1;
+}
+
+static int stbi__free_jpeg_components(stbi__jpeg *z, int ncomp, int why)
+{
+   int i;
+   for (i=0; i < ncomp; ++i) {
+      if (z->img_comp[i].raw_data) {
+         STBI_FREE(z->img_comp[i].raw_data);
+         z->img_comp[i].raw_data = NULL;
+         z->img_comp[i].data = NULL;
+      }
+      if (z->img_comp[i].raw_coeff) {
+         STBI_FREE(z->img_comp[i].raw_coeff);
+         z->img_comp[i].raw_coeff = 0;
+         z->img_comp[i].coeff = 0;
+      }
+      if (z->img_comp[i].linebuf) {
+         STBI_FREE(z->img_comp[i].linebuf);
+         z->img_comp[i].linebuf = NULL;
+      }
+   }
+   return why;
+}
+
+static int stbi__process_frame_header(stbi__jpeg *z, int scan)
+{
+   stbi__context *s = z->s;
+   int Lf,p,i,q, h_max=1,v_max=1,c;
+   Lf = stbi__get16be(s);         if (Lf < 11) return stbi__err("bad SOF len","Corrupt JPEG"); // JPEG
+   p  = stbi__get8(s);            if (p != 8) return stbi__err("only 8-bit","JPEG format not supported: 8-bit only"); // JPEG baseline
+   s->img_y = stbi__get16be(s);   if (s->img_y == 0) return stbi__err("no header height", "JPEG format not supported: delayed height"); // Legal, but we don't handle it--but neither does IJG
+   s->img_x = stbi__get16be(s);   if (s->img_x == 0) return stbi__err("0 width","Corrupt JPEG"); // JPEG requires
+   if (s->img_y > STBI_MAX_DIMENSIONS) return stbi__err("too large","Very large image (corrupt?)");
+   if (s->img_x > STBI_MAX_DIMENSIONS) return stbi__err("too large","Very large image (corrupt?)");
+   c = stbi__get8(s);
+   if (c != 3 && c != 1 && c != 4) return stbi__err("bad component count","Corrupt JPEG");
+   s->img_n = c;
+   for (i=0; i < c; ++i) {
+      z->img_comp[i].data = NULL;
+      z->img_comp[i].linebuf = NULL;
+   }
+
+   if (Lf != 8+3*s->img_n) return stbi__err("bad SOF len","Corrupt JPEG");
+
+   z->rgb = 0;
+   for (i=0; i < s->img_n; ++i) {
+      static const unsigned char rgb[3] = { 'R', 'G', 'B' };
+      z->img_comp[i].id = stbi__get8(s);
+      if (s->img_n == 3 && z->img_comp[i].id == rgb[i])
+         ++z->rgb;
+      q = stbi__get8(s);
+      z->img_comp[i].h = (q >> 4);  if (!z->img_comp[i].h || z->img_comp[i].h > 4) return stbi__err("bad H","Corrupt JPEG");
+      z->img_comp[i].v = q & 15;    if (!z->img_comp[i].v || z->img_comp[i].v > 4) return stbi__err("bad V","Corrupt JPEG");
+      z->img_comp[i].tq = stbi__get8(s);  if (z->img_comp[i].tq > 3) return stbi__err("bad TQ","Corrupt JPEG");
+   }
+
+   if (scan != STBI__SCAN_load) return 1;
+
+   if (!stbi__mad3sizes_valid(s->img_x, s->img_y, s->img_n, 0)) return stbi__err("too large", "Image too large to decode");
+
+   for (i=0; i < s->img_n; ++i) {
+      if (z->img_comp[i].h > h_max) h_max = z->img_comp[i].h;
+      if (z->img_comp[i].v > v_max) v_max = z->img_comp[i].v;
+   }
+
+   // check that plane subsampling factors are integer ratios; our resamplers can't deal with fractional ratios
+   // and I've never seen a non-corrupted JPEG file actually use them
+   for (i=0; i < s->img_n; ++i) {
+      if (h_max % z->img_comp[i].h != 0) return stbi__err("bad H","Corrupt JPEG");
+      if (v_max % z->img_comp[i].v != 0) return stbi__err("bad V","Corrupt JPEG");
+   }
+
+   // compute interleaved mcu info
+   z->img_h_max = h_max;
+   z->img_v_max = v_max;
+   z->img_mcu_w = h_max * 8;
+   z->img_mcu_h = v_max * 8;
+   // these sizes can't be more than 17 bits
+   z->img_mcu_x = (s->img_x + z->img_mcu_w-1) / z->img_mcu_w;
+   z->img_mcu_y = (s->img_y + z->img_mcu_h-1) / z->img_mcu_h;
+
+   for (i=0; i < s->img_n; ++i) {
+      // number of effective pixels (e.g. for non-interleaved MCU)
+      z->img_comp[i].x = (s->img_x * z->img_comp[i].h + h_max-1) / h_max;
+      z->img_comp[i].y = (s->img_y * z->img_comp[i].v + v_max-1) / v_max;
+      // to simplify generation, we'll allocate enough memory to decode
+      // the bogus oversized data from using interleaved MCUs and their
+      // big blocks (e.g. a 16x16 iMCU on an image of width 33); we won't
+      // discard the extra data until colorspace conversion
+      //
+      // img_mcu_x, img_mcu_y: <=17 bits; comp[i].h and .v are <=4 (checked earlier)
+      // so these muls can't overflow with 32-bit ints (which we require)
+      z->img_comp[i].w2 = z->img_mcu_x * z->img_comp[i].h * 8;
+      z->img_comp[i].h2 = z->img_mcu_y * z->img_comp[i].v * 8;
+      z->img_comp[i].coeff = 0;
+      z->img_comp[i].raw_coeff = 0;
+      z->img_comp[i].linebuf = NULL;
+      z->img_comp[i].raw_data = stbi__malloc_mad2(z->img_comp[i].w2, z->img_comp[i].h2, 15);
+      if (z->img_comp[i].raw_data == NULL)
+         return stbi__free_jpeg_components(z, i+1, stbi__err("outofmem", "Out of memory"));
+      // align blocks for idct using mmx/sse
+      z->img_comp[i].data = (stbi_uc*) (((size_t) z->img_comp[i].raw_data + 15) & ~15);
+      if (z->progressive) {
+         // w2, h2 are multiples of 8 (see above)
+         z->img_comp[i].coeff_w = z->img_comp[i].w2 / 8;
+         z->img_comp[i].coeff_h = z->img_comp[i].h2 / 8;
+         z->img_comp[i].raw_coeff = stbi__malloc_mad3(z->img_comp[i].w2, z->img_comp[i].h2, sizeof(short), 15);
+         if (z->img_comp[i].raw_coeff == NULL)
+            return stbi__free_jpeg_components(z, i+1, stbi__err("outofmem", "Out of memory"));
+         z->img_comp[i].coeff = (short*) (((size_t) z->img_comp[i].raw_coeff + 15) & ~15);
+      }
+   }
+
+   return 1;
+}
+
+// use comparisons since in some cases we handle more than one case (e.g. SOF)
+#define stbi__DNL(x)         ((x) == 0xdc)
+#define stbi__SOI(x)         ((x) == 0xd8)
+#define stbi__EOI(x)         ((x) == 0xd9)
+#define stbi__SOF(x)         ((x) == 0xc0 || (x) == 0xc1 || (x) == 0xc2)
+#define stbi__SOS(x)         ((x) == 0xda)
+
+#define stbi__SOF_progressive(x)   ((x) == 0xc2)
+
+static int stbi__decode_jpeg_header(stbi__jpeg *z, int scan)
+{
+   int m;
+   z->jfif = 0;
+   z->app14_color_transform = -1; // valid values are 0,1,2
+   z->marker = STBI__MARKER_none; // initialize cached marker to empty
+   m = stbi__get_marker(z);
+   if (!stbi__SOI(m)) return stbi__err("no SOI","Corrupt JPEG");
+   if (scan == STBI__SCAN_type) return 1;
+   m = stbi__get_marker(z);
+   while (!stbi__SOF(m)) {
+      if (!stbi__process_marker(z,m)) return 0;
+      m = stbi__get_marker(z);
+      while (m == STBI__MARKER_none) {
+         // some files have extra padding after their blocks, so ok, we'll scan
+         if (stbi__at_eof(z->s)) return stbi__err("no SOF", "Corrupt JPEG");
+         m = stbi__get_marker(z);
+      }
+   }
+   z->progressive = stbi__SOF_progressive(m);
+   if (!stbi__process_frame_header(z, scan)) return 0;
+   return 1;
+}
+
+static int stbi__skip_jpeg_junk_at_end(stbi__jpeg *j)
+{
+   // some JPEGs have junk at end, skip over it but if we find what looks
+   // like a valid marker, resume there
+   while (!stbi__at_eof(j->s)) {
+      int x = stbi__get8(j->s);
+      while (x == 255) { // might be a marker
+         if (stbi__at_eof(j->s)) return STBI__MARKER_none;
+         x = stbi__get8(j->s);
+         if (x != 0x00 && x != 0xff) {
+            // not a stuffed zero or lead-in to another marker, looks
+            // like an actual marker, return it
+            return x;
+         }
+         // stuffed zero has x=0 now which ends the loop, meaning we go
+         // back to regular scan loop.
+         // repeated 0xff keeps trying to read the next byte of the marker.
+      }
+   }
+   return STBI__MARKER_none;
+}
+
+// decode image to YCbCr format
+static int stbi__decode_jpeg_image(stbi__jpeg *j)
+{
+   int m;
+   for (m = 0; m < 4; m++) {
+      j->img_comp[m].raw_data = NULL;
+      j->img_comp[m].raw_coeff = NULL;
+   }
+   j->restart_interval = 0;
+   if (!stbi__decode_jpeg_header(j, STBI__SCAN_load)) return 0;
+   m = stbi__get_marker(j);
+   while (!stbi__EOI(m)) {
+      if (stbi__SOS(m)) {
+         if (!stbi__process_scan_header(j)) return 0;
+         if (!stbi__parse_entropy_coded_data(j)) return 0;
+         if (j->marker == STBI__MARKER_none ) {
+         j->marker = stbi__skip_jpeg_junk_at_end(j);
+            // if we reach eof without hitting a marker, stbi__get_marker() below will fail and we'll eventually return 0
+         }
+         m = stbi__get_marker(j);
+         if (STBI__RESTART(m))
+            m = stbi__get_marker(j);
+      } else if (stbi__DNL(m)) {
+         int Ld = stbi__get16be(j->s);
+         stbi__uint32 NL = stbi__get16be(j->s);
+         if (Ld != 4) return stbi__err("bad DNL len", "Corrupt JPEG");
+         if (NL != j->s->img_y) return stbi__err("bad DNL height", "Corrupt JPEG");
+         m = stbi__get_marker(j);
+      } else {
+         if (!stbi__process_marker(j, m)) return 1;
+         m = stbi__get_marker(j);
+      }
+   }
+   if (j->progressive)
+      stbi__jpeg_finish(j);
+   return 1;
+}
+
+// static jfif-centered resampling (across block boundaries)
+
+typedef stbi_uc *(*resample_row_func)(stbi_uc *out, stbi_uc *in0, stbi_uc *in1,
+                                    int w, int hs);
+
+#define stbi__div4(x) ((stbi_uc) ((x) >> 2))
+
+static stbi_uc *resample_row_1(stbi_uc *out, stbi_uc *in_near, stbi_uc *in_far, int w, int hs)
+{
+   STBI_NOTUSED(out);
+   STBI_NOTUSED(in_far);
+   STBI_NOTUSED(w);
+   STBI_NOTUSED(hs);
+   return in_near;
+}
+
+static stbi_uc* stbi__resample_row_v_2(stbi_uc *out, stbi_uc *in_near, stbi_uc *in_far, int w, int hs)
+{
+   // need to generate two samples vertically for every one in input
+   int i;
+   STBI_NOTUSED(hs);
+   for (i=0; i < w; ++i)
+      out[i] = stbi__div4(3*in_near[i] + in_far[i] + 2);
+   return out;
+}
+
+static stbi_uc*  stbi__resample_row_h_2(stbi_uc *out, stbi_uc *in_near, stbi_uc *in_far, int w, int hs)
+{
+   // need to generate two samples horizontally for every one in input
+   int i;
+   stbi_uc *input = in_near;
+
+   if (w == 1) {
+      // if only one sample, can't do any interpolation
+      out[0] = out[1] = input[0];
+      return out;
+   }
+
+   out[0] = input[0];
+   out[1] = stbi__div4(input[0]*3 + input[1] + 2);
+   for (i=1; i < w-1; ++i) {
+      int n = 3*input[i]+2;
+      out[i*2+0] = stbi__div4(n+input[i-1]);
+      out[i*2+1] = stbi__div4(n+input[i+1]);
+   }
+   out[i*2+0] = stbi__div4(input[w-2]*3 + input[w-1] + 2);
+   out[i*2+1] = input[w-1];
+
+   STBI_NOTUSED(in_far);
+   STBI_NOTUSED(hs);
+
+   return out;
+}
+
+#define stbi__div16(x) ((stbi_uc) ((x) >> 4))
+
+static stbi_uc *stbi__resample_row_hv_2(stbi_uc *out, stbi_uc *in_near, stbi_uc *in_far, int w, int hs)
+{
+   // need to generate 2x2 samples for every one in input
+   int i,t0,t1;
+   if (w == 1) {
+      out[0] = out[1] = stbi__div4(3*in_near[0] + in_far[0] + 2);
+      return out;
+   }
+
+   t1 = 3*in_near[0] + in_far[0];
+   out[0] = stbi__div4(t1+2);
+   for (i=1; i < w; ++i) {
+      t0 = t1;
+      t1 = 3*in_near[i]+in_far[i];
+      out[i*2-1] = stbi__div16(3*t0 + t1 + 8);
+      out[i*2  ] = stbi__div16(3*t1 + t0 + 8);
+   }
+   out[w*2-1] = stbi__div4(t1+2);
+
+   STBI_NOTUSED(hs);
+
+   return out;
+}
+
+#if defined(STBI_SSE2) || defined(STBI_NEON)
+static stbi_uc *stbi__resample_row_hv_2_simd(stbi_uc *out, stbi_uc *in_near, stbi_uc *in_far, int w, int hs)
+{
+   // need to generate 2x2 samples for every one in input
+   int i=0,t0,t1;
+
+   if (w == 1) {
+      out[0] = out[1] = stbi__div4(3*in_near[0] + in_far[0] + 2);
+      return out;
+   }
+
+   t1 = 3*in_near[0] + in_far[0];
+   // process groups of 8 pixels for as long as we can.
+   // note we can't handle the last pixel in a row in this loop
+   // because we need to handle the filter boundary conditions.
+   for (; i < ((w-1) & ~7); i += 8) {
+#if defined(STBI_SSE2)
+      // load and perform the vertical filtering pass
+      // this uses 3*x + y = 4*x + (y - x)
+      __m128i zero  = _mm_setzero_si128();
+      __m128i farb  = _mm_loadl_epi64((__m128i *) (in_far + i));
+      __m128i nearb = _mm_loadl_epi64((__m128i *) (in_near + i));
+      __m128i farw  = _mm_unpacklo_epi8(farb, zero);
+      __m128i nearw = _mm_unpacklo_epi8(nearb, zero);
+      __m128i diff  = _mm_sub_epi16(farw, nearw);
+      __m128i nears = _mm_slli_epi16(nearw, 2);
+      __m128i curr  = _mm_add_epi16(nears, diff); // current row
+
+      // horizontal filter works the same based on shifted vers of current
+      // row. "prev" is current row shifted right by 1 pixel; we need to
+      // insert the previous pixel value (from t1).
+      // "next" is current row shifted left by 1 pixel, with first pixel
+      // of next block of 8 pixels added in.
+      __m128i prv0 = _mm_slli_si128(curr, 2);
+      __m128i nxt0 = _mm_srli_si128(curr, 2);
+      __m128i prev = _mm_insert_epi16(prv0, t1, 0);
+      __m128i next = _mm_insert_epi16(nxt0, 3*in_near[i+8] + in_far[i+8], 7);
+
+      // horizontal filter, polyphase implementation since it's convenient:
+      // even pixels = 3*cur + prev = cur*4 + (prev - cur)
+      // odd  pixels = 3*cur + next = cur*4 + (next - cur)
+      // note the shared term.
+      __m128i bias  = _mm_set1_epi16(8);
+      __m128i curs = _mm_slli_epi16(curr, 2);
+      __m128i prvd = _mm_sub_epi16(prev, curr);
+      __m128i nxtd = _mm_sub_epi16(next, curr);
+      __m128i curb = _mm_add_epi16(curs, bias);
+      __m128i even = _mm_add_epi16(prvd, curb);
+      __m128i odd  = _mm_add_epi16(nxtd, curb);
+
+      // interleave even and odd pixels, then undo scaling.
+      __m128i int0 = _mm_unpacklo_epi16(even, odd);
+      __m128i int1 = _mm_unpackhi_epi16(even, odd);
+      __m128i de0  = _mm_srli_epi16(int0, 4);
+      __m128i de1  = _mm_srli_epi16(int1, 4);
+
+      // pack and write output
+      __m128i outv = _mm_packus_epi16(de0, de1);
+      _mm_storeu_si128((__m128i *) (out + i*2), outv);
+#elif defined(STBI_NEON)
+      // load and perform the vertical filtering pass
+      // this uses 3*x + y = 4*x + (y - x)
+      uint8x8_t farb  = vld1_u8(in_far + i);
+      uint8x8_t nearb = vld1_u8(in_near + i);
+      int16x8_t diff  = vreinterpretq_s16_u16(vsubl_u8(farb, nearb));
+      int16x8_t nears = vreinterpretq_s16_u16(vshll_n_u8(nearb, 2));
+      int16x8_t curr  = vaddq_s16(nears, diff); // current row
+
+      // horizontal filter works the same based on shifted vers of current
+      // row. "prev" is current row shifted right by 1 pixel; we need to
+      // insert the previous pixel value (from t1).
+      // "next" is current row shifted left by 1 pixel, with first pixel
+      // of next block of 8 pixels added in.
+      int16x8_t prv0 = vextq_s16(curr, curr, 7);
+      int16x8_t nxt0 = vextq_s16(curr, curr, 1);
+      int16x8_t prev = vsetq_lane_s16(t1, prv0, 0);
+      int16x8_t next = vsetq_lane_s16(3*in_near[i+8] + in_far[i+8], nxt0, 7);
+
+      // horizontal filter, polyphase implementation since it's convenient:
+      // even pixels = 3*cur + prev = cur*4 + (prev - cur)
+      // odd  pixels = 3*cur + next = cur*4 + (next - cur)
+      // note the shared term.
+      int16x8_t curs = vshlq_n_s16(curr, 2);
+      int16x8_t prvd = vsubq_s16(prev, curr);
+      int16x8_t nxtd = vsubq_s16(next, curr);
+      int16x8_t even = vaddq_s16(curs, prvd);
+      int16x8_t odd  = vaddq_s16(curs, nxtd);
+
+      // undo scaling and round, then store with even/odd phases interleaved
+      uint8x8x2_t o;
+      o.val[0] = vqrshrun_n_s16(even, 4);
+      o.val[1] = vqrshrun_n_s16(odd,  4);
+      vst2_u8(out + i*2, o);
+#endif
+
+      // "previous" value for next iter
+      t1 = 3*in_near[i+7] + in_far[i+7];
+   }
+
+   t0 = t1;
+   t1 = 3*in_near[i] + in_far[i];
+   out[i*2] = stbi__div16(3*t1 + t0 + 8);
+
+   for (++i; i < w; ++i) {
+      t0 = t1;
+      t1 = 3*in_near[i]+in_far[i];
+      out[i*2-1] = stbi__div16(3*t0 + t1 + 8);
+      out[i*2  ] = stbi__div16(3*t1 + t0 + 8);
+   }
+   out[w*2-1] = stbi__div4(t1+2);
+
+   STBI_NOTUSED(hs);
+
+   return out;
+}
+#endif
+
+static stbi_uc *stbi__resample_row_generic(stbi_uc *out, stbi_uc *in_near, stbi_uc *in_far, int w, int hs)
+{
+   // resample with nearest-neighbor
+   int i,j;
+   STBI_NOTUSED(in_far);
+   for (i=0; i < w; ++i)
+      for (j=0; j < hs; ++j)
+         out[i*hs+j] = in_near[i];
+   return out;
+}
+
+// this is a reduced-precision calculation of YCbCr-to-RGB introduced
+// to make sure the code produces the same results in both SIMD and scalar
+#define stbi__float2fixed(x)  (((int) ((x) * 4096.0f + 0.5f)) << 8)
+static void stbi__YCbCr_to_RGB_row(stbi_uc *out, const stbi_uc *y, const stbi_uc *pcb, const stbi_uc *pcr, int count, int step)
+{
+   int i;
+   for (i=0; i < count; ++i) {
+      int y_fixed = (y[i] << 20) + (1<<19); // rounding
+      int r,g,b;
+      int cr = pcr[i] - 128;
+      int cb = pcb[i] - 128;
+      r = y_fixed +  cr* stbi__float2fixed(1.40200f);
+      g = y_fixed + (cr*-stbi__float2fixed(0.71414f)) + ((cb*-stbi__float2fixed(0.34414f)) & 0xffff0000);
+      b = y_fixed                                     +   cb* stbi__float2fixed(1.77200f);
+      r >>= 20;
+      g >>= 20;
+      b >>= 20;
+      if ((unsigned) r > 255) { if (r < 0) r = 0; else r = 255; }
+      if ((unsigned) g > 255) { if (g < 0) g = 0; else g = 255; }
+      if ((unsigned) b > 255) { if (b < 0) b = 0; else b = 255; }
+      out[0] = (stbi_uc)r;
+      out[1] = (stbi_uc)g;
+      out[2] = (stbi_uc)b;
+      out[3] = 255;
+      out += step;
+   }
+}
+
+#if defined(STBI_SSE2) || defined(STBI_NEON)
+static void stbi__YCbCr_to_RGB_simd(stbi_uc *out, stbi_uc const *y, stbi_uc const *pcb, stbi_uc const *pcr, int count, int step)
+{
+   int i = 0;
+
+#ifdef STBI_SSE2
+   // step == 3 is pretty ugly on the final interleave, and i'm not convinced
+   // it's useful in practice (you wouldn't use it for textures, for example).
+   // so just accelerate step == 4 case.
+   if (step == 4) {
+      // this is a fairly straightforward implementation and not super-optimized.
+      __m128i signflip  = _mm_set1_epi8(-0x80);
+      __m128i cr_const0 = _mm_set1_epi16(   (short) ( 1.40200f*4096.0f+0.5f));
+      __m128i cr_const1 = _mm_set1_epi16( - (short) ( 0.71414f*4096.0f+0.5f));
+      __m128i cb_const0 = _mm_set1_epi16( - (short) ( 0.34414f*4096.0f+0.5f));
+      __m128i cb_const1 = _mm_set1_epi16(   (short) ( 1.77200f*4096.0f+0.5f));
+      __m128i y_bias = _mm_set1_epi8((char) (unsigned char) 128);
+      __m128i xw = _mm_set1_epi16(255); // alpha channel
+
+      for (; i+7 < count; i += 8) {
+         // load
+         __m128i y_bytes = _mm_loadl_epi64((__m128i *) (y+i));
+         __m128i cr_bytes = _mm_loadl_epi64((__m128i *) (pcr+i));
+         __m128i cb_bytes = _mm_loadl_epi64((__m128i *) (pcb+i));
+         __m128i cr_biased = _mm_xor_si128(cr_bytes, signflip); // -128
+         __m128i cb_biased = _mm_xor_si128(cb_bytes, signflip); // -128
+
+         // unpack to short (and left-shift cr, cb by 8)
+         __m128i yw  = _mm_unpacklo_epi8(y_bias, y_bytes);
+         __m128i crw = _mm_unpacklo_epi8(_mm_setzero_si128(), cr_biased);
+         __m128i cbw = _mm_unpacklo_epi8(_mm_setzero_si128(), cb_biased);
+
+         // color transform
+         __m128i yws = _mm_srli_epi16(yw, 4);
+         __m128i cr0 = _mm_mulhi_epi16(cr_const0, crw);
+         __m128i cb0 = _mm_mulhi_epi16(cb_const0, cbw);
+         __m128i cb1 = _mm_mulhi_epi16(cbw, cb_const1);
+         __m128i cr1 = _mm_mulhi_epi16(crw, cr_const1);
+         __m128i rws = _mm_add_epi16(cr0, yws);
+         __m128i gwt = _mm_add_epi16(cb0, yws);
+         __m128i bws = _mm_add_epi16(yws, cb1);
+         __m128i gws = _mm_add_epi16(gwt, cr1);
+
+         // descale
+         __m128i rw = _mm_srai_epi16(rws, 4);
+         __m128i bw = _mm_srai_epi16(bws, 4);
+         __m128i gw = _mm_srai_epi16(gws, 4);
+
+         // back to byte, set up for transpose
+         __m128i brb = _mm_packus_epi16(rw, bw);
+         __m128i gxb = _mm_packus_epi16(gw, xw);
+
+         // transpose to interleave channels
+         __m128i t0 = _mm_unpacklo_epi8(brb, gxb);
+         __m128i t1 = _mm_unpackhi_epi8(brb, gxb);
+         __m128i o0 = _mm_unpacklo_epi16(t0, t1);
+         __m128i o1 = _mm_unpackhi_epi16(t0, t1);
+
+         // store
+         _mm_storeu_si128((__m128i *) (out + 0), o0);
+         _mm_storeu_si128((__m128i *) (out + 16), o1);
+         out += 32;
+      }
+   }
+#endif
+
+#ifdef STBI_NEON
+   // in this version, step=3 support would be easy to add. but is there demand?
+   if (step == 4) {
+      // this is a fairly straightforward implementation and not super-optimized.
+      uint8x8_t signflip = vdup_n_u8(0x80);
+      int16x8_t cr_const0 = vdupq_n_s16(   (short) ( 1.40200f*4096.0f+0.5f));
+      int16x8_t cr_const1 = vdupq_n_s16( - (short) ( 0.71414f*4096.0f+0.5f));
+      int16x8_t cb_const0 = vdupq_n_s16( - (short) ( 0.34414f*4096.0f+0.5f));
+      int16x8_t cb_const1 = vdupq_n_s16(   (short) ( 1.77200f*4096.0f+0.5f));
+
+      for (; i+7 < count; i += 8) {
+         // load
+         uint8x8_t y_bytes  = vld1_u8(y + i);
+         uint8x8_t cr_bytes = vld1_u8(pcr + i);
+         uint8x8_t cb_bytes = vld1_u8(pcb + i);
+         int8x8_t cr_biased = vreinterpret_s8_u8(vsub_u8(cr_bytes, signflip));
+         int8x8_t cb_biased = vreinterpret_s8_u8(vsub_u8(cb_bytes, signflip));
+
+         // expand to s16
+         int16x8_t yws = vreinterpretq_s16_u16(vshll_n_u8(y_bytes, 4));
+         int16x8_t crw = vshll_n_s8(cr_biased, 7);
+         int16x8_t cbw = vshll_n_s8(cb_biased, 7);
+
+         // color transform
+         int16x8_t cr0 = vqdmulhq_s16(crw, cr_const0);
+         int16x8_t cb0 = vqdmulhq_s16(cbw, cb_const0);
+         int16x8_t cr1 = vqdmulhq_s16(crw, cr_const1);
+         int16x8_t cb1 = vqdmulhq_s16(cbw, cb_const1);
+         int16x8_t rws = vaddq_s16(yws, cr0);
+         int16x8_t gws = vaddq_s16(vaddq_s16(yws, cb0), cr1);
+         int16x8_t bws = vaddq_s16(yws, cb1);
+
+         // undo scaling, round, convert to byte
+         uint8x8x4_t o;
+         o.val[0] = vqrshrun_n_s16(rws, 4);
+         o.val[1] = vqrshrun_n_s16(gws, 4);
+         o.val[2] = vqrshrun_n_s16(bws, 4);
+         o.val[3] = vdup_n_u8(255);
+
+         // store, interleaving r/g/b/a
+         vst4_u8(out, o);
+         out += 8*4;
+      }
+   }
+#endif
+
+   for (; i < count; ++i) {
+      int y_fixed = (y[i] << 20) + (1<<19); // rounding
+      int r,g,b;
+      int cr = pcr[i] - 128;
+      int cb = pcb[i] - 128;
+      r = y_fixed + cr* stbi__float2fixed(1.40200f);
+      g = y_fixed + cr*-stbi__float2fixed(0.71414f) + ((cb*-stbi__float2fixed(0.34414f)) & 0xffff0000);
+      b = y_fixed                                   +   cb* stbi__float2fixed(1.77200f);
+      r >>= 20;
+      g >>= 20;
+      b >>= 20;
+      if ((unsigned) r > 255) { if (r < 0) r = 0; else r = 255; }
+      if ((unsigned) g > 255) { if (g < 0) g = 0; else g = 255; }
+      if ((unsigned) b > 255) { if (b < 0) b = 0; else b = 255; }
+      out[0] = (stbi_uc)r;
+      out[1] = (stbi_uc)g;
+      out[2] = (stbi_uc)b;
+      out[3] = 255;
+      out += step;
+   }
+}
+#endif
+
+// set up the kernels
+static void stbi__setup_jpeg(stbi__jpeg *j)
+{
+   j->idct_block_kernel = stbi__idct_block;
+   j->YCbCr_to_RGB_kernel = stbi__YCbCr_to_RGB_row;
+   j->resample_row_hv_2_kernel = stbi__resample_row_hv_2;
+
+#ifdef STBI_SSE2
+   if (stbi__sse2_available()) {
+      j->idct_block_kernel = stbi__idct_simd;
+      j->YCbCr_to_RGB_kernel = stbi__YCbCr_to_RGB_simd;
+      j->resample_row_hv_2_kernel = stbi__resample_row_hv_2_simd;
+   }
+#endif
+
+#ifdef STBI_NEON
+   j->idct_block_kernel = stbi__idct_simd;
+   j->YCbCr_to_RGB_kernel = stbi__YCbCr_to_RGB_simd;
+   j->resample_row_hv_2_kernel = stbi__resample_row_hv_2_simd;
+#endif
+}
+
+// clean up the temporary component buffers
+static void stbi__cleanup_jpeg(stbi__jpeg *j)
+{
+   stbi__free_jpeg_components(j, j->s->img_n, 0);
+}
+
+typedef struct
+{
+   resample_row_func resample;
+   stbi_uc *line0,*line1;
+   int hs,vs;   // expansion factor in each axis
+   int w_lores; // horizontal pixels pre-expansion
+   int ystep;   // how far through vertical expansion we are
+   int ypos;    // which pre-expansion row we're on
+} stbi__resample;
+
+// fast 0..255 * 0..255 => 0..255 rounded multiplication
+static stbi_uc stbi__blinn_8x8(stbi_uc x, stbi_uc y)
+{
+   unsigned int t = x*y + 128;
+   return (stbi_uc) ((t + (t >>8)) >> 8);
+}
+
+static stbi_uc *load_jpeg_image(stbi__jpeg *z, int *out_x, int *out_y, int *comp, int req_comp)
+{
+   int n, decode_n, is_rgb;
+   z->s->img_n = 0; // make stbi__cleanup_jpeg safe
+
+   // validate req_comp
+   if (req_comp < 0 || req_comp > 4) return stbi__errpuc("bad req_comp", "Internal error");
+
+   // load a jpeg image from whichever source, but leave in YCbCr format
+   if (!stbi__decode_jpeg_image(z)) { stbi__cleanup_jpeg(z); return NULL; }
+
+   // determine actual number of components to generate
+   n = req_comp ? req_comp : z->s->img_n >= 3 ? 3 : 1;
+
+   is_rgb = z->s->img_n == 3 && (z->rgb == 3 || (z->app14_color_transform == 0 && !z->jfif));
+
+   if (z->s->img_n == 3 && n < 3 && !is_rgb)
+      decode_n = 1;
+   else
+      decode_n = z->s->img_n;
+
+   // nothing to do if no components requested; check this now to avoid
+   // accessing uninitialized coutput[0] later
+   if (decode_n <= 0) { stbi__cleanup_jpeg(z); return NULL; }
+
+   // resample and color-convert
+   {
+      int k;
+      unsigned int i,j;
+      stbi_uc *output;
+      stbi_uc *coutput[4] = { NULL, NULL, NULL, NULL };
+
+      stbi__resample res_comp[4];
+
+      for (k=0; k < decode_n; ++k) {
+         stbi__resample *r = &res_comp[k];
+
+         // allocate line buffer big enough for upsampling off the edges
+         // with upsample factor of 4
+         z->img_comp[k].linebuf = (stbi_uc *) stbi__malloc(z->s->img_x + 3);
+         if (!z->img_comp[k].linebuf) { stbi__cleanup_jpeg(z); return stbi__errpuc("outofmem", "Out of memory"); }
+
+         r->hs      = z->img_h_max / z->img_comp[k].h;
+         r->vs      = z->img_v_max / z->img_comp[k].v;
+         r->ystep   = r->vs >> 1;
+         r->w_lores = (z->s->img_x + r->hs-1) / r->hs;
+         r->ypos    = 0;
+         r->line0   = r->line1 = z->img_comp[k].data;
+
+         if      (r->hs == 1 && r->vs == 1) r->resample = resample_row_1;
+         else if (r->hs == 1 && r->vs == 2) r->resample = stbi__resample_row_v_2;
+         else if (r->hs == 2 && r->vs == 1) r->resample = stbi__resample_row_h_2;
+         else if (r->hs == 2 && r->vs == 2) r->resample = z->resample_row_hv_2_kernel;
+         else                               r->resample = stbi__resample_row_generic;
+      }
+
+      // can't error after this so, this is safe
+      output = (stbi_uc *) stbi__malloc_mad3(n, z->s->img_x, z->s->img_y, 1);
+      if (!output) { stbi__cleanup_jpeg(z); return stbi__errpuc("outofmem", "Out of memory"); }
+
+      // now go ahead and resample
+      for (j=0; j < z->s->img_y; ++j) {
+         stbi_uc *out = output + n * z->s->img_x * j;
+         for (k=0; k < decode_n; ++k) {
+            stbi__resample *r = &res_comp[k];
+            int y_bot = r->ystep >= (r->vs >> 1);
+            coutput[k] = r->resample(z->img_comp[k].linebuf,
+                                     y_bot ? r->line1 : r->line0,
+                                     y_bot ? r->line0 : r->line1,
+                                     r->w_lores, r->hs);
+            if (++r->ystep >= r->vs) {
+               r->ystep = 0;
+               r->line0 = r->line1;
+               if (++r->ypos < z->img_comp[k].y)
+                  r->line1 += z->img_comp[k].w2;
+            }
+         }
+         if (n >= 3) {
+            stbi_uc *y = coutput[0];
+            if (z->s->img_n == 3) {
+               if (is_rgb) {
+                  for (i=0; i < z->s->img_x; ++i) {
+                     out[0] = y[i];
+                     out[1] = coutput[1][i];
+                     out[2] = coutput[2][i];
+                     out[3] = 255;
+                     out += n;
+                  }
+               } else {
+                  z->YCbCr_to_RGB_kernel(out, y, coutput[1], coutput[2], z->s->img_x, n);
+               }
+            } else if (z->s->img_n == 4) {
+               if (z->app14_color_transform == 0) { // CMYK
+                  for (i=0; i < z->s->img_x; ++i) {
+                     stbi_uc m = coutput[3][i];
+                     out[0] = stbi__blinn_8x8(coutput[0][i], m);
+                     out[1] = stbi__blinn_8x8(coutput[1][i], m);
+                     out[2] = stbi__blinn_8x8(coutput[2][i], m);
+                     out[3] = 255;
+                     out += n;
+                  }
+               } else if (z->app14_color_transform == 2) { // YCCK
+                  z->YCbCr_to_RGB_kernel(out, y, coutput[1], coutput[2], z->s->img_x, n);
+                  for (i=0; i < z->s->img_x; ++i) {
+                     stbi_uc m = coutput[3][i];
+                     out[0] = stbi__blinn_8x8(255 - out[0], m);
+                     out[1] = stbi__blinn_8x8(255 - out[1], m);
+                     out[2] = stbi__blinn_8x8(255 - out[2], m);
+                     out += n;
+                  }
+               } else { // YCbCr + alpha?  Ignore the fourth channel for now
+                  z->YCbCr_to_RGB_kernel(out, y, coutput[1], coutput[2], z->s->img_x, n);
+               }
+            } else
+               for (i=0; i < z->s->img_x; ++i) {
+                  out[0] = out[1] = out[2] = y[i];
+                  out[3] = 255; // not used if n==3
+                  out += n;
+               }
+         } else {
+            if (is_rgb) {
+               if (n == 1)
+                  for (i=0; i < z->s->img_x; ++i)
+                     *out++ = stbi__compute_y(coutput[0][i], coutput[1][i], coutput[2][i]);
+               else {
+                  for (i=0; i < z->s->img_x; ++i, out += 2) {
+                     out[0] = stbi__compute_y(coutput[0][i], coutput[1][i], coutput[2][i]);
+                     out[1] = 255;
+                  }
+               }
+            } else if (z->s->img_n == 4 && z->app14_color_transform == 0) {
+               for (i=0; i < z->s->img_x; ++i) {
+                  stbi_uc m = coutput[3][i];
+                  stbi_uc r = stbi__blinn_8x8(coutput[0][i], m);
+                  stbi_uc g = stbi__blinn_8x8(coutput[1][i], m);
+                  stbi_uc b = stbi__blinn_8x8(coutput[2][i], m);
+                  out[0] = stbi__compute_y(r, g, b);
+                  out[1] = 255;
+                  out += n;
+               }
+            } else if (z->s->img_n == 4 && z->app14_color_transform == 2) {
+               for (i=0; i < z->s->img_x; ++i) {
+                  out[0] = stbi__blinn_8x8(255 - coutput[0][i], coutput[3][i]);
+                  out[1] = 255;
+                  out += n;
+               }
+            } else {
+               stbi_uc *y = coutput[0];
+               if (n == 1)
+                  for (i=0; i < z->s->img_x; ++i) out[i] = y[i];
+               else
+                  for (i=0; i < z->s->img_x; ++i) { *out++ = y[i]; *out++ = 255; }
+            }
+         }
+      }
+      stbi__cleanup_jpeg(z);
+      *out_x = z->s->img_x;
+      *out_y = z->s->img_y;
+      if (comp) *comp = z->s->img_n >= 3 ? 3 : 1; // report original components, not output
+      return output;
+   }
+}
+
+static void *stbi__jpeg_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri)
+{
+   unsigned char* result;
+   stbi__jpeg* j = (stbi__jpeg*) stbi__malloc(sizeof(stbi__jpeg));
+   if (!j) return stbi__errpuc("outofmem", "Out of memory");
+   memset(j, 0, sizeof(stbi__jpeg));
+   STBI_NOTUSED(ri);
+   j->s = s;
+   stbi__setup_jpeg(j);
+   result = load_jpeg_image(j, x,y,comp,req_comp);
+   STBI_FREE(j);
+   return result;
+}
+
+static int stbi__jpeg_test(stbi__context *s)
+{
+   int r;
+   stbi__jpeg* j = (stbi__jpeg*)stbi__malloc(sizeof(stbi__jpeg));
+   if (!j) return stbi__err("outofmem", "Out of memory");
+   memset(j, 0, sizeof(stbi__jpeg));
+   j->s = s;
+   stbi__setup_jpeg(j);
+   r = stbi__decode_jpeg_header(j, STBI__SCAN_type);
+   stbi__rewind(s);
+   STBI_FREE(j);
+   return r;
+}
+
+static int stbi__jpeg_info_raw(stbi__jpeg *j, int *x, int *y, int *comp)
+{
+   if (!stbi__decode_jpeg_header(j, STBI__SCAN_header)) {
+      stbi__rewind( j->s );
+      return 0;
+   }
+   if (x) *x = j->s->img_x;
+   if (y) *y = j->s->img_y;
+   if (comp) *comp = j->s->img_n >= 3 ? 3 : 1;
+   return 1;
+}
+
+static int stbi__jpeg_info(stbi__context *s, int *x, int *y, int *comp)
+{
+   int result;
+   stbi__jpeg* j = (stbi__jpeg*) (stbi__malloc(sizeof(stbi__jpeg)));
+   if (!j) return stbi__err("outofmem", "Out of memory");
+   memset(j, 0, sizeof(stbi__jpeg));
+   j->s = s;
+   result = stbi__jpeg_info_raw(j, x, y, comp);
+   STBI_FREE(j);
+   return result;
+}
+#endif
+
+// public domain zlib decode    v0.2  Sean Barrett 2006-11-18
+//    simple implementation
+//      - all input must be provided in an upfront buffer
+//      - all output is written to a single output buffer (can malloc/realloc)
+//    performance
+//      - fast huffman
+
+#ifndef STBI_NO_ZLIB
+
+// fast-way is faster to check than jpeg huffman, but slow way is slower
+#define STBI__ZFAST_BITS  9 // accelerate all cases in default tables
+#define STBI__ZFAST_MASK  ((1 << STBI__ZFAST_BITS) - 1)
+#define STBI__ZNSYMS 288 // number of symbols in literal/length alphabet
+
+// zlib-style huffman encoding
+// (jpegs packs from left, zlib from right, so can't share code)
+typedef struct
+{
+   stbi__uint16 fast[1 << STBI__ZFAST_BITS];
+   stbi__uint16 firstcode[16];
+   int maxcode[17];
+   stbi__uint16 firstsymbol[16];
+   stbi_uc  size[STBI__ZNSYMS];
+   stbi__uint16 value[STBI__ZNSYMS];
+} stbi__zhuffman;
+
+stbi_inline static int stbi__bitreverse16(int n)
+{
+  n = ((n & 0xAAAA) >>  1) | ((n & 0x5555) << 1);
+  n = ((n & 0xCCCC) >>  2) | ((n & 0x3333) << 2);
+  n = ((n & 0xF0F0) >>  4) | ((n & 0x0F0F) << 4);
+  n = ((n & 0xFF00) >>  8) | ((n & 0x00FF) << 8);
+  return n;
+}
+
+stbi_inline static int stbi__bit_reverse(int v, int bits)
+{
+   STBI_ASSERT(bits <= 16);
+   // to bit reverse n bits, reverse 16 and shift
+   // e.g. 11 bits, bit reverse and shift away 5
+   return stbi__bitreverse16(v) >> (16-bits);
+}
+
+static int stbi__zbuild_huffman(stbi__zhuffman *z, const stbi_uc *sizelist, int num)
+{
+   int i,k=0;
+   int code, next_code[16], sizes[17];
+
+   // DEFLATE spec for generating codes
+   memset(sizes, 0, sizeof(sizes));
+   memset(z->fast, 0, sizeof(z->fast));
+   for (i=0; i < num; ++i)
+      ++sizes[sizelist[i]];
+   sizes[0] = 0;
+   for (i=1; i < 16; ++i)
+      if (sizes[i] > (1 << i))
+         return stbi__err("bad sizes", "Corrupt PNG");
+   code = 0;
+   for (i=1; i < 16; ++i) {
+      next_code[i] = code;
+      z->firstcode[i] = (stbi__uint16) code;
+      z->firstsymbol[i] = (stbi__uint16) k;
+      code = (code + sizes[i]);
+      if (sizes[i])
+         if (code-1 >= (1 << i)) return stbi__err("bad codelengths","Corrupt PNG");
+      z->maxcode[i] = code << (16-i); // preshift for inner loop
+      code <<= 1;
+      k += sizes[i];
+   }
+   z->maxcode[16] = 0x10000; // sentinel
+   for (i=0; i < num; ++i) {
+      int s = sizelist[i];
+      if (s) {
+         int c = next_code[s] - z->firstcode[s] + z->firstsymbol[s];
+         stbi__uint16 fastv = (stbi__uint16) ((s << 9) | i);
+         z->size [c] = (stbi_uc     ) s;
+         z->value[c] = (stbi__uint16) i;
+         if (s <= STBI__ZFAST_BITS) {
+            int j = stbi__bit_reverse(next_code[s],s);
+            while (j < (1 << STBI__ZFAST_BITS)) {
+               z->fast[j] = fastv;
+               j += (1 << s);
+            }
+         }
+         ++next_code[s];
+      }
+   }
+   return 1;
+}
+
+// zlib-from-memory implementation for PNG reading
+//    because PNG allows splitting the zlib stream arbitrarily,
+//    and it's annoying structurally to have PNG call ZLIB call PNG,
+//    we require PNG read all the IDATs and combine them into a single
+//    memory buffer
+
+typedef struct
+{
+   stbi_uc *zbuffer, *zbuffer_end;
+   int num_bits;
+   stbi__uint32 code_buffer;
+
+   char *zout;
+   char *zout_start;
+   char *zout_end;
+   int   z_expandable;
+
+   stbi__zhuffman z_length, z_distance;
+} stbi__zbuf;
+
+stbi_inline static int stbi__zeof(stbi__zbuf *z)
+{
+   return (z->zbuffer >= z->zbuffer_end);
+}
+
+stbi_inline static stbi_uc stbi__zget8(stbi__zbuf *z)
+{
+   return stbi__zeof(z) ? 0 : *z->zbuffer++;
+}
+
+static void stbi__fill_bits(stbi__zbuf *z)
+{
+   do {
+      if (z->code_buffer >= (1U << z->num_bits)) {
+        z->zbuffer = z->zbuffer_end;  /* treat this as EOF so we fail. */
+        return;
+      }
+      z->code_buffer |= (unsigned int) stbi__zget8(z) << z->num_bits;
+      z->num_bits += 8;
+   } while (z->num_bits <= 24);
+}
+
+stbi_inline static unsigned int stbi__zreceive(stbi__zbuf *z, int n)
+{
+   unsigned int k;
+   if (z->num_bits < n) stbi__fill_bits(z);
+   k = z->code_buffer & ((1 << n) - 1);
+   z->code_buffer >>= n;
+   z->num_bits -= n;
+   return k;
+}
+
+static int stbi__zhuffman_decode_slowpath(stbi__zbuf *a, stbi__zhuffman *z)
+{
+   int b,s,k;
+   // not resolved by fast table, so compute it the slow way
+   // use jpeg approach, which requires MSbits at top
+   k = stbi__bit_reverse(a->code_buffer, 16);
+   for (s=STBI__ZFAST_BITS+1; ; ++s)
+      if (k < z->maxcode[s])
+         break;
+   if (s >= 16) return -1; // invalid code!
+   // code size is s, so:
+   b = (k >> (16-s)) - z->firstcode[s] + z->firstsymbol[s];
+   if (b >= STBI__ZNSYMS) return -1; // some data was corrupt somewhere!
+   if (z->size[b] != s) return -1;  // was originally an assert, but report failure instead.
+   a->code_buffer >>= s;
+   a->num_bits -= s;
+   return z->value[b];
+}
+
+stbi_inline static int stbi__zhuffman_decode(stbi__zbuf *a, stbi__zhuffman *z)
+{
+   int b,s;
+   if (a->num_bits < 16) {
+      if (stbi__zeof(a)) {
+         return -1;   /* report error for unexpected end of data. */
+      }
+      stbi__fill_bits(a);
+   }
+   b = z->fast[a->code_buffer & STBI__ZFAST_MASK];
+   if (b) {
+      s = b >> 9;
+      a->code_buffer >>= s;
+      a->num_bits -= s;
+      return b & 511;
+   }
+   return stbi__zhuffman_decode_slowpath(a, z);
+}
+
+static int stbi__zexpand(stbi__zbuf *z, char *zout, int n)  // need to make room for n bytes
+{
+   char *q;
+   unsigned int cur, limit, old_limit;
+   z->zout = zout;
+   if (!z->z_expandable) return stbi__err("output buffer limit","Corrupt PNG");
+   cur   = (unsigned int) (z->zout - z->zout_start);
+   limit = old_limit = (unsigned) (z->zout_end - z->zout_start);
+   if (UINT_MAX - cur < (unsigned) n) return stbi__err("outofmem", "Out of memory");
+   while (cur + n > limit) {
+      if(limit > UINT_MAX / 2) return stbi__err("outofmem", "Out of memory");
+      limit *= 2;
+   }
+   q = (char *) STBI_REALLOC_SIZED(z->zout_start, old_limit, limit);
+   STBI_NOTUSED(old_limit);
+   if (q == NULL) return stbi__err("outofmem", "Out of memory");
+   z->zout_start = q;
+   z->zout       = q + cur;
+   z->zout_end   = q + limit;
+   return 1;
+}
+
+static const int stbi__zlength_base[31] = {
+   3,4,5,6,7,8,9,10,11,13,
+   15,17,19,23,27,31,35,43,51,59,
+   67,83,99,115,131,163,195,227,258,0,0 };
+
+static const int stbi__zlength_extra[31]=
+{ 0,0,0,0,0,0,0,0,1,1,1,1,2,2,2,2,3,3,3,3,4,4,4,4,5,5,5,5,0,0,0 };
+
+static const int stbi__zdist_base[32] = { 1,2,3,4,5,7,9,13,17,25,33,49,65,97,129,193,
+257,385,513,769,1025,1537,2049,3073,4097,6145,8193,12289,16385,24577,0,0};
+
+static const int stbi__zdist_extra[32] =
+{ 0,0,0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7,8,8,9,9,10,10,11,11,12,12,13,13};
+
+static int stbi__parse_huffman_block(stbi__zbuf *a)
+{
+   char *zout = a->zout;
+   for(;;) {
+      int z = stbi__zhuffman_decode(a, &a->z_length);
+      if (z < 256) {
+         if (z < 0) return stbi__err("bad huffman code","Corrupt PNG"); // error in huffman codes
+         if (zout >= a->zout_end) {
+            if (!stbi__zexpand(a, zout, 1)) return 0;
+            zout = a->zout;
+         }
+         *zout++ = (char) z;
+      } else {
+         stbi_uc *p;
+         int len,dist;
+         if (z == 256) {
+            a->zout = zout;
+            return 1;
+         }
+         if (z >= 286) return stbi__err("bad huffman code","Corrupt PNG"); // per DEFLATE, length codes 286 and 287 must not appear in compressed data
+         z -= 257;
+         len = stbi__zlength_base[z];
+         if (stbi__zlength_extra[z]) len += stbi__zreceive(a, stbi__zlength_extra[z]);
+         z = stbi__zhuffman_decode(a, &a->z_distance);
+         if (z < 0 || z >= 30) return stbi__err("bad huffman code","Corrupt PNG"); // per DEFLATE, distance codes 30 and 31 must not appear in compressed data
+         dist = stbi__zdist_base[z];
+         if (stbi__zdist_extra[z]) dist += stbi__zreceive(a, stbi__zdist_extra[z]);
+         if (zout - a->zout_start < dist) return stbi__err("bad dist","Corrupt PNG");
+         if (zout + len > a->zout_end) {
+            if (!stbi__zexpand(a, zout, len)) return 0;
+            zout = a->zout;
+         }
+         p = (stbi_uc *) (zout - dist);
+         if (dist == 1) { // run of one byte; common in images.
+            stbi_uc v = *p;
+            if (len) { do *zout++ = v; while (--len); }
+         } else {
+            if (len) { do *zout++ = *p++; while (--len); }
+         }
+      }
+   }
+}
+
+static int stbi__compute_huffman_codes(stbi__zbuf *a)
+{
+   static const stbi_uc length_dezigzag[19] = { 16,17,18,0,8,7,9,6,10,5,11,4,12,3,13,2,14,1,15 };
+   stbi__zhuffman z_codelength;
+   stbi_uc lencodes[286+32+137];//padding for maximum single op
+   stbi_uc codelength_sizes[19];
+   int i,n;
+
+   int hlit  = stbi__zreceive(a,5) + 257;
+   int hdist = stbi__zreceive(a,5) + 1;
+   int hclen = stbi__zreceive(a,4) + 4;
+   int ntot  = hlit + hdist;
+
+   memset(codelength_sizes, 0, sizeof(codelength_sizes));
+   for (i=0; i < hclen; ++i) {
+      int s = stbi__zreceive(a,3);
+      codelength_sizes[length_dezigzag[i]] = (stbi_uc) s;
+   }
+   if (!stbi__zbuild_huffman(&z_codelength, codelength_sizes, 19)) return 0;
+
+   n = 0;
+   while (n < ntot) {
+      int c = stbi__zhuffman_decode(a, &z_codelength);
+      if (c < 0 || c >= 19) return stbi__err("bad codelengths", "Corrupt PNG");
+      if (c < 16)
+         lencodes[n++] = (stbi_uc) c;
+      else {
+         stbi_uc fill = 0;
+         if (c == 16) {
+            c = stbi__zreceive(a,2)+3;
+            if (n == 0) return stbi__err("bad codelengths", "Corrupt PNG");
+            fill = lencodes[n-1];
+         } else if (c == 17) {
+            c = stbi__zreceive(a,3)+3;
+         } else if (c == 18) {
+            c = stbi__zreceive(a,7)+11;
+         } else {
+            return stbi__err("bad codelengths", "Corrupt PNG");
+         }
+         if (ntot - n < c) return stbi__err("bad codelengths", "Corrupt PNG");
+         memset(lencodes+n, fill, c);
+         n += c;
+      }
+   }
+   if (n != ntot) return stbi__err("bad codelengths","Corrupt PNG");
+   if (!stbi__zbuild_huffman(&a->z_length, lencodes, hlit)) return 0;
+   if (!stbi__zbuild_huffman(&a->z_distance, lencodes+hlit, hdist)) return 0;
+   return 1;
+}
+
+static int stbi__parse_uncompressed_block(stbi__zbuf *a)
+{
+   stbi_uc header[4];
+   int len,nlen,k;
+   if (a->num_bits & 7)
+      stbi__zreceive(a, a->num_bits & 7); // discard
+   // drain the bit-packed data into header
+   k = 0;
+   while (a->num_bits > 0) {
+      header[k++] = (stbi_uc) (a->code_buffer & 255); // suppress MSVC run-time check
+      a->code_buffer >>= 8;
+      a->num_bits -= 8;
+   }
+   if (a->num_bits < 0) return stbi__err("zlib corrupt","Corrupt PNG");
+   // now fill header the normal way
+   while (k < 4)
+      header[k++] = stbi__zget8(a);
+   len  = header[1] * 256 + header[0];
+   nlen = header[3] * 256 + header[2];
+   if (nlen != (len ^ 0xffff)) return stbi__err("zlib corrupt","Corrupt PNG");
+   if (a->zbuffer + len > a->zbuffer_end) return stbi__err("read past buffer","Corrupt PNG");
+   if (a->zout + len > a->zout_end)
+      if (!stbi__zexpand(a, a->zout, len)) return 0;
+   memcpy(a->zout, a->zbuffer, len);
+   a->zbuffer += len;
+   a->zout += len;
+   return 1;
+}
+
+static int stbi__parse_zlib_header(stbi__zbuf *a)
+{
+   int cmf   = stbi__zget8(a);
+   int cm    = cmf & 15;
+   /* int cinfo = cmf >> 4; */
+   int flg   = stbi__zget8(a);
+   if (stbi__zeof(a)) return stbi__err("bad zlib header","Corrupt PNG"); // zlib spec
+   if ((cmf*256+flg) % 31 != 0) return stbi__err("bad zlib header","Corrupt PNG"); // zlib spec
+   if (flg & 32) return stbi__err("no preset dict","Corrupt PNG"); // preset dictionary not allowed in png
+   if (cm != 8) return stbi__err("bad compression","Corrupt PNG"); // DEFLATE required for png
+   // window = 1 << (8 + cinfo)... but who cares, we fully buffer output
+   return 1;
+}
+
+static const stbi_uc stbi__zdefault_length[STBI__ZNSYMS] =
+{
+   8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8, 8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,
+   8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8, 8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,
+   8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8, 8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,
+   8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8, 8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,
+   8,8,8,8,8,8,8,8,8,8,8,8,8,8,8,8, 9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,
+   9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,9, 9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,
+   9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,9, 9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,
+   9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,9, 9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,9,
+   7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7, 7,7,7,7,7,7,7,7,8,8,8,8,8,8,8,8
+};
+static const stbi_uc stbi__zdefault_distance[32] =
+{
+   5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5,5
+};
+/*
+Init algorithm:
+{
+   int i;   // use <= to match clearly with spec
+   for (i=0; i <= 143; ++i)     stbi__zdefault_length[i]   = 8;
+   for (   ; i <= 255; ++i)     stbi__zdefault_length[i]   = 9;
+   for (   ; i <= 279; ++i)     stbi__zdefault_length[i]   = 7;
+   for (   ; i <= 287; ++i)     stbi__zdefault_length[i]   = 8;
+
+   for (i=0; i <=  31; ++i)     stbi__zdefault_distance[i] = 5;
+}
+*/
+
+static int stbi__parse_zlib(stbi__zbuf *a, int parse_header)
+{
+   int final, type;
+   if (parse_header)
+      if (!stbi__parse_zlib_header(a)) return 0;
+   a->num_bits = 0;
+   a->code_buffer = 0;
+   do {
+      final = stbi__zreceive(a,1);
+      type = stbi__zreceive(a,2);
+      if (type == 0) {
+         if (!stbi__parse_uncompressed_block(a)) return 0;
+      } else if (type == 3) {
+         return 0;
+      } else {
+         if (type == 1) {
+            // use fixed code lengths
+            if (!stbi__zbuild_huffman(&a->z_length  , stbi__zdefault_length  , STBI__ZNSYMS)) return 0;
+            if (!stbi__zbuild_huffman(&a->z_distance, stbi__zdefault_distance,  32)) return 0;
+         } else {
+            if (!stbi__compute_huffman_codes(a)) return 0;
+         }
+         if (!stbi__parse_huffman_block(a)) return 0;
+      }
+   } while (!final);
+   return 1;
+}
+
+static int stbi__do_zlib(stbi__zbuf *a, char *obuf, int olen, int exp, int parse_header)
+{
+   a->zout_start = obuf;
+   a->zout       = obuf;
+   a->zout_end   = obuf + olen;
+   a->z_expandable = exp;
+
+   return stbi__parse_zlib(a, parse_header);
+}
+
+STBIDEF char *stbi_zlib_decode_malloc_guesssize(const char *buffer, int len, int initial_size, int *outlen)
+{
+   stbi__zbuf a;
+   char *p = (char *) stbi__malloc(initial_size);
+   if (p == NULL) return NULL;
+   a.zbuffer = (stbi_uc *) buffer;
+   a.zbuffer_end = (stbi_uc *) buffer + len;
+   if (stbi__do_zlib(&a, p, initial_size, 1, 1)) {
+      if (outlen) *outlen = (int) (a.zout - a.zout_start);
+      return a.zout_start;
+   } else {
+      STBI_FREE(a.zout_start);
+      return NULL;
+   }
+}
+
+STBIDEF char *stbi_zlib_decode_malloc(char const *buffer, int len, int *outlen)
+{
+   return stbi_zlib_decode_malloc_guesssize(buffer, len, 16384, outlen);
+}
+
+STBIDEF char *stbi_zlib_decode_malloc_guesssize_headerflag(const char *buffer, int len, int initial_size, int *outlen, int parse_header)
+{
+   stbi__zbuf a;
+   char *p = (char *) stbi__malloc(initial_size);
+   if (p == NULL) return NULL;
+   a.zbuffer = (stbi_uc *) buffer;
+   a.zbuffer_end = (stbi_uc *) buffer + len;
+   if (stbi__do_zlib(&a, p, initial_size, 1, parse_header)) {
+      if (outlen) *outlen = (int) (a.zout - a.zout_start);
+      return a.zout_start;
+   } else {
+      STBI_FREE(a.zout_start);
+      return NULL;
+   }
+}
+
+STBIDEF int stbi_zlib_decode_buffer(char *obuffer, int olen, char const *ibuffer, int ilen)
+{
+   stbi__zbuf a;
+   a.zbuffer = (stbi_uc *) ibuffer;
+   a.zbuffer_end = (stbi_uc *) ibuffer + ilen;
+   if (stbi__do_zlib(&a, obuffer, olen, 0, 1))
+      return (int) (a.zout - a.zout_start);
+   else
+      return -1;
+}
+
+STBIDEF char *stbi_zlib_decode_noheader_malloc(char const *buffer, int len, int *outlen)
+{
+   stbi__zbuf a;
+   char *p = (char *) stbi__malloc(16384);
+   if (p == NULL) return NULL;
+   a.zbuffer = (stbi_uc *) buffer;
+   a.zbuffer_end = (stbi_uc *) buffer+len;
+   if (stbi__do_zlib(&a, p, 16384, 1, 0)) {
+      if (outlen) *outlen = (int) (a.zout - a.zout_start);
+      return a.zout_start;
+   } else {
+      STBI_FREE(a.zout_start);
+      return NULL;
+   }
+}
+
+STBIDEF int stbi_zlib_decode_noheader_buffer(char *obuffer, int olen, const char *ibuffer, int ilen)
+{
+   stbi__zbuf a;
+   a.zbuffer = (stbi_uc *) ibuffer;
+   a.zbuffer_end = (stbi_uc *) ibuffer + ilen;
+   if (stbi__do_zlib(&a, obuffer, olen, 0, 0))
+      return (int) (a.zout - a.zout_start);
+   else
+      return -1;
+}
+#endif
+
+// public domain "baseline" PNG decoder   v0.10  Sean Barrett 2006-11-18
+//    simple implementation
+//      - only 8-bit samples
+//      - no CRC checking
+//      - allocates lots of intermediate memory
+//        - avoids problem of streaming data between subsystems
+//        - avoids explicit window management
+//    performance
+//      - uses stb_zlib, a PD zlib implementation with fast huffman decoding
+
+#ifndef STBI_NO_PNG
+typedef struct
+{
+   stbi__uint32 length;
+   stbi__uint32 type;
+} stbi__pngchunk;
+
+static stbi__pngchunk stbi__get_chunk_header(stbi__context *s)
+{
+   stbi__pngchunk c;
+   c.length = stbi__get32be(s);
+   c.type   = stbi__get32be(s);
+   return c;
+}
+
+static int stbi__check_png_header(stbi__context *s)
+{
+   static const stbi_uc png_sig[8] = { 137,80,78,71,13,10,26,10 };
+   int i;
+   for (i=0; i < 8; ++i)
+      if (stbi__get8(s) != png_sig[i]) return stbi__err("bad png sig","Not a PNG");
+   return 1;
+}
+
+typedef struct
+{
+   stbi__context *s;
+   stbi_uc *idata, *expanded, *out;
+   int depth;
+} stbi__png;
+
+
+enum {
+   STBI__F_none=0,
+   STBI__F_sub=1,
+   STBI__F_up=2,
+   STBI__F_avg=3,
+   STBI__F_paeth=4,
+   // synthetic filters used for first scanline to avoid needing a dummy row of 0s
+   STBI__F_avg_first,
+   STBI__F_paeth_first
+};
+
+static stbi_uc first_row_filter[5] =
+{
+   STBI__F_none,
+   STBI__F_sub,
+   STBI__F_none,
+   STBI__F_avg_first,
+   STBI__F_paeth_first
+};
+
+static int stbi__paeth(int a, int b, int c)
+{
+   int p = a + b - c;
+   int pa = abs(p-a);
+   int pb = abs(p-b);
+   int pc = abs(p-c);
+   if (pa <= pb && pa <= pc) return a;
+   if (pb <= pc) return b;
+   return c;
+}
+
+static const stbi_uc stbi__depth_scale_table[9] = { 0, 0xff, 0x55, 0, 0x11, 0,0,0, 0x01 };
+
+// create the png data from post-deflated data
+static int stbi__create_png_image_raw(stbi__png *a, stbi_uc *raw, stbi__uint32 raw_len, int out_n, stbi__uint32 x, stbi__uint32 y, int depth, int color)
+{
+   int bytes = (depth == 16? 2 : 1);
+   stbi__context *s = a->s;
+   stbi__uint32 i,j,stride = x*out_n*bytes;
+   stbi__uint32 img_len, img_width_bytes;
+   int k;
+   int img_n = s->img_n; // copy it into a local for later
+
+   int output_bytes = out_n*bytes;
+   int filter_bytes = img_n*bytes;
+   int width = x;
+
+   STBI_ASSERT(out_n == s->img_n || out_n == s->img_n+1);
+   a->out = (stbi_uc *) stbi__malloc_mad3(x, y, output_bytes, 0); // extra bytes to write off the end into
+   if (!a->out) return stbi__err("outofmem", "Out of memory");
+
+   if (!stbi__mad3sizes_valid(img_n, x, depth, 7)) return stbi__err("too large", "Corrupt PNG");
+   img_width_bytes = (((img_n * x * depth) + 7) >> 3);
+   img_len = (img_width_bytes + 1) * y;
+
+   // we used to check for exact match between raw_len and img_len on non-interlaced PNGs,
+   // but issue #276 reported a PNG in the wild that had extra data at the end (all zeros),
+   // so just check for raw_len < img_len always.
+   if (raw_len < img_len) return stbi__err("not enough pixels","Corrupt PNG");
+
+   for (j=0; j < y; ++j) {
+      stbi_uc *cur = a->out + stride*j;
+      stbi_uc *prior;
+      int filter = *raw++;
+
+      if (filter > 4)
+         return stbi__err("invalid filter","Corrupt PNG");
+
+      if (depth < 8) {
+         if (img_width_bytes > x) return stbi__err("invalid width","Corrupt PNG");
+         cur += x*out_n - img_width_bytes; // store output to the rightmost img_len bytes, so we can decode in place
+         filter_bytes = 1;
+         width = img_width_bytes;
+      }
+      prior = cur - stride; // bugfix: need to compute this after 'cur +=' computation above
+
+      // if first row, use special filter that doesn't sample previous row
+      if (j == 0) filter = first_row_filter[filter];
+
+      // handle first byte explicitly
+      for (k=0; k < filter_bytes; ++k) {
+         switch (filter) {
+            case STBI__F_none       : cur[k] = raw[k]; break;
+            case STBI__F_sub        : cur[k] = raw[k]; break;
+            case STBI__F_up         : cur[k] = STBI__BYTECAST(raw[k] + prior[k]); break;
+            case STBI__F_avg        : cur[k] = STBI__BYTECAST(raw[k] + (prior[k]>>1)); break;
+            case STBI__F_paeth      : cur[k] = STBI__BYTECAST(raw[k] + stbi__paeth(0,prior[k],0)); break;
+            case STBI__F_avg_first  : cur[k] = raw[k]; break;
+            case STBI__F_paeth_first: cur[k] = raw[k]; break;
+         }
+      }
+
+      if (depth == 8) {
+         if (img_n != out_n)
+            cur[img_n] = 255; // first pixel
+         raw += img_n;
+         cur += out_n;
+         prior += out_n;
+      } else if (depth == 16) {
+         if (img_n != out_n) {
+            cur[filter_bytes]   = 255; // first pixel top byte
+            cur[filter_bytes+1] = 255; // first pixel bottom byte
+         }
+         raw += filter_bytes;
+         cur += output_bytes;
+         prior += output_bytes;
+      } else {
+         raw += 1;
+         cur += 1;
+         prior += 1;
+      }
+
+      // this is a little gross, so that we don't switch per-pixel or per-component
+      if (depth < 8 || img_n == out_n) {
+         int nk = (width - 1)*filter_bytes;
+         #define STBI__CASE(f) \
+             case f:     \
+                for (k=0; k < nk; ++k)
+         switch (filter) {
+            // "none" filter turns into a memcpy here; make that explicit.
+            case STBI__F_none:         memcpy(cur, raw, nk); break;
+            STBI__CASE(STBI__F_sub)          { cur[k] = STBI__BYTECAST(raw[k] + cur[k-filter_bytes]); } break;
+            STBI__CASE(STBI__F_up)           { cur[k] = STBI__BYTECAST(raw[k] + prior[k]); } break;
+            STBI__CASE(STBI__F_avg)          { cur[k] = STBI__BYTECAST(raw[k] + ((prior[k] + cur[k-filter_bytes])>>1)); } break;
+            STBI__CASE(STBI__F_paeth)        { cur[k] = STBI__BYTECAST(raw[k] + stbi__paeth(cur[k-filter_bytes],prior[k],prior[k-filter_bytes])); } break;
+            STBI__CASE(STBI__F_avg_first)    { cur[k] = STBI__BYTECAST(raw[k] + (cur[k-filter_bytes] >> 1)); } break;
+            STBI__CASE(STBI__F_paeth_first)  { cur[k] = STBI__BYTECAST(raw[k] + stbi__paeth(cur[k-filter_bytes],0,0)); } break;
+         }
+         #undef STBI__CASE
+         raw += nk;
+      } else {
+         STBI_ASSERT(img_n+1 == out_n);
+         #define STBI__CASE(f) \
+             case f:     \
+                for (i=x-1; i >= 1; --i, cur[filter_bytes]=255,raw+=filter_bytes,cur+=output_bytes,prior+=output_bytes) \
+                   for (k=0; k < filter_bytes; ++k)
+         switch (filter) {
+            STBI__CASE(STBI__F_none)         { cur[k] = raw[k]; } break;
+            STBI__CASE(STBI__F_sub)          { cur[k] = STBI__BYTECAST(raw[k] + cur[k- output_bytes]); } break;
+            STBI__CASE(STBI__F_up)           { cur[k] = STBI__BYTECAST(raw[k] + prior[k]); } break;
+            STBI__CASE(STBI__F_avg)          { cur[k] = STBI__BYTECAST(raw[k] + ((prior[k] + cur[k- output_bytes])>>1)); } break;
+            STBI__CASE(STBI__F_paeth)        { cur[k] = STBI__BYTECAST(raw[k] + stbi__paeth(cur[k- output_bytes],prior[k],prior[k- output_bytes])); } break;
+            STBI__CASE(STBI__F_avg_first)    { cur[k] = STBI__BYTECAST(raw[k] + (cur[k- output_bytes] >> 1)); } break;
+            STBI__CASE(STBI__F_paeth_first)  { cur[k] = STBI__BYTECAST(raw[k] + stbi__paeth(cur[k- output_bytes],0,0)); } break;
+         }
+         #undef STBI__CASE
+
+         // the loop above sets the high byte of the pixels' alpha, but for
+         // 16 bit png files we also need the low byte set. we'll do that here.
+         if (depth == 16) {
+            cur = a->out + stride*j; // start at the beginning of the row again
+            for (i=0; i < x; ++i,cur+=output_bytes) {
+               cur[filter_bytes+1] = 255;
+            }
+         }
+      }
+   }
+
+   // we make a separate pass to expand bits to pixels; for performance,
+   // this could run two scanlines behind the above code, so it won't
+   // intefere with filtering but will still be in the cache.
+   if (depth < 8) {
+      for (j=0; j < y; ++j) {
+         stbi_uc *cur = a->out + stride*j;
+         stbi_uc *in  = a->out + stride*j + x*out_n - img_width_bytes;
+         // unpack 1/2/4-bit into a 8-bit buffer. allows us to keep the common 8-bit path optimal at minimal cost for 1/2/4-bit
+         // png guarante byte alignment, if width is not multiple of 8/4/2 we'll decode dummy trailing data that will be skipped in the later loop
+         stbi_uc scale = (color == 0) ? stbi__depth_scale_table[depth] : 1; // scale grayscale values to 0..255 range
+
+         // note that the final byte might overshoot and write more data than desired.
+         // we can allocate enough data that this never writes out of memory, but it
+         // could also overwrite the next scanline. can it overwrite non-empty data
+         // on the next scanline? yes, consider 1-pixel-wide scanlines with 1-bit-per-pixel.
+         // so we need to explicitly clamp the final ones
+
+         if (depth == 4) {
+            for (k=x*img_n; k >= 2; k-=2, ++in) {
+               *cur++ = scale * ((*in >> 4)       );
+               *cur++ = scale * ((*in     ) & 0x0f);
+            }
+            if (k > 0) *cur++ = scale * ((*in >> 4)       );
+         } else if (depth == 2) {
+            for (k=x*img_n; k >= 4; k-=4, ++in) {
+               *cur++ = scale * ((*in >> 6)       );
+               *cur++ = scale * ((*in >> 4) & 0x03);
+               *cur++ = scale * ((*in >> 2) & 0x03);
+               *cur++ = scale * ((*in     ) & 0x03);
+            }
+            if (k > 0) *cur++ = scale * ((*in >> 6)       );
+            if (k > 1) *cur++ = scale * ((*in >> 4) & 0x03);
+            if (k > 2) *cur++ = scale * ((*in >> 2) & 0x03);
+         } else if (depth == 1) {
+            for (k=x*img_n; k >= 8; k-=8, ++in) {
+               *cur++ = scale * ((*in >> 7)       );
+               *cur++ = scale * ((*in >> 6) & 0x01);
+               *cur++ = scale * ((*in >> 5) & 0x01);
+               *cur++ = scale * ((*in >> 4) & 0x01);
+               *cur++ = scale * ((*in >> 3) & 0x01);
+               *cur++ = scale * ((*in >> 2) & 0x01);
+               *cur++ = scale * ((*in >> 1) & 0x01);
+               *cur++ = scale * ((*in     ) & 0x01);
+            }
+            if (k > 0) *cur++ = scale * ((*in >> 7)       );
+            if (k > 1) *cur++ = scale * ((*in >> 6) & 0x01);
+            if (k > 2) *cur++ = scale * ((*in >> 5) & 0x01);
+            if (k > 3) *cur++ = scale * ((*in >> 4) & 0x01);
+            if (k > 4) *cur++ = scale * ((*in >> 3) & 0x01);
+            if (k > 5) *cur++ = scale * ((*in >> 2) & 0x01);
+            if (k > 6) *cur++ = scale * ((*in >> 1) & 0x01);
+         }
+         if (img_n != out_n) {
+            int q;
+            // insert alpha = 255
+            cur = a->out + stride*j;
+            if (img_n == 1) {
+               for (q=x-1; q >= 0; --q) {
+                  cur[q*2+1] = 255;
+                  cur[q*2+0] = cur[q];
+               }
+            } else {
+               STBI_ASSERT(img_n == 3);
+               for (q=x-1; q >= 0; --q) {
+                  cur[q*4+3] = 255;
+                  cur[q*4+2] = cur[q*3+2];
+                  cur[q*4+1] = cur[q*3+1];
+                  cur[q*4+0] = cur[q*3+0];
+               }
+            }
+         }
+      }
+   } else if (depth == 16) {
+      // force the image data from big-endian to platform-native.
+      // this is done in a separate pass due to the decoding relying
+      // on the data being untouched, but could probably be done
+      // per-line during decode if care is taken.
+      stbi_uc *cur = a->out;
+      stbi__uint16 *cur16 = (stbi__uint16*)cur;
+
+      for(i=0; i < x*y*out_n; ++i,cur16++,cur+=2) {
+         *cur16 = (cur[0] << 8) | cur[1];
+      }
+   }
+
+   return 1;
+}
+
+static int stbi__create_png_image(stbi__png *a, stbi_uc *image_data, stbi__uint32 image_data_len, int out_n, int depth, int color, int interlaced)
+{
+   int bytes = (depth == 16 ? 2 : 1);
+   int out_bytes = out_n * bytes;
+   stbi_uc *final;
+   int p;
+   if (!interlaced)
+      return stbi__create_png_image_raw(a, image_data, image_data_len, out_n, a->s->img_x, a->s->img_y, depth, color);
+
+   // de-interlacing
+   final = (stbi_uc *) stbi__malloc_mad3(a->s->img_x, a->s->img_y, out_bytes, 0);
+   if (!final) return stbi__err("outofmem", "Out of memory");
+   for (p=0; p < 7; ++p) {
+      int xorig[] = { 0,4,0,2,0,1,0 };
+      int yorig[] = { 0,0,4,0,2,0,1 };
+      int xspc[]  = { 8,8,4,4,2,2,1 };
+      int yspc[]  = { 8,8,8,4,4,2,2 };
+      int i,j,x,y;
+      // pass1_x[4] = 0, pass1_x[5] = 1, pass1_x[12] = 1
+      x = (a->s->img_x - xorig[p] + xspc[p]-1) / xspc[p];
+      y = (a->s->img_y - yorig[p] + yspc[p]-1) / yspc[p];
+      if (x && y) {
+         stbi__uint32 img_len = ((((a->s->img_n * x * depth) + 7) >> 3) + 1) * y;
+         if (!stbi__create_png_image_raw(a, image_data, image_data_len, out_n, x, y, depth, color)) {
+            STBI_FREE(final);
+            return 0;
+         }
+         for (j=0; j < y; ++j) {
+            for (i=0; i < x; ++i) {
+               int out_y = j*yspc[p]+yorig[p];
+               int out_x = i*xspc[p]+xorig[p];
+               memcpy(final + out_y*a->s->img_x*out_bytes + out_x*out_bytes,
+                      a->out + (j*x+i)*out_bytes, out_bytes);
+            }
+         }
+         STBI_FREE(a->out);
+         image_data += img_len;
+         image_data_len -= img_len;
+      }
+   }
+   a->out = final;
+
+   return 1;
+}
+
+static int stbi__compute_transparency(stbi__png *z, stbi_uc tc[3], int out_n)
+{
+   stbi__context *s = z->s;
+   stbi__uint32 i, pixel_count = s->img_x * s->img_y;
+   stbi_uc *p = z->out;
+
+   // compute color-based transparency, assuming we've
+   // already got 255 as the alpha value in the output
+   STBI_ASSERT(out_n == 2 || out_n == 4);
+
+   if (out_n == 2) {
+      for (i=0; i < pixel_count; ++i) {
+         p[1] = (p[0] == tc[0] ? 0 : 255);
+         p += 2;
+      }
+   } else {
+      for (i=0; i < pixel_count; ++i) {
+         if (p[0] == tc[0] && p[1] == tc[1] && p[2] == tc[2])
+            p[3] = 0;
+         p += 4;
+      }
+   }
+   return 1;
+}
+
+static int stbi__compute_transparency16(stbi__png *z, stbi__uint16 tc[3], int out_n)
+{
+   stbi__context *s = z->s;
+   stbi__uint32 i, pixel_count = s->img_x * s->img_y;
+   stbi__uint16 *p = (stbi__uint16*) z->out;
+
+   // compute color-based transparency, assuming we've
+   // already got 65535 as the alpha value in the output
+   STBI_ASSERT(out_n == 2 || out_n == 4);
+
+   if (out_n == 2) {
+      for (i = 0; i < pixel_count; ++i) {
+         p[1] = (p[0] == tc[0] ? 0 : 65535);
+         p += 2;
+      }
+   } else {
+      for (i = 0; i < pixel_count; ++i) {
+         if (p[0] == tc[0] && p[1] == tc[1] && p[2] == tc[2])
+            p[3] = 0;
+         p += 4;
+      }
+   }
+   return 1;
+}
+
+static int stbi__expand_png_palette(stbi__png *a, stbi_uc *palette, int len, int pal_img_n)
+{
+   stbi__uint32 i, pixel_count = a->s->img_x * a->s->img_y;
+   stbi_uc *p, *temp_out, *orig = a->out;
+
+   p = (stbi_uc *) stbi__malloc_mad2(pixel_count, pal_img_n, 0);
+   if (p == NULL) return stbi__err("outofmem", "Out of memory");
+
+   // between here and free(out) below, exitting would leak
+   temp_out = p;
+
+   if (pal_img_n == 3) {
+      for (i=0; i < pixel_count; ++i) {
+         int n = orig[i]*4;
+         p[0] = palette[n  ];
+         p[1] = palette[n+1];
+         p[2] = palette[n+2];
+         p += 3;
+      }
+   } else {
+      for (i=0; i < pixel_count; ++i) {
+         int n = orig[i]*4;
+         p[0] = palette[n  ];
+         p[1] = palette[n+1];
+         p[2] = palette[n+2];
+         p[3] = palette[n+3];
+         p += 4;
+      }
+   }
+   STBI_FREE(a->out);
+   a->out = temp_out;
+
+   STBI_NOTUSED(len);
+
+   return 1;
+}
+
+static int stbi__unpremultiply_on_load_global = 0;
+static int stbi__de_iphone_flag_global = 0;
+
+STBIDEF void stbi_set_unpremultiply_on_load(int flag_true_if_should_unpremultiply)
+{
+   stbi__unpremultiply_on_load_global = flag_true_if_should_unpremultiply;
+}
+
+STBIDEF void stbi_convert_iphone_png_to_rgb(int flag_true_if_should_convert)
+{
+   stbi__de_iphone_flag_global = flag_true_if_should_convert;
+}
+
+#ifndef STBI_THREAD_LOCAL
+#define stbi__unpremultiply_on_load  stbi__unpremultiply_on_load_global
+#define stbi__de_iphone_flag  stbi__de_iphone_flag_global
+#else
+static STBI_THREAD_LOCAL int stbi__unpremultiply_on_load_local, stbi__unpremultiply_on_load_set;
+static STBI_THREAD_LOCAL int stbi__de_iphone_flag_local, stbi__de_iphone_flag_set;
+
+STBIDEF void stbi_set_unpremultiply_on_load_thread(int flag_true_if_should_unpremultiply)
+{
+   stbi__unpremultiply_on_load_local = flag_true_if_should_unpremultiply;
+   stbi__unpremultiply_on_load_set = 1;
+}
+
+STBIDEF void stbi_convert_iphone_png_to_rgb_thread(int flag_true_if_should_convert)
+{
+   stbi__de_iphone_flag_local = flag_true_if_should_convert;
+   stbi__de_iphone_flag_set = 1;
+}
+
+#define stbi__unpremultiply_on_load  (stbi__unpremultiply_on_load_set           \
+                                       ? stbi__unpremultiply_on_load_local      \
+                                       : stbi__unpremultiply_on_load_global)
+#define stbi__de_iphone_flag  (stbi__de_iphone_flag_set                         \
+                                ? stbi__de_iphone_flag_local                    \
+                                : stbi__de_iphone_flag_global)
+#endif // STBI_THREAD_LOCAL
+
+static void stbi__de_iphone(stbi__png *z)
+{
+   stbi__context *s = z->s;
+   stbi__uint32 i, pixel_count = s->img_x * s->img_y;
+   stbi_uc *p = z->out;
+
+   if (s->img_out_n == 3) {  // convert bgr to rgb
+      for (i=0; i < pixel_count; ++i) {
+         stbi_uc t = p[0];
+         p[0] = p[2];
+         p[2] = t;
+         p += 3;
+      }
+   } else {
+      STBI_ASSERT(s->img_out_n == 4);
+      if (stbi__unpremultiply_on_load) {
+         // convert bgr to rgb and unpremultiply
+         for (i=0; i < pixel_count; ++i) {
+            stbi_uc a = p[3];
+            stbi_uc t = p[0];
+            if (a) {
+               stbi_uc half = a / 2;
+               p[0] = (p[2] * 255 + half) / a;
+               p[1] = (p[1] * 255 + half) / a;
+               p[2] = ( t   * 255 + half) / a;
+            } else {
+               p[0] = p[2];
+               p[2] = t;
+            }
+            p += 4;
+         }
+      } else {
+         // convert bgr to rgb
+         for (i=0; i < pixel_count; ++i) {
+            stbi_uc t = p[0];
+            p[0] = p[2];
+            p[2] = t;
+            p += 4;
+         }
+      }
+   }
+}
+
+#define STBI__PNG_TYPE(a,b,c,d)  (((unsigned) (a) << 24) + ((unsigned) (b) << 16) + ((unsigned) (c) << 8) + (unsigned) (d))
+
+static int stbi__parse_png_file(stbi__png *z, int scan, int req_comp)
+{
+   stbi_uc palette[1024], pal_img_n=0;
+   stbi_uc has_trans=0, tc[3]={0};
+   stbi__uint16 tc16[3];
+   stbi__uint32 ioff=0, idata_limit=0, i, pal_len=0;
+   int first=1,k,interlace=0, color=0, is_iphone=0;
+   stbi__context *s = z->s;
+
+   z->expanded = NULL;
+   z->idata = NULL;
+   z->out = NULL;
+
+   if (!stbi__check_png_header(s)) return 0;
+
+   if (scan == STBI__SCAN_type) return 1;
+
+   for (;;) {
+      stbi__pngchunk c = stbi__get_chunk_header(s);
+      switch (c.type) {
+         case STBI__PNG_TYPE('C','g','B','I'):
+            is_iphone = 1;
+            stbi__skip(s, c.length);
+            break;
+         case STBI__PNG_TYPE('I','H','D','R'): {
+            int comp,filter;
+            if (!first) return stbi__err("multiple IHDR","Corrupt PNG");
+            first = 0;
+            if (c.length != 13) return stbi__err("bad IHDR len","Corrupt PNG");
+            s->img_x = stbi__get32be(s);
+            s->img_y = stbi__get32be(s);
+            if (s->img_y > STBI_MAX_DIMENSIONS) return stbi__err("too large","Very large image (corrupt?)");
+            if (s->img_x > STBI_MAX_DIMENSIONS) return stbi__err("too large","Very large image (corrupt?)");
+            z->depth = stbi__get8(s);  if (z->depth != 1 && z->depth != 2 && z->depth != 4 && z->depth != 8 && z->depth != 16)  return stbi__err("1/2/4/8/16-bit only","PNG not supported: 1/2/4/8/16-bit only");
+            color = stbi__get8(s);  if (color > 6)         return stbi__err("bad ctype","Corrupt PNG");
+            if (color == 3 && z->depth == 16)                  return stbi__err("bad ctype","Corrupt PNG");
+            if (color == 3) pal_img_n = 3; else if (color & 1) return stbi__err("bad ctype","Corrupt PNG");
+            comp  = stbi__get8(s);  if (comp) return stbi__err("bad comp method","Corrupt PNG");
+            filter= stbi__get8(s);  if (filter) return stbi__err("bad filter method","Corrupt PNG");
+            interlace = stbi__get8(s); if (interlace>1) return stbi__err("bad interlace method","Corrupt PNG");
+            if (!s->img_x || !s->img_y) return stbi__err("0-pixel image","Corrupt PNG");
+            if (!pal_img_n) {
+               s->img_n = (color & 2 ? 3 : 1) + (color & 4 ? 1 : 0);
+               if ((1 << 30) / s->img_x / s->img_n < s->img_y) return stbi__err("too large", "Image too large to decode");
+            } else {
+               // if paletted, then pal_n is our final components, and
+               // img_n is # components to decompress/filter.
+               s->img_n = 1;
+               if ((1 << 30) / s->img_x / 4 < s->img_y) return stbi__err("too large","Corrupt PNG");
+            }
+            // even with SCAN_header, have to scan to see if we have a tRNS
+            break;
+         }
+
+         case STBI__PNG_TYPE('P','L','T','E'):  {
+            if (first) return stbi__err("first not IHDR", "Corrupt PNG");
+            if (c.length > 256*3) return stbi__err("invalid PLTE","Corrupt PNG");
+            pal_len = c.length / 3;
+            if (pal_len * 3 != c.length) return stbi__err("invalid PLTE","Corrupt PNG");
+            for (i=0; i < pal_len; ++i) {
+               palette[i*4+0] = stbi__get8(s);
+               palette[i*4+1] = stbi__get8(s);
+               palette[i*4+2] = stbi__get8(s);
+               palette[i*4+3] = 255;
+            }
+            break;
+         }
+
+         case STBI__PNG_TYPE('t','R','N','S'): {
+            if (first) return stbi__err("first not IHDR", "Corrupt PNG");
+            if (z->idata) return stbi__err("tRNS after IDAT","Corrupt PNG");
+            if (pal_img_n) {
+               if (scan == STBI__SCAN_header) { s->img_n = 4; return 1; }
+               if (pal_len == 0) return stbi__err("tRNS before PLTE","Corrupt PNG");
+               if (c.length > pal_len) return stbi__err("bad tRNS len","Corrupt PNG");
+               pal_img_n = 4;
+               for (i=0; i < c.length; ++i)
+                  palette[i*4+3] = stbi__get8(s);
+            } else {
+               if (!(s->img_n & 1)) return stbi__err("tRNS with alpha","Corrupt PNG");
+               if (c.length != (stbi__uint32) s->img_n*2) return stbi__err("bad tRNS len","Corrupt PNG");
+               has_trans = 1;
+               // non-paletted with tRNS = constant alpha. if header-scanning, we can stop now.
+               if (scan == STBI__SCAN_header) { ++s->img_n; return 1; }
+               if (z->depth == 16) {
+                  for (k = 0; k < s->img_n; ++k) tc16[k] = (stbi__uint16)stbi__get16be(s); // copy the values as-is
+               } else {
+                  for (k = 0; k < s->img_n; ++k) tc[k] = (stbi_uc)(stbi__get16be(s) & 255) * stbi__depth_scale_table[z->depth]; // non 8-bit images will be larger
+               }
+            }
+            break;
+         }
+
+         case STBI__PNG_TYPE('I','D','A','T'): {
+            if (first) return stbi__err("first not IHDR", "Corrupt PNG");
+            if (pal_img_n && !pal_len) return stbi__err("no PLTE","Corrupt PNG");
+            if (scan == STBI__SCAN_header) {
+               // header scan definitely stops at first IDAT
+               if (pal_img_n)
+                  s->img_n = pal_img_n;
+               return 1;
+            }
+            if (c.length > (1u << 30)) return stbi__err("IDAT size limit", "IDAT section larger than 2^30 bytes");
+            if ((int)(ioff + c.length) < (int)ioff) return 0;
+            if (ioff + c.length > idata_limit) {
+               stbi__uint32 idata_limit_old = idata_limit;
+               stbi_uc *p;
+               if (idata_limit == 0) idata_limit = c.length > 4096 ? c.length : 4096;
+               while (ioff + c.length > idata_limit)
+                  idata_limit *= 2;
+               STBI_NOTUSED(idata_limit_old);
+               p = (stbi_uc *) STBI_REALLOC_SIZED(z->idata, idata_limit_old, idata_limit); if (p == NULL) return stbi__err("outofmem", "Out of memory");
+               z->idata = p;
+            }
+            if (!stbi__getn(s, z->idata+ioff,c.length)) return stbi__err("outofdata","Corrupt PNG");
+            ioff += c.length;
+            break;
+         }
+
+         case STBI__PNG_TYPE('I','E','N','D'): {
+            stbi__uint32 raw_len, bpl;
+            if (first) return stbi__err("first not IHDR", "Corrupt PNG");
+            if (scan != STBI__SCAN_load) return 1;
+            if (z->idata == NULL) return stbi__err("no IDAT","Corrupt PNG");
+            // initial guess for decoded data size to avoid unnecessary reallocs
+            bpl = (s->img_x * z->depth + 7) / 8; // bytes per line, per component
+            raw_len = bpl * s->img_y * s->img_n /* pixels */ + s->img_y /* filter mode per row */;
+            z->expanded = (stbi_uc *) stbi_zlib_decode_malloc_guesssize_headerflag((char *) z->idata, ioff, raw_len, (int *) &raw_len, !is_iphone);
+            if (z->expanded == NULL) return 0; // zlib should set error
+            STBI_FREE(z->idata); z->idata = NULL;
+            if ((req_comp == s->img_n+1 && req_comp != 3 && !pal_img_n) || has_trans)
+               s->img_out_n = s->img_n+1;
+            else
+               s->img_out_n = s->img_n;
+            if (!stbi__create_png_image(z, z->expanded, raw_len, s->img_out_n, z->depth, color, interlace)) return 0;
+            if (has_trans) {
+               if (z->depth == 16) {
+                  if (!stbi__compute_transparency16(z, tc16, s->img_out_n)) return 0;
+               } else {
+                  if (!stbi__compute_transparency(z, tc, s->img_out_n)) return 0;
+               }
+            }
+            if (is_iphone && stbi__de_iphone_flag && s->img_out_n > 2)
+               stbi__de_iphone(z);
+            if (pal_img_n) {
+               // pal_img_n == 3 or 4
+               s->img_n = pal_img_n; // record the actual colors we had
+               s->img_out_n = pal_img_n;
+               if (req_comp >= 3) s->img_out_n = req_comp;
+               if (!stbi__expand_png_palette(z, palette, pal_len, s->img_out_n))
+                  return 0;
+            } else if (has_trans) {
+               // non-paletted image with tRNS -> source image has (constant) alpha
+               ++s->img_n;
+            }
+            STBI_FREE(z->expanded); z->expanded = NULL;
+            // end of PNG chunk, read and skip CRC
+            stbi__get32be(s);
+            return 1;
+         }
+
+         default:
+            // if critical, fail
+            if (first) return stbi__err("first not IHDR", "Corrupt PNG");
+            if ((c.type & (1 << 29)) == 0) {
+               #ifndef STBI_NO_FAILURE_STRINGS
+               // not threadsafe
+               static char invalid_chunk[] = "XXXX PNG chunk not known";
+               invalid_chunk[0] = STBI__BYTECAST(c.type >> 24);
+               invalid_chunk[1] = STBI__BYTECAST(c.type >> 16);
+               invalid_chunk[2] = STBI__BYTECAST(c.type >>  8);
+               invalid_chunk[3] = STBI__BYTECAST(c.type >>  0);
+               #endif
+               return stbi__err(invalid_chunk, "PNG not supported: unknown PNG chunk type");
+            }
+            stbi__skip(s, c.length);
+            break;
+      }
+      // end of PNG chunk, read and skip CRC
+      stbi__get32be(s);
+   }
+}
+
+static void *stbi__do_png(stbi__png *p, int *x, int *y, int *n, int req_comp, stbi__result_info *ri)
+{
+   void *result=NULL;
+   if (req_comp < 0 || req_comp > 4) return stbi__errpuc("bad req_comp", "Internal error");
+   if (stbi__parse_png_file(p, STBI__SCAN_load, req_comp)) {
+      if (p->depth <= 8)
+         ri->bits_per_channel = 8;
+      else if (p->depth == 16)
+         ri->bits_per_channel = 16;
+      else
+         return stbi__errpuc("bad bits_per_channel", "PNG not supported: unsupported color depth");
+      result = p->out;
+      p->out = NULL;
+      if (req_comp && req_comp != p->s->img_out_n) {
+         if (ri->bits_per_channel == 8)
+            result = stbi__convert_format((unsigned char *) result, p->s->img_out_n, req_comp, p->s->img_x, p->s->img_y);
+         else
+            result = stbi__convert_format16((stbi__uint16 *) result, p->s->img_out_n, req_comp, p->s->img_x, p->s->img_y);
+         p->s->img_out_n = req_comp;
+         if (result == NULL) return result;
+      }
+      *x = p->s->img_x;
+      *y = p->s->img_y;
+      if (n) *n = p->s->img_n;
+   }
+   STBI_FREE(p->out);      p->out      = NULL;
+   STBI_FREE(p->expanded); p->expanded = NULL;
+   STBI_FREE(p->idata);    p->idata    = NULL;
+
+   return result;
+}
+
+static void *stbi__png_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri)
+{
+   stbi__png p;
+   p.s = s;
+   return stbi__do_png(&p, x,y,comp,req_comp, ri);
+}
+
+static int stbi__png_test(stbi__context *s)
+{
+   int r;
+   r = stbi__check_png_header(s);
+   stbi__rewind(s);
+   return r;
+}
+
+static int stbi__png_info_raw(stbi__png *p, int *x, int *y, int *comp)
+{
+   if (!stbi__parse_png_file(p, STBI__SCAN_header, 0)) {
+      stbi__rewind( p->s );
+      return 0;
+   }
+   if (x) *x = p->s->img_x;
+   if (y) *y = p->s->img_y;
+   if (comp) *comp = p->s->img_n;
+   return 1;
+}
+
+static int stbi__png_info(stbi__context *s, int *x, int *y, int *comp)
+{
+   stbi__png p;
+   p.s = s;
+   return stbi__png_info_raw(&p, x, y, comp);
+}
+
+static int stbi__png_is16(stbi__context *s)
+{
+   stbi__png p;
+   p.s = s;
+   if (!stbi__png_info_raw(&p, NULL, NULL, NULL))
+	   return 0;
+   if (p.depth != 16) {
+      stbi__rewind(p.s);
+      return 0;
+   }
+   return 1;
+}
+#endif
+
+// Microsoft/Windows BMP image
+
+#ifndef STBI_NO_BMP
+static int stbi__bmp_test_raw(stbi__context *s)
+{
+   int r;
+   int sz;
+   if (stbi__get8(s) != 'B') return 0;
+   if (stbi__get8(s) != 'M') return 0;
+   stbi__get32le(s); // discard filesize
+   stbi__get16le(s); // discard reserved
+   stbi__get16le(s); // discard reserved
+   stbi__get32le(s); // discard data offset
+   sz = stbi__get32le(s);
+   r = (sz == 12 || sz == 40 || sz == 56 || sz == 108 || sz == 124);
+   return r;
+}
+
+static int stbi__bmp_test(stbi__context *s)
+{
+   int r = stbi__bmp_test_raw(s);
+   stbi__rewind(s);
+   return r;
+}
+
+
+// returns 0..31 for the highest set bit
+static int stbi__high_bit(unsigned int z)
+{
+   int n=0;
+   if (z == 0) return -1;
+   if (z >= 0x10000) { n += 16; z >>= 16; }
+   if (z >= 0x00100) { n +=  8; z >>=  8; }
+   if (z >= 0x00010) { n +=  4; z >>=  4; }
+   if (z >= 0x00004) { n +=  2; z >>=  2; }
+   if (z >= 0x00002) { n +=  1;/* >>=  1;*/ }
+   return n;
+}
+
+static int stbi__bitcount(unsigned int a)
+{
+   a = (a & 0x55555555) + ((a >>  1) & 0x55555555); // max 2
+   a = (a & 0x33333333) + ((a >>  2) & 0x33333333); // max 4
+   a = (a + (a >> 4)) & 0x0f0f0f0f; // max 8 per 4, now 8 bits
+   a = (a + (a >> 8)); // max 16 per 8 bits
+   a = (a + (a >> 16)); // max 32 per 8 bits
+   return a & 0xff;
+}
+
+// extract an arbitrarily-aligned N-bit value (N=bits)
+// from v, and then make it 8-bits long and fractionally
+// extend it to full full range.
+static int stbi__shiftsigned(unsigned int v, int shift, int bits)
+{
+   static unsigned int mul_table[9] = {
+      0,
+      0xff/*0b11111111*/, 0x55/*0b01010101*/, 0x49/*0b01001001*/, 0x11/*0b00010001*/,
+      0x21/*0b00100001*/, 0x41/*0b01000001*/, 0x81/*0b10000001*/, 0x01/*0b00000001*/,
+   };
+   static unsigned int shift_table[9] = {
+      0, 0,0,1,0,2,4,6,0,
+   };
+   if (shift < 0)
+      v <<= -shift;
+   else
+      v >>= shift;
+   STBI_ASSERT(v < 256);
+   v >>= (8-bits);
+   STBI_ASSERT(bits >= 0 && bits <= 8);
+   return (int) ((unsigned) v * mul_table[bits]) >> shift_table[bits];
+}
+
+typedef struct
+{
+   int bpp, offset, hsz;
+   unsigned int mr,mg,mb,ma, all_a;
+   int extra_read;
+} stbi__bmp_data;
+
+static int stbi__bmp_set_mask_defaults(stbi__bmp_data *info, int compress)
+{
+   // BI_BITFIELDS specifies masks explicitly, don't override
+   if (compress == 3)
+      return 1;
+
+   if (compress == 0) {
+      if (info->bpp == 16) {
+         info->mr = 31u << 10;
+         info->mg = 31u <<  5;
+         info->mb = 31u <<  0;
+      } else if (info->bpp == 32) {
+         info->mr = 0xffu << 16;
+         info->mg = 0xffu <<  8;
+         info->mb = 0xffu <<  0;
+         info->ma = 0xffu << 24;
+         info->all_a = 0; // if all_a is 0 at end, then we loaded alpha channel but it was all 0
+      } else {
+         // otherwise, use defaults, which is all-0
+         info->mr = info->mg = info->mb = info->ma = 0;
+      }
+      return 1;
+   }
+   return 0; // error
+}
+
+static void *stbi__bmp_parse_header(stbi__context *s, stbi__bmp_data *info)
+{
+   int hsz;
+   if (stbi__get8(s) != 'B' || stbi__get8(s) != 'M') return stbi__errpuc("not BMP", "Corrupt BMP");
+   stbi__get32le(s); // discard filesize
+   stbi__get16le(s); // discard reserved
+   stbi__get16le(s); // discard reserved
+   info->offset = stbi__get32le(s);
+   info->hsz = hsz = stbi__get32le(s);
+   info->mr = info->mg = info->mb = info->ma = 0;
+   info->extra_read = 14;
+
+   if (info->offset < 0) return stbi__errpuc("bad BMP", "bad BMP");
+
+   if (hsz != 12 && hsz != 40 && hsz != 56 && hsz != 108 && hsz != 124) return stbi__errpuc("unknown BMP", "BMP type not supported: unknown");
+   if (hsz == 12) {
+      s->img_x = stbi__get16le(s);
+      s->img_y = stbi__get16le(s);
+   } else {
+      s->img_x = stbi__get32le(s);
+      s->img_y = stbi__get32le(s);
+   }
+   if (stbi__get16le(s) != 1) return stbi__errpuc("bad BMP", "bad BMP");
+   info->bpp = stbi__get16le(s);
+   if (hsz != 12) {
+      int compress = stbi__get32le(s);
+      if (compress == 1 || compress == 2) return stbi__errpuc("BMP RLE", "BMP type not supported: RLE");
+      if (compress >= 4) return stbi__errpuc("BMP JPEG/PNG", "BMP type not supported: unsupported compression"); // this includes PNG/JPEG modes
+      if (compress == 3 && info->bpp != 16 && info->bpp != 32) return stbi__errpuc("bad BMP", "bad BMP"); // bitfields requires 16 or 32 bits/pixel
+      stbi__get32le(s); // discard sizeof
+      stbi__get32le(s); // discard hres
+      stbi__get32le(s); // discard vres
+      stbi__get32le(s); // discard colorsused
+      stbi__get32le(s); // discard max important
+      if (hsz == 40 || hsz == 56) {
+         if (hsz == 56) {
+            stbi__get32le(s);
+            stbi__get32le(s);
+            stbi__get32le(s);
+            stbi__get32le(s);
+         }
+         if (info->bpp == 16 || info->bpp == 32) {
+            if (compress == 0) {
+               stbi__bmp_set_mask_defaults(info, compress);
+            } else if (compress == 3) {
+               info->mr = stbi__get32le(s);
+               info->mg = stbi__get32le(s);
+               info->mb = stbi__get32le(s);
+               info->extra_read += 12;
+               // not documented, but generated by photoshop and handled by mspaint
+               if (info->mr == info->mg && info->mg == info->mb) {
+                  // ?!?!?
+                  return stbi__errpuc("bad BMP", "bad BMP");
+               }
+            } else
+               return stbi__errpuc("bad BMP", "bad BMP");
+         }
+      } else {
+         // V4/V5 header
+         int i;
+         if (hsz != 108 && hsz != 124)
+            return stbi__errpuc("bad BMP", "bad BMP");
+         info->mr = stbi__get32le(s);
+         info->mg = stbi__get32le(s);
+         info->mb = stbi__get32le(s);
+         info->ma = stbi__get32le(s);
+         if (compress != 3) // override mr/mg/mb unless in BI_BITFIELDS mode, as per docs
+            stbi__bmp_set_mask_defaults(info, compress);
+         stbi__get32le(s); // discard color space
+         for (i=0; i < 12; ++i)
+            stbi__get32le(s); // discard color space parameters
+         if (hsz == 124) {
+            stbi__get32le(s); // discard rendering intent
+            stbi__get32le(s); // discard offset of profile data
+            stbi__get32le(s); // discard size of profile data
+            stbi__get32le(s); // discard reserved
+         }
+      }
+   }
+   return (void *) 1;
+}
+
+
+static void *stbi__bmp_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri)
+{
+   stbi_uc *out;
+   unsigned int mr=0,mg=0,mb=0,ma=0, all_a;
+   stbi_uc pal[256][4];
+   int psize=0,i,j,width;
+   int flip_vertically, pad, target;
+   stbi__bmp_data info;
+   STBI_NOTUSED(ri);
+
+   info.all_a = 255;
+   if (stbi__bmp_parse_header(s, &info) == NULL)
+      return NULL; // error code already set
+
+   flip_vertically = ((int) s->img_y) > 0;
+   s->img_y = abs((int) s->img_y);
+
+   if (s->img_y > STBI_MAX_DIMENSIONS) return stbi__errpuc("too large","Very large image (corrupt?)");
+   if (s->img_x > STBI_MAX_DIMENSIONS) return stbi__errpuc("too large","Very large image (corrupt?)");
+
+   mr = info.mr;
+   mg = info.mg;
+   mb = info.mb;
+   ma = info.ma;
+   all_a = info.all_a;
+
+   if (info.hsz == 12) {
+      if (info.bpp < 24)
+         psize = (info.offset - info.extra_read - 24) / 3;
+   } else {
+      if (info.bpp < 16)
+         psize = (info.offset - info.extra_read - info.hsz) >> 2;
+   }
+   if (psize == 0) {
+      // accept some number of extra bytes after the header, but if the offset points either to before
+      // the header ends or implies a large amount of extra data, reject the file as malformed
+      int bytes_read_so_far = s->callback_already_read + (int)(s->img_buffer - s->img_buffer_original);
+      int header_limit = 1024; // max we actually read is below 256 bytes currently.
+      int extra_data_limit = 256*4; // what ordinarily goes here is a palette; 256 entries*4 bytes is its max size.
+      if (bytes_read_so_far <= 0 || bytes_read_so_far > header_limit) {
+         return stbi__errpuc("bad header", "Corrupt BMP");
+      }
+      // we established that bytes_read_so_far is positive and sensible.
+      // the first half of this test rejects offsets that are either too small positives, or
+      // negative, and guarantees that info.offset >= bytes_read_so_far > 0. this in turn
+      // ensures the number computed in the second half of the test can't overflow.
+      if (info.offset < bytes_read_so_far || info.offset - bytes_read_so_far > extra_data_limit) {
+         return stbi__errpuc("bad offset", "Corrupt BMP");
+      } else {
+         stbi__skip(s, info.offset - bytes_read_so_far);
+      }
+   }
+
+   if (info.bpp == 24 && ma == 0xff000000)
+      s->img_n = 3;
+   else
+      s->img_n = ma ? 4 : 3;
+   if (req_comp && req_comp >= 3) // we can directly decode 3 or 4
+      target = req_comp;
+   else
+      target = s->img_n; // if they want monochrome, we'll post-convert
+
+   // sanity-check size
+   if (!stbi__mad3sizes_valid(target, s->img_x, s->img_y, 0))
+      return stbi__errpuc("too large", "Corrupt BMP");
+
+   out = (stbi_uc *) stbi__malloc_mad3(target, s->img_x, s->img_y, 0);
+   if (!out) return stbi__errpuc("outofmem", "Out of memory");
+   if (info.bpp < 16) {
+      int z=0;
+      if (psize == 0 || psize > 256) { STBI_FREE(out); return stbi__errpuc("invalid", "Corrupt BMP"); }
+      for (i=0; i < psize; ++i) {
+         pal[i][2] = stbi__get8(s);
+         pal[i][1] = stbi__get8(s);
+         pal[i][0] = stbi__get8(s);
+         if (info.hsz != 12) stbi__get8(s);
+         pal[i][3] = 255;
+      }
+      stbi__skip(s, info.offset - info.extra_read - info.hsz - psize * (info.hsz == 12 ? 3 : 4));
+      if (info.bpp == 1) width = (s->img_x + 7) >> 3;
+      else if (info.bpp == 4) width = (s->img_x + 1) >> 1;
+      else if (info.bpp == 8) width = s->img_x;
+      else { STBI_FREE(out); return stbi__errpuc("bad bpp", "Corrupt BMP"); }
+      pad = (-width)&3;
+      if (info.bpp == 1) {
+         for (j=0; j < (int) s->img_y; ++j) {
+            int bit_offset = 7, v = stbi__get8(s);
+            for (i=0; i < (int) s->img_x; ++i) {
+               int color = (v>>bit_offset)&0x1;
+               out[z++] = pal[color][0];
+               out[z++] = pal[color][1];
+               out[z++] = pal[color][2];
+               if (target == 4) out[z++] = 255;
+               if (i+1 == (int) s->img_x) break;
+               if((--bit_offset) < 0) {
+                  bit_offset = 7;
+                  v = stbi__get8(s);
+               }
+            }
+            stbi__skip(s, pad);
+         }
+      } else {
+         for (j=0; j < (int) s->img_y; ++j) {
+            for (i=0; i < (int) s->img_x; i += 2) {
+               int v=stbi__get8(s),v2=0;
+               if (info.bpp == 4) {
+                  v2 = v & 15;
+                  v >>= 4;
+               }
+               out[z++] = pal[v][0];
+               out[z++] = pal[v][1];
+               out[z++] = pal[v][2];
+               if (target == 4) out[z++] = 255;
+               if (i+1 == (int) s->img_x) break;
+               v = (info.bpp == 8) ? stbi__get8(s) : v2;
+               out[z++] = pal[v][0];
+               out[z++] = pal[v][1];
+               out[z++] = pal[v][2];
+               if (target == 4) out[z++] = 255;
+            }
+            stbi__skip(s, pad);
+         }
+      }
+   } else {
+      int rshift=0,gshift=0,bshift=0,ashift=0,rcount=0,gcount=0,bcount=0,acount=0;
+      int z = 0;
+      int easy=0;
+      stbi__skip(s, info.offset - info.extra_read - info.hsz);
+      if (info.bpp == 24) width = 3 * s->img_x;
+      else if (info.bpp == 16) width = 2*s->img_x;
+      else /* bpp = 32 and pad = 0 */ width=0;
+      pad = (-width) & 3;
+      if (info.bpp == 24) {
+         easy = 1;
+      } else if (info.bpp == 32) {
+         if (mb == 0xff && mg == 0xff00 && mr == 0x00ff0000 && ma == 0xff000000)
+            easy = 2;
+      }
+      if (!easy) {
+         if (!mr || !mg || !mb) { STBI_FREE(out); return stbi__errpuc("bad masks", "Corrupt BMP"); }
+         // right shift amt to put high bit in position #7
+         rshift = stbi__high_bit(mr)-7; rcount = stbi__bitcount(mr);
+         gshift = stbi__high_bit(mg)-7; gcount = stbi__bitcount(mg);
+         bshift = stbi__high_bit(mb)-7; bcount = stbi__bitcount(mb);
+         ashift = stbi__high_bit(ma)-7; acount = stbi__bitcount(ma);
+         if (rcount > 8 || gcount > 8 || bcount > 8 || acount > 8) { STBI_FREE(out); return stbi__errpuc("bad masks", "Corrupt BMP"); }
+      }
+      for (j=0; j < (int) s->img_y; ++j) {
+         if (easy) {
+            for (i=0; i < (int) s->img_x; ++i) {
+               unsigned char a;
+               out[z+2] = stbi__get8(s);
+               out[z+1] = stbi__get8(s);
+               out[z+0] = stbi__get8(s);
+               z += 3;
+               a = (easy == 2 ? stbi__get8(s) : 255);
+               all_a |= a;
+               if (target == 4) out[z++] = a;
+            }
+         } else {
+            int bpp = info.bpp;
+            for (i=0; i < (int) s->img_x; ++i) {
+               stbi__uint32 v = (bpp == 16 ? (stbi__uint32) stbi__get16le(s) : stbi__get32le(s));
+               unsigned int a;
+               out[z++] = STBI__BYTECAST(stbi__shiftsigned(v & mr, rshift, rcount));
+               out[z++] = STBI__BYTECAST(stbi__shiftsigned(v & mg, gshift, gcount));
+               out[z++] = STBI__BYTECAST(stbi__shiftsigned(v & mb, bshift, bcount));
+               a = (ma ? stbi__shiftsigned(v & ma, ashift, acount) : 255);
+               all_a |= a;
+               if (target == 4) out[z++] = STBI__BYTECAST(a);
+            }
+         }
+         stbi__skip(s, pad);
+      }
+   }
+
+   // if alpha channel is all 0s, replace with all 255s
+   if (target == 4 && all_a == 0)
+      for (i=4*s->img_x*s->img_y-1; i >= 0; i -= 4)
+         out[i] = 255;
+
+   if (flip_vertically) {
+      stbi_uc t;
+      for (j=0; j < (int) s->img_y>>1; ++j) {
+         stbi_uc *p1 = out +      j     *s->img_x*target;
+         stbi_uc *p2 = out + (s->img_y-1-j)*s->img_x*target;
+         for (i=0; i < (int) s->img_x*target; ++i) {
+            t = p1[i]; p1[i] = p2[i]; p2[i] = t;
+         }
+      }
+   }
+
+   if (req_comp && req_comp != target) {
+      out = stbi__convert_format(out, target, req_comp, s->img_x, s->img_y);
+      if (out == NULL) return out; // stbi__convert_format frees input on failure
+   }
+
+   *x = s->img_x;
+   *y = s->img_y;
+   if (comp) *comp = s->img_n;
+   return out;
+}
+#endif
+
+// Targa Truevision - TGA
+// by Jonathan Dummer
+#ifndef STBI_NO_TGA
+// returns STBI_rgb or whatever, 0 on error
+static int stbi__tga_get_comp(int bits_per_pixel, int is_grey, int* is_rgb16)
+{
+   // only RGB or RGBA (incl. 16bit) or grey allowed
+   if (is_rgb16) *is_rgb16 = 0;
+   switch(bits_per_pixel) {
+      case 8:  return STBI_grey;
+      case 16: if(is_grey) return STBI_grey_alpha;
+               // fallthrough
+      case 15: if(is_rgb16) *is_rgb16 = 1;
+               return STBI_rgb;
+      case 24: // fallthrough
+      case 32: return bits_per_pixel/8;
+      default: return 0;
+   }
+}
+
+static int stbi__tga_info(stbi__context *s, int *x, int *y, int *comp)
+{
+    int tga_w, tga_h, tga_comp, tga_image_type, tga_bits_per_pixel, tga_colormap_bpp;
+    int sz, tga_colormap_type;
+    stbi__get8(s);                   // discard Offset
+    tga_colormap_type = stbi__get8(s); // colormap type
+    if( tga_colormap_type > 1 ) {
+        stbi__rewind(s);
+        return 0;      // only RGB or indexed allowed
+    }
+    tga_image_type = stbi__get8(s); // image type
+    if ( tga_colormap_type == 1 ) { // colormapped (paletted) image
+        if (tga_image_type != 1 && tga_image_type != 9) {
+            stbi__rewind(s);
+            return 0;
+        }
+        stbi__skip(s,4);       // skip index of first colormap entry and number of entries
+        sz = stbi__get8(s);    //   check bits per palette color entry
+        if ( (sz != 8) && (sz != 15) && (sz != 16) && (sz != 24) && (sz != 32) ) {
+            stbi__rewind(s);
+            return 0;
+        }
+        stbi__skip(s,4);       // skip image x and y origin
+        tga_colormap_bpp = sz;
+    } else { // "normal" image w/o colormap - only RGB or grey allowed, +/- RLE
+        if ( (tga_image_type != 2) && (tga_image_type != 3) && (tga_image_type != 10) && (tga_image_type != 11) ) {
+            stbi__rewind(s);
+            return 0; // only RGB or grey allowed, +/- RLE
+        }
+        stbi__skip(s,9); // skip colormap specification and image x/y origin
+        tga_colormap_bpp = 0;
+    }
+    tga_w = stbi__get16le(s);
+    if( tga_w < 1 ) {
+        stbi__rewind(s);
+        return 0;   // test width
+    }
+    tga_h = stbi__get16le(s);
+    if( tga_h < 1 ) {
+        stbi__rewind(s);
+        return 0;   // test height
+    }
+    tga_bits_per_pixel = stbi__get8(s); // bits per pixel
+    stbi__get8(s); // ignore alpha bits
+    if (tga_colormap_bpp != 0) {
+        if((tga_bits_per_pixel != 8) && (tga_bits_per_pixel != 16)) {
+            // when using a colormap, tga_bits_per_pixel is the size of the indexes
+            // I don't think anything but 8 or 16bit indexes makes sense
+            stbi__rewind(s);
+            return 0;
+        }
+        tga_comp = stbi__tga_get_comp(tga_colormap_bpp, 0, NULL);
+    } else {
+        tga_comp = stbi__tga_get_comp(tga_bits_per_pixel, (tga_image_type == 3) || (tga_image_type == 11), NULL);
+    }
+    if(!tga_comp) {
+      stbi__rewind(s);
+      return 0;
+    }
+    if (x) *x = tga_w;
+    if (y) *y = tga_h;
+    if (comp) *comp = tga_comp;
+    return 1;                   // seems to have passed everything
+}
+
+static int stbi__tga_test(stbi__context *s)
+{
+   int res = 0;
+   int sz, tga_color_type;
+   stbi__get8(s);      //   discard Offset
+   tga_color_type = stbi__get8(s);   //   color type
+   if ( tga_color_type > 1 ) goto errorEnd;   //   only RGB or indexed allowed
+   sz = stbi__get8(s);   //   image type
+   if ( tga_color_type == 1 ) { // colormapped (paletted) image
+      if (sz != 1 && sz != 9) goto errorEnd; // colortype 1 demands image type 1 or 9
+      stbi__skip(s,4);       // skip index of first colormap entry and number of entries
+      sz = stbi__get8(s);    //   check bits per palette color entry
+      if ( (sz != 8) && (sz != 15) && (sz != 16) && (sz != 24) && (sz != 32) ) goto errorEnd;
+      stbi__skip(s,4);       // skip image x and y origin
+   } else { // "normal" image w/o colormap
+      if ( (sz != 2) && (sz != 3) && (sz != 10) && (sz != 11) ) goto errorEnd; // only RGB or grey allowed, +/- RLE
+      stbi__skip(s,9); // skip colormap specification and image x/y origin
+   }
+   if ( stbi__get16le(s) < 1 ) goto errorEnd;      //   test width
+   if ( stbi__get16le(s) < 1 ) goto errorEnd;      //   test height
+   sz = stbi__get8(s);   //   bits per pixel
+   if ( (tga_color_type == 1) && (sz != 8) && (sz != 16) ) goto errorEnd; // for colormapped images, bpp is size of an index
+   if ( (sz != 8) && (sz != 15) && (sz != 16) && (sz != 24) && (sz != 32) ) goto errorEnd;
+
+   res = 1; // if we got this far, everything's good and we can return 1 instead of 0
+
+errorEnd:
+   stbi__rewind(s);
+   return res;
+}
+
+// read 16bit value and convert to 24bit RGB
+static void stbi__tga_read_rgb16(stbi__context *s, stbi_uc* out)
+{
+   stbi__uint16 px = (stbi__uint16)stbi__get16le(s);
+   stbi__uint16 fiveBitMask = 31;
+   // we have 3 channels with 5bits each
+   int r = (px >> 10) & fiveBitMask;
+   int g = (px >> 5) & fiveBitMask;
+   int b = px & fiveBitMask;
+   // Note that this saves the data in RGB(A) order, so it doesn't need to be swapped later
+   out[0] = (stbi_uc)((r * 255)/31);
+   out[1] = (stbi_uc)((g * 255)/31);
+   out[2] = (stbi_uc)((b * 255)/31);
+
+   // some people claim that the most significant bit might be used for alpha
+   // (possibly if an alpha-bit is set in the "image descriptor byte")
+   // but that only made 16bit test images completely translucent..
+   // so let's treat all 15 and 16bit TGAs as RGB with no alpha.
+}
+
+static void *stbi__tga_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri)
+{
+   //   read in the TGA header stuff
+   int tga_offset = stbi__get8(s);
+   int tga_indexed = stbi__get8(s);
+   int tga_image_type = stbi__get8(s);
+   int tga_is_RLE = 0;
+   int tga_palette_start = stbi__get16le(s);
+   int tga_palette_len = stbi__get16le(s);
+   int tga_palette_bits = stbi__get8(s);
+   int tga_x_origin = stbi__get16le(s);
+   int tga_y_origin = stbi__get16le(s);
+   int tga_width = stbi__get16le(s);
+   int tga_height = stbi__get16le(s);
+   int tga_bits_per_pixel = stbi__get8(s);
+   int tga_comp, tga_rgb16=0;
+   int tga_inverted = stbi__get8(s);
+   // int tga_alpha_bits = tga_inverted & 15; // the 4 lowest bits - unused (useless?)
+   //   image data
+   unsigned char *tga_data;
+   unsigned char *tga_palette = NULL;
+   int i, j;
+   unsigned char raw_data[4] = {0};
+   int RLE_count = 0;
+   int RLE_repeating = 0;
+   int read_next_pixel = 1;
+   STBI_NOTUSED(ri);
+   STBI_NOTUSED(tga_x_origin); // @TODO
+   STBI_NOTUSED(tga_y_origin); // @TODO
+
+   if (tga_height > STBI_MAX_DIMENSIONS) return stbi__errpuc("too large","Very large image (corrupt?)");
+   if (tga_width > STBI_MAX_DIMENSIONS) return stbi__errpuc("too large","Very large image (corrupt?)");
+
+   //   do a tiny bit of precessing
+   if ( tga_image_type >= 8 )
+   {
+      tga_image_type -= 8;
+      tga_is_RLE = 1;
+   }
+   tga_inverted = 1 - ((tga_inverted >> 5) & 1);
+
+   //   If I'm paletted, then I'll use the number of bits from the palette
+   if ( tga_indexed ) tga_comp = stbi__tga_get_comp(tga_palette_bits, 0, &tga_rgb16);
+   else tga_comp = stbi__tga_get_comp(tga_bits_per_pixel, (tga_image_type == 3), &tga_rgb16);
+
+   if(!tga_comp) // shouldn't really happen, stbi__tga_test() should have ensured basic consistency
+      return stbi__errpuc("bad format", "Can't find out TGA pixelformat");
+
+   //   tga info
+   *x = tga_width;
+   *y = tga_height;
+   if (comp) *comp = tga_comp;
+
+   if (!stbi__mad3sizes_valid(tga_width, tga_height, tga_comp, 0))
+      return stbi__errpuc("too large", "Corrupt TGA");
+
+   tga_data = (unsigned char*)stbi__malloc_mad3(tga_width, tga_height, tga_comp, 0);
+   if (!tga_data) return stbi__errpuc("outofmem", "Out of memory");
+
+   // skip to the data's starting position (offset usually = 0)
+   stbi__skip(s, tga_offset );
+
+   if ( !tga_indexed && !tga_is_RLE && !tga_rgb16 ) {
+      for (i=0; i < tga_height; ++i) {
+         int row = tga_inverted ? tga_height -i - 1 : i;
+         stbi_uc *tga_row = tga_data + row*tga_width*tga_comp;
+         stbi__getn(s, tga_row, tga_width * tga_comp);
+      }
+   } else  {
+      //   do I need to load a palette?
+      if ( tga_indexed)
+      {
+         if (tga_palette_len == 0) {  /* you have to have at least one entry! */
+            STBI_FREE(tga_data);
+            return stbi__errpuc("bad palette", "Corrupt TGA");
+         }
+
+         //   any data to skip? (offset usually = 0)
+         stbi__skip(s, tga_palette_start );
+         //   load the palette
+         tga_palette = (unsigned char*)stbi__malloc_mad2(tga_palette_len, tga_comp, 0);
+         if (!tga_palette) {
+            STBI_FREE(tga_data);
+            return stbi__errpuc("outofmem", "Out of memory");
+         }
+         if (tga_rgb16) {
+            stbi_uc *pal_entry = tga_palette;
+            STBI_ASSERT(tga_comp == STBI_rgb);
+            for (i=0; i < tga_palette_len; ++i) {
+               stbi__tga_read_rgb16(s, pal_entry);
+               pal_entry += tga_comp;
+            }
+         } else if (!stbi__getn(s, tga_palette, tga_palette_len * tga_comp)) {
+               STBI_FREE(tga_data);
+               STBI_FREE(tga_palette);
+               return stbi__errpuc("bad palette", "Corrupt TGA");
+         }
+      }
+      //   load the data
+      for (i=0; i < tga_width * tga_height; ++i)
+      {
+         //   if I'm in RLE mode, do I need to get a RLE stbi__pngchunk?
+         if ( tga_is_RLE )
+         {
+            if ( RLE_count == 0 )
+            {
+               //   yep, get the next byte as a RLE command
+               int RLE_cmd = stbi__get8(s);
+               RLE_count = 1 + (RLE_cmd & 127);
+               RLE_repeating = RLE_cmd >> 7;
+               read_next_pixel = 1;
+            } else if ( !RLE_repeating )
+            {
+               read_next_pixel = 1;
+            }
+         } else
+         {
+            read_next_pixel = 1;
+         }
+         //   OK, if I need to read a pixel, do it now
+         if ( read_next_pixel )
+         {
+            //   load however much data we did have
+            if ( tga_indexed )
+            {
+               // read in index, then perform the lookup
+               int pal_idx = (tga_bits_per_pixel == 8) ? stbi__get8(s) : stbi__get16le(s);
+               if ( pal_idx >= tga_palette_len ) {
+                  // invalid index
+                  pal_idx = 0;
+               }
+               pal_idx *= tga_comp;
+               for (j = 0; j < tga_comp; ++j) {
+                  raw_data[j] = tga_palette[pal_idx+j];
+               }
+            } else if(tga_rgb16) {
+               STBI_ASSERT(tga_comp == STBI_rgb);
+               stbi__tga_read_rgb16(s, raw_data);
+            } else {
+               //   read in the data raw
+               for (j = 0; j < tga_comp; ++j) {
+                  raw_data[j] = stbi__get8(s);
+               }
+            }
+            //   clear the reading flag for the next pixel
+            read_next_pixel = 0;
+         } // end of reading a pixel
+
+         // copy data
+         for (j = 0; j < tga_comp; ++j)
+           tga_data[i*tga_comp+j] = raw_data[j];
+
+         //   in case we're in RLE mode, keep counting down
+         --RLE_count;
+      }
+      //   do I need to invert the image?
+      if ( tga_inverted )
+      {
+         for (j = 0; j*2 < tga_height; ++j)
+         {
+            int index1 = j * tga_width * tga_comp;
+            int index2 = (tga_height - 1 - j) * tga_width * tga_comp;
+            for (i = tga_width * tga_comp; i > 0; --i)
+            {
+               unsigned char temp = tga_data[index1];
+               tga_data[index1] = tga_data[index2];
+               tga_data[index2] = temp;
+               ++index1;
+               ++index2;
+            }
+         }
+      }
+      //   clear my palette, if I had one
+      if ( tga_palette != NULL )
+      {
+         STBI_FREE( tga_palette );
+      }
+   }
+
+   // swap RGB - if the source data was RGB16, it already is in the right order
+   if (tga_comp >= 3 && !tga_rgb16)
+   {
+      unsigned char* tga_pixel = tga_data;
+      for (i=0; i < tga_width * tga_height; ++i)
+      {
+         unsigned char temp = tga_pixel[0];
+         tga_pixel[0] = tga_pixel[2];
+         tga_pixel[2] = temp;
+         tga_pixel += tga_comp;
+      }
+   }
+
+   // convert to target component count
+   if (req_comp && req_comp != tga_comp)
+      tga_data = stbi__convert_format(tga_data, tga_comp, req_comp, tga_width, tga_height);
+
+   //   the things I do to get rid of an error message, and yet keep
+   //   Microsoft's C compilers happy... [8^(
+   tga_palette_start = tga_palette_len = tga_palette_bits =
+         tga_x_origin = tga_y_origin = 0;
+   STBI_NOTUSED(tga_palette_start);
+   //   OK, done
+   return tga_data;
+}
+#endif
+
+// *************************************************************************************************
+// Photoshop PSD loader -- PD by Thatcher Ulrich, integration by Nicolas Schulz, tweaked by STB
+
+#ifndef STBI_NO_PSD
+static int stbi__psd_test(stbi__context *s)
+{
+   int r = (stbi__get32be(s) == 0x38425053);
+   stbi__rewind(s);
+   return r;
+}
+
+static int stbi__psd_decode_rle(stbi__context *s, stbi_uc *p, int pixelCount)
+{
+   int count, nleft, len;
+
+   count = 0;
+   while ((nleft = pixelCount - count) > 0) {
+      len = stbi__get8(s);
+      if (len == 128) {
+         // No-op.
+      } else if (len < 128) {
+         // Copy next len+1 bytes literally.
+         len++;
+         if (len > nleft) return 0; // corrupt data
+         count += len;
+         while (len) {
+            *p = stbi__get8(s);
+            p += 4;
+            len--;
+         }
+      } else if (len > 128) {
+         stbi_uc   val;
+         // Next -len+1 bytes in the dest are replicated from next source byte.
+         // (Interpret len as a negative 8-bit int.)
+         len = 257 - len;
+         if (len > nleft) return 0; // corrupt data
+         val = stbi__get8(s);
+         count += len;
+         while (len) {
+            *p = val;
+            p += 4;
+            len--;
+         }
+      }
+   }
+
+   return 1;
+}
+
+static void *stbi__psd_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri, int bpc)
+{
+   int pixelCount;
+   int channelCount, compression;
+   int channel, i;
+   int bitdepth;
+   int w,h;
+   stbi_uc *out;
+   STBI_NOTUSED(ri);
+
+   // Check identifier
+   if (stbi__get32be(s) != 0x38425053)   // "8BPS"
+      return stbi__errpuc("not PSD", "Corrupt PSD image");
+
+   // Check file type version.
+   if (stbi__get16be(s) != 1)
+      return stbi__errpuc("wrong version", "Unsupported version of PSD image");
+
+   // Skip 6 reserved bytes.
+   stbi__skip(s, 6 );
+
+   // Read the number of channels (R, G, B, A, etc).
+   channelCount = stbi__get16be(s);
+   if (channelCount < 0 || channelCount > 16)
+      return stbi__errpuc("wrong channel count", "Unsupported number of channels in PSD image");
+
+   // Read the rows and columns of the image.
+   h = stbi__get32be(s);
+   w = stbi__get32be(s);
+
+   if (h > STBI_MAX_DIMENSIONS) return stbi__errpuc("too large","Very large image (corrupt?)");
+   if (w > STBI_MAX_DIMENSIONS) return stbi__errpuc("too large","Very large image (corrupt?)");
+
+   // Make sure the depth is 8 bits.
+   bitdepth = stbi__get16be(s);
+   if (bitdepth != 8 && bitdepth != 16)
+      return stbi__errpuc("unsupported bit depth", "PSD bit depth is not 8 or 16 bit");
+
+   // Make sure the color mode is RGB.
+   // Valid options are:
+   //   0: Bitmap
+   //   1: Grayscale
+   //   2: Indexed color
+   //   3: RGB color
+   //   4: CMYK color
+   //   7: Multichannel
+   //   8: Duotone
+   //   9: Lab color
+   if (stbi__get16be(s) != 3)
+      return stbi__errpuc("wrong color format", "PSD is not in RGB color format");
+
+   // Skip the Mode Data.  (It's the palette for indexed color; other info for other modes.)
+   stbi__skip(s,stbi__get32be(s) );
+
+   // Skip the image resources.  (resolution, pen tool paths, etc)
+   stbi__skip(s, stbi__get32be(s) );
+
+   // Skip the reserved data.
+   stbi__skip(s, stbi__get32be(s) );
+
+   // Find out if the data is compressed.
+   // Known values:
+   //   0: no compression
+   //   1: RLE compressed
+   compression = stbi__get16be(s);
+   if (compression > 1)
+      return stbi__errpuc("bad compression", "PSD has an unknown compression format");
+
+   // Check size
+   if (!stbi__mad3sizes_valid(4, w, h, 0))
+      return stbi__errpuc("too large", "Corrupt PSD");
+
+   // Create the destination image.
+
+   if (!compression && bitdepth == 16 && bpc == 16) {
+      out = (stbi_uc *) stbi__malloc_mad3(8, w, h, 0);
+      ri->bits_per_channel = 16;
+   } else
+      out = (stbi_uc *) stbi__malloc(4 * w*h);
+
+   if (!out) return stbi__errpuc("outofmem", "Out of memory");
+   pixelCount = w*h;
+
+   // Initialize the data to zero.
+   //memset( out, 0, pixelCount * 4 );
+
+   // Finally, the image data.
+   if (compression) {
+      // RLE as used by .PSD and .TIFF
+      // Loop until you get the number of unpacked bytes you are expecting:
+      //     Read the next source byte into n.
+      //     If n is between 0 and 127 inclusive, copy the next n+1 bytes literally.
+      //     Else if n is between -127 and -1 inclusive, copy the next byte -n+1 times.
+      //     Else if n is 128, noop.
+      // Endloop
+
+      // The RLE-compressed data is preceded by a 2-byte data count for each row in the data,
+      // which we're going to just skip.
+      stbi__skip(s, h * channelCount * 2 );
+
+      // Read the RLE data by channel.
+      for (channel = 0; channel < 4; channel++) {
+         stbi_uc *p;
+
+         p = out+channel;
+         if (channel >= channelCount) {
+            // Fill this channel with default data.
+            for (i = 0; i < pixelCount; i++, p += 4)
+               *p = (channel == 3 ? 255 : 0);
+         } else {
+            // Read the RLE data.
+            if (!stbi__psd_decode_rle(s, p, pixelCount)) {
+               STBI_FREE(out);
+               return stbi__errpuc("corrupt", "bad RLE data");
+            }
+         }
+      }
+
+   } else {
+      // We're at the raw image data.  It's each channel in order (Red, Green, Blue, Alpha, ...)
+      // where each channel consists of an 8-bit (or 16-bit) value for each pixel in the image.
+
+      // Read the data by channel.
+      for (channel = 0; channel < 4; channel++) {
+         if (channel >= channelCount) {
+            // Fill this channel with default data.
+            if (bitdepth == 16 && bpc == 16) {
+               stbi__uint16 *q = ((stbi__uint16 *) out) + channel;
+               stbi__uint16 val = channel == 3 ? 65535 : 0;
+               for (i = 0; i < pixelCount; i++, q += 4)
+                  *q = val;
+            } else {
+               stbi_uc *p = out+channel;
+               stbi_uc val = channel == 3 ? 255 : 0;
+               for (i = 0; i < pixelCount; i++, p += 4)
+                  *p = val;
+            }
+         } else {
+            if (ri->bits_per_channel == 16) {    // output bpc
+               stbi__uint16 *q = ((stbi__uint16 *) out) + channel;
+               for (i = 0; i < pixelCount; i++, q += 4)
+                  *q = (stbi__uint16) stbi__get16be(s);
+            } else {
+               stbi_uc *p = out+channel;
+               if (bitdepth == 16) {  // input bpc
+                  for (i = 0; i < pixelCount; i++, p += 4)
+                     *p = (stbi_uc) (stbi__get16be(s) >> 8);
+               } else {
+                  for (i = 0; i < pixelCount; i++, p += 4)
+                     *p = stbi__get8(s);
+               }
+            }
+         }
+      }
+   }
+
+   // remove weird white matte from PSD
+   if (channelCount >= 4) {
+      if (ri->bits_per_channel == 16) {
+         for (i=0; i < w*h; ++i) {
+            stbi__uint16 *pixel = (stbi__uint16 *) out + 4*i;
+            if (pixel[3] != 0 && pixel[3] != 65535) {
+               float a = pixel[3] / 65535.0f;
+               float ra = 1.0f / a;
+               float inv_a = 65535.0f * (1 - ra);
+               pixel[0] = (stbi__uint16) (pixel[0]*ra + inv_a);
+               pixel[1] = (stbi__uint16) (pixel[1]*ra + inv_a);
+               pixel[2] = (stbi__uint16) (pixel[2]*ra + inv_a);
+            }
+         }
+      } else {
+         for (i=0; i < w*h; ++i) {
+            unsigned char *pixel = out + 4*i;
+            if (pixel[3] != 0 && pixel[3] != 255) {
+               float a = pixel[3] / 255.0f;
+               float ra = 1.0f / a;
+               float inv_a = 255.0f * (1 - ra);
+               pixel[0] = (unsigned char) (pixel[0]*ra + inv_a);
+               pixel[1] = (unsigned char) (pixel[1]*ra + inv_a);
+               pixel[2] = (unsigned char) (pixel[2]*ra + inv_a);
+            }
+         }
+      }
+   }
+
+   // convert to desired output format
+   if (req_comp && req_comp != 4) {
+      if (ri->bits_per_channel == 16)
+         out = (stbi_uc *) stbi__convert_format16((stbi__uint16 *) out, 4, req_comp, w, h);
+      else
+         out = stbi__convert_format(out, 4, req_comp, w, h);
+      if (out == NULL) return out; // stbi__convert_format frees input on failure
+   }
+
+   if (comp) *comp = 4;
+   *y = h;
+   *x = w;
+
+   return out;
+}
+#endif
+
+// *************************************************************************************************
+// Softimage PIC loader
+// by Tom Seddon
+//
+// See http://softimage.wiki.softimage.com/index.php/INFO:_PIC_file_format
+// See http://ozviz.wasp.uwa.edu.au/~pbourke/dataformats/softimagepic/
+
+#ifndef STBI_NO_PIC
+static int stbi__pic_is4(stbi__context *s,const char *str)
+{
+   int i;
+   for (i=0; i<4; ++i)
+      if (stbi__get8(s) != (stbi_uc)str[i])
+         return 0;
+
+   return 1;
+}
+
+static int stbi__pic_test_core(stbi__context *s)
+{
+   int i;
+
+   if (!stbi__pic_is4(s,"\x53\x80\xF6\x34"))
+      return 0;
+
+   for(i=0;i<84;++i)
+      stbi__get8(s);
+
+   if (!stbi__pic_is4(s,"PICT"))
+      return 0;
+
+   return 1;
+}
+
+typedef struct
+{
+   stbi_uc size,type,channel;
+} stbi__pic_packet;
+
+static stbi_uc *stbi__readval(stbi__context *s, int channel, stbi_uc *dest)
+{
+   int mask=0x80, i;
+
+   for (i=0; i<4; ++i, mask>>=1) {
+      if (channel & mask) {
+         if (stbi__at_eof(s)) return stbi__errpuc("bad file","PIC file too short");
+         dest[i]=stbi__get8(s);
+      }
+   }
+
+   return dest;
+}
+
+static void stbi__copyval(int channel,stbi_uc *dest,const stbi_uc *src)
+{
+   int mask=0x80,i;
+
+   for (i=0;i<4; ++i, mask>>=1)
+      if (channel&mask)
+         dest[i]=src[i];
+}
+
+static stbi_uc *stbi__pic_load_core(stbi__context *s,int width,int height,int *comp, stbi_uc *result)
+{
+   int act_comp=0,num_packets=0,y,chained;
+   stbi__pic_packet packets[10];
+
+   // this will (should...) cater for even some bizarre stuff like having data
+    // for the same channel in multiple packets.
+   do {
+      stbi__pic_packet *packet;
+
+      if (num_packets==sizeof(packets)/sizeof(packets[0]))
+         return stbi__errpuc("bad format","too many packets");
+
+      packet = &packets[num_packets++];
+
+      chained = stbi__get8(s);
+      packet->size    = stbi__get8(s);
+      packet->type    = stbi__get8(s);
+      packet->channel = stbi__get8(s);
+
+      act_comp |= packet->channel;
+
+      if (stbi__at_eof(s))          return stbi__errpuc("bad file","file too short (reading packets)");
+      if (packet->size != 8)  return stbi__errpuc("bad format","packet isn't 8bpp");
+   } while (chained);
+
+   *comp = (act_comp & 0x10 ? 4 : 3); // has alpha channel?
+
+   for(y=0; y<height; ++y) {
+      int packet_idx;
+
+      for(packet_idx=0; packet_idx < num_packets; ++packet_idx) {
+         stbi__pic_packet *packet = &packets[packet_idx];
+         stbi_uc *dest = result+y*width*4;
+
+         switch (packet->type) {
+            default:
+               return stbi__errpuc("bad format","packet has bad compression type");
+
+            case 0: {//uncompressed
+               int x;
+
+               for(x=0;x<width;++x, dest+=4)
+                  if (!stbi__readval(s,packet->channel,dest))
+                     return 0;
+               break;
+            }
+
+            case 1://Pure RLE
+               {
+                  int left=width, i;
+
+                  while (left>0) {
+                     stbi_uc count,value[4];
+
+                     count=stbi__get8(s);
+                     if (stbi__at_eof(s))   return stbi__errpuc("bad file","file too short (pure read count)");
+
+                     if (count > left)
+                        count = (stbi_uc) left;
+
+                     if (!stbi__readval(s,packet->channel,value))  return 0;
+
+                     for(i=0; i<count; ++i,dest+=4)
+                        stbi__copyval(packet->channel,dest,value);
+                     left -= count;
+                  }
+               }
+               break;
+
+            case 2: {//Mixed RLE
+               int left=width;
+               while (left>0) {
+                  int count = stbi__get8(s), i;
+                  if (stbi__at_eof(s))  return stbi__errpuc("bad file","file too short (mixed read count)");
+
+                  if (count >= 128) { // Repeated
+                     stbi_uc value[4];
+
+                     if (count==128)
+                        count = stbi__get16be(s);
+                     else
+                        count -= 127;
+                     if (count > left)
+                        return stbi__errpuc("bad file","scanline overrun");
+
+                     if (!stbi__readval(s,packet->channel,value))
+                        return 0;
+
+                     for(i=0;i<count;++i, dest += 4)
+                        stbi__copyval(packet->channel,dest,value);
+                  } else { // Raw
+                     ++count;
+                     if (count>left) return stbi__errpuc("bad file","scanline overrun");
+
+                     for(i=0;i<count;++i, dest+=4)
+                        if (!stbi__readval(s,packet->channel,dest))
+                           return 0;
+                  }
+                  left-=count;
+               }
+               break;
+            }
+         }
+      }
+   }
+
+   return result;
+}
+
+static void *stbi__pic_load(stbi__context *s,int *px,int *py,int *comp,int req_comp, stbi__result_info *ri)
+{
+   stbi_uc *result;
+   int i, x,y, internal_comp;
+   STBI_NOTUSED(ri);
+
+   if (!comp) comp = &internal_comp;
+
+   for (i=0; i<92; ++i)
+      stbi__get8(s);
+
+   x = stbi__get16be(s);
+   y = stbi__get16be(s);
+
+   if (y > STBI_MAX_DIMENSIONS) return stbi__errpuc("too large","Very large image (corrupt?)");
+   if (x > STBI_MAX_DIMENSIONS) return stbi__errpuc("too large","Very large image (corrupt?)");
+
+   if (stbi__at_eof(s))  return stbi__errpuc("bad file","file too short (pic header)");
+   if (!stbi__mad3sizes_valid(x, y, 4, 0)) return stbi__errpuc("too large", "PIC image too large to decode");
+
+   stbi__get32be(s); //skip `ratio'
+   stbi__get16be(s); //skip `fields'
+   stbi__get16be(s); //skip `pad'
+
+   // intermediate buffer is RGBA
+   result = (stbi_uc *) stbi__malloc_mad3(x, y, 4, 0);
+   if (!result) return stbi__errpuc("outofmem", "Out of memory");
+   memset(result, 0xff, x*y*4);
+
+   if (!stbi__pic_load_core(s,x,y,comp, result)) {
+      STBI_FREE(result);
+      result=0;
+   }
+   *px = x;
+   *py = y;
+   if (req_comp == 0) req_comp = *comp;
+   result=stbi__convert_format(result,4,req_comp,x,y);
+
+   return result;
+}
+
+static int stbi__pic_test(stbi__context *s)
+{
+   int r = stbi__pic_test_core(s);
+   stbi__rewind(s);
+   return r;
+}
+#endif
+
+// *************************************************************************************************
+// GIF loader -- public domain by Jean-Marc Lienher -- simplified/shrunk by stb
+
+#ifndef STBI_NO_GIF
+typedef struct
+{
+   stbi__int16 prefix;
+   stbi_uc first;
+   stbi_uc suffix;
+} stbi__gif_lzw;
+
+typedef struct
+{
+   int w,h;
+   stbi_uc *out;                 // output buffer (always 4 components)
+   stbi_uc *background;          // The current "background" as far as a gif is concerned
+   stbi_uc *history;
+   int flags, bgindex, ratio, transparent, eflags;
+   stbi_uc  pal[256][4];
+   stbi_uc lpal[256][4];
+   stbi__gif_lzw codes[8192];
+   stbi_uc *color_table;
+   int parse, step;
+   int lflags;
+   int start_x, start_y;
+   int max_x, max_y;
+   int cur_x, cur_y;
+   int line_size;
+   int delay;
+} stbi__gif;
+
+static int stbi__gif_test_raw(stbi__context *s)
+{
+   int sz;
+   if (stbi__get8(s) != 'G' || stbi__get8(s) != 'I' || stbi__get8(s) != 'F' || stbi__get8(s) != '8') return 0;
+   sz = stbi__get8(s);
+   if (sz != '9' && sz != '7') return 0;
+   if (stbi__get8(s) != 'a') return 0;
+   return 1;
+}
+
+static int stbi__gif_test(stbi__context *s)
+{
+   int r = stbi__gif_test_raw(s);
+   stbi__rewind(s);
+   return r;
+}
+
+static void stbi__gif_parse_colortable(stbi__context *s, stbi_uc pal[256][4], int num_entries, int transp)
+{
+   int i;
+   for (i=0; i < num_entries; ++i) {
+      pal[i][2] = stbi__get8(s);
+      pal[i][1] = stbi__get8(s);
+      pal[i][0] = stbi__get8(s);
+      pal[i][3] = transp == i ? 0 : 255;
+   }
+}
+
+static int stbi__gif_header(stbi__context *s, stbi__gif *g, int *comp, int is_info)
+{
+   stbi_uc version;
+   if (stbi__get8(s) != 'G' || stbi__get8(s) != 'I' || stbi__get8(s) != 'F' || stbi__get8(s) != '8')
+      return stbi__err("not GIF", "Corrupt GIF");
+
+   version = stbi__get8(s);
+   if (version != '7' && version != '9')    return stbi__err("not GIF", "Corrupt GIF");
+   if (stbi__get8(s) != 'a')                return stbi__err("not GIF", "Corrupt GIF");
+
+   stbi__g_failure_reason = "";
+   g->w = stbi__get16le(s);
+   g->h = stbi__get16le(s);
+   g->flags = stbi__get8(s);
+   g->bgindex = stbi__get8(s);
+   g->ratio = stbi__get8(s);
+   g->transparent = -1;
+
+   if (g->w > STBI_MAX_DIMENSIONS) return stbi__err("too large","Very large image (corrupt?)");
+   if (g->h > STBI_MAX_DIMENSIONS) return stbi__err("too large","Very large image (corrupt?)");
+
+   if (comp != 0) *comp = 4;  // can't actually tell whether it's 3 or 4 until we parse the comments
+
+   if (is_info) return 1;
+
+   if (g->flags & 0x80)
+      stbi__gif_parse_colortable(s,g->pal, 2 << (g->flags & 7), -1);
+
+   return 1;
+}
+
+static int stbi__gif_info_raw(stbi__context *s, int *x, int *y, int *comp)
+{
+   stbi__gif* g = (stbi__gif*) stbi__malloc(sizeof(stbi__gif));
+   if (!g) return stbi__err("outofmem", "Out of memory");
+   if (!stbi__gif_header(s, g, comp, 1)) {
+      STBI_FREE(g);
+      stbi__rewind( s );
+      return 0;
+   }
+   if (x) *x = g->w;
+   if (y) *y = g->h;
+   STBI_FREE(g);
+   return 1;
+}
+
+static void stbi__out_gif_code(stbi__gif *g, stbi__uint16 code)
+{
+   stbi_uc *p, *c;
+   int idx;
+
+   // recurse to decode the prefixes, since the linked-list is backwards,
+   // and working backwards through an interleaved image would be nasty
+   if (g->codes[code].prefix >= 0)
+      stbi__out_gif_code(g, g->codes[code].prefix);
+
+   if (g->cur_y >= g->max_y) return;
+
+   idx = g->cur_x + g->cur_y;
+   p = &g->out[idx];
+   g->history[idx / 4] = 1;
+
+   c = &g->color_table[g->codes[code].suffix * 4];
+   if (c[3] > 128) { // don't render transparent pixels;
+      p[0] = c[2];
+      p[1] = c[1];
+      p[2] = c[0];
+      p[3] = c[3];
+   }
+   g->cur_x += 4;
+
+   if (g->cur_x >= g->max_x) {
+      g->cur_x = g->start_x;
+      g->cur_y += g->step;
+
+      while (g->cur_y >= g->max_y && g->parse > 0) {
+         g->step = (1 << g->parse) * g->line_size;
+         g->cur_y = g->start_y + (g->step >> 1);
+         --g->parse;
+      }
+   }
+}
+
+static stbi_uc *stbi__process_gif_raster(stbi__context *s, stbi__gif *g)
+{
+   stbi_uc lzw_cs;
+   stbi__int32 len, init_code;
+   stbi__uint32 first;
+   stbi__int32 codesize, codemask, avail, oldcode, bits, valid_bits, clear;
+   stbi__gif_lzw *p;
+
+   lzw_cs = stbi__get8(s);
+   if (lzw_cs > 12) return NULL;
+   clear = 1 << lzw_cs;
+   first = 1;
+   codesize = lzw_cs + 1;
+   codemask = (1 << codesize) - 1;
+   bits = 0;
+   valid_bits = 0;
+   for (init_code = 0; init_code < clear; init_code++) {
+      g->codes[init_code].prefix = -1;
+      g->codes[init_code].first = (stbi_uc) init_code;
+      g->codes[init_code].suffix = (stbi_uc) init_code;
+   }
+
+   // support no starting clear code
+   avail = clear+2;
+   oldcode = -1;
+
+   len = 0;
+   for(;;) {
+      if (valid_bits < codesize) {
+         if (len == 0) {
+            len = stbi__get8(s); // start new block
+            if (len == 0)
+               return g->out;
+         }
+         --len;
+         bits |= (stbi__int32) stbi__get8(s) << valid_bits;
+         valid_bits += 8;
+      } else {
+         stbi__int32 code = bits & codemask;
+         bits >>= codesize;
+         valid_bits -= codesize;
+         // @OPTIMIZE: is there some way we can accelerate the non-clear path?
+         if (code == clear) {  // clear code
+            codesize = lzw_cs + 1;
+            codemask = (1 << codesize) - 1;
+            avail = clear + 2;
+            oldcode = -1;
+            first = 0;
+         } else if (code == clear + 1) { // end of stream code
+            stbi__skip(s, len);
+            while ((len = stbi__get8(s)) > 0)
+               stbi__skip(s,len);
+            return g->out;
+         } else if (code <= avail) {
+            if (first) {
+               return stbi__errpuc("no clear code", "Corrupt GIF");
+            }
+
+            if (oldcode >= 0) {
+               p = &g->codes[avail++];
+               if (avail > 8192) {
+                  return stbi__errpuc("too many codes", "Corrupt GIF");
+               }
+
+               p->prefix = (stbi__int16) oldcode;
+               p->first = g->codes[oldcode].first;
+               p->suffix = (code == avail) ? p->first : g->codes[code].first;
+            } else if (code == avail)
+               return stbi__errpuc("illegal code in raster", "Corrupt GIF");
+
+            stbi__out_gif_code(g, (stbi__uint16) code);
+
+            if ((avail & codemask) == 0 && avail <= 0x0FFF) {
+               codesize++;
+               codemask = (1 << codesize) - 1;
+            }
+
+            oldcode = code;
+         } else {
+            return stbi__errpuc("illegal code in raster", "Corrupt GIF");
+         }
+      }
+   }
+}
+
+// this function is designed to support animated gifs, although stb_image doesn't support it
+// two back is the image from two frames ago, used for a very specific disposal format
+static stbi_uc *stbi__gif_load_next(stbi__context *s, stbi__gif *g, int *comp, int req_comp, stbi_uc *two_back)
+{
+   int dispose;
+   int first_frame;
+   int pi;
+   int pcount;
+   STBI_NOTUSED(req_comp);
+
+   // on first frame, any non-written pixels get the background colour (non-transparent)
+   first_frame = 0;
+   if (g->out == 0) {
+      if (!stbi__gif_header(s, g, comp,0)) return 0; // stbi__g_failure_reason set by stbi__gif_header
+      if (!stbi__mad3sizes_valid(4, g->w, g->h, 0))
+         return stbi__errpuc("too large", "GIF image is too large");
+      pcount = g->w * g->h;
+      g->out = (stbi_uc *) stbi__malloc(4 * pcount);
+      g->background = (stbi_uc *) stbi__malloc(4 * pcount);
+      g->history = (stbi_uc *) stbi__malloc(pcount);
+      if (!g->out || !g->background || !g->history)
+         return stbi__errpuc("outofmem", "Out of memory");
+
+      // image is treated as "transparent" at the start - ie, nothing overwrites the current background;
+      // background colour is only used for pixels that are not rendered first frame, after that "background"
+      // color refers to the color that was there the previous frame.
+      memset(g->out, 0x00, 4 * pcount);
+      memset(g->background, 0x00, 4 * pcount); // state of the background (starts transparent)
+      memset(g->history, 0x00, pcount);        // pixels that were affected previous frame
+      first_frame = 1;
+   } else {
+      // second frame - how do we dispose of the previous one?
+      dispose = (g->eflags & 0x1C) >> 2;
+      pcount = g->w * g->h;
+
+      if ((dispose == 3) && (two_back == 0)) {
+         dispose = 2; // if I don't have an image to revert back to, default to the old background
+      }
+
+      if (dispose == 3) { // use previous graphic
+         for (pi = 0; pi < pcount; ++pi) {
+            if (g->history[pi]) {
+               memcpy( &g->out[pi * 4], &two_back[pi * 4], 4 );
+            }
+         }
+      } else if (dispose == 2) {
+         // restore what was changed last frame to background before that frame;
+         for (pi = 0; pi < pcount; ++pi) {
+            if (g->history[pi]) {
+               memcpy( &g->out[pi * 4], &g->background[pi * 4], 4 );
+            }
+         }
+      } else {
+         // This is a non-disposal case eithe way, so just
+         // leave the pixels as is, and they will become the new background
+         // 1: do not dispose
+         // 0:  not specified.
+      }
+
+      // background is what out is after the undoing of the previou frame;
+      memcpy( g->background, g->out, 4 * g->w * g->h );
+   }
+
+   // clear my history;
+   memset( g->history, 0x00, g->w * g->h );        // pixels that were affected previous frame
+
+   for (;;) {
+      int tag = stbi__get8(s);
+      switch (tag) {
+         case 0x2C: /* Image Descriptor */
+         {
+            stbi__int32 x, y, w, h;
+            stbi_uc *o;
+
+            x = stbi__get16le(s);
+            y = stbi__get16le(s);
+            w = stbi__get16le(s);
+            h = stbi__get16le(s);
+            if (((x + w) > (g->w)) || ((y + h) > (g->h)))
+               return stbi__errpuc("bad Image Descriptor", "Corrupt GIF");
+
+            g->line_size = g->w * 4;
+            g->start_x = x * 4;
+            g->start_y = y * g->line_size;
+            g->max_x   = g->start_x + w * 4;
+            g->max_y   = g->start_y + h * g->line_size;
+            g->cur_x   = g->start_x;
+            g->cur_y   = g->start_y;
+
+            // if the width of the specified rectangle is 0, that means
+            // we may not see *any* pixels or the image is malformed;
+            // to make sure this is caught, move the current y down to
+            // max_y (which is what out_gif_code checks).
+            if (w == 0)
+               g->cur_y = g->max_y;
+
+            g->lflags = stbi__get8(s);
+
+            if (g->lflags & 0x40) {
+               g->step = 8 * g->line_size; // first interlaced spacing
+               g->parse = 3;
+            } else {
+               g->step = g->line_size;
+               g->parse = 0;
+            }
+
+            if (g->lflags & 0x80) {
+               stbi__gif_parse_colortable(s,g->lpal, 2 << (g->lflags & 7), g->eflags & 0x01 ? g->transparent : -1);
+               g->color_table = (stbi_uc *) g->lpal;
+            } else if (g->flags & 0x80) {
+               g->color_table = (stbi_uc *) g->pal;
+            } else
+               return stbi__errpuc("missing color table", "Corrupt GIF");
+
+            o = stbi__process_gif_raster(s, g);
+            if (!o) return NULL;
+
+            // if this was the first frame,
+            pcount = g->w * g->h;
+            if (first_frame && (g->bgindex > 0)) {
+               // if first frame, any pixel not drawn to gets the background color
+               for (pi = 0; pi < pcount; ++pi) {
+                  if (g->history[pi] == 0) {
+                     g->pal[g->bgindex][3] = 255; // just in case it was made transparent, undo that; It will be reset next frame if need be;
+                     memcpy( &g->out[pi * 4], &g->pal[g->bgindex], 4 );
+                  }
+               }
+            }
+
+            return o;
+         }
+
+         case 0x21: // Comment Extension.
+         {
+            int len;
+            int ext = stbi__get8(s);
+            if (ext == 0xF9) { // Graphic Control Extension.
+               len = stbi__get8(s);
+               if (len == 4) {
+                  g->eflags = stbi__get8(s);
+                  g->delay = 10 * stbi__get16le(s); // delay - 1/100th of a second, saving as 1/1000ths.
+
+                  // unset old transparent
+                  if (g->transparent >= 0) {
+                     g->pal[g->transparent][3] = 255;
+                  }
+                  if (g->eflags & 0x01) {
+                     g->transparent = stbi__get8(s);
+                     if (g->transparent >= 0) {
+                        g->pal[g->transparent][3] = 0;
+                     }
+                  } else {
+                     // don't need transparent
+                     stbi__skip(s, 1);
+                     g->transparent = -1;
+                  }
+               } else {
+                  stbi__skip(s, len);
+                  break;
+               }
+            }
+            while ((len = stbi__get8(s)) != 0) {
+               stbi__skip(s, len);
+            }
+            break;
+         }
+
+         case 0x3B: // gif stream termination code
+            return (stbi_uc *) s; // using '1' causes warning on some compilers
+
+         default:
+            return stbi__errpuc("unknown code", "Corrupt GIF");
+      }
+   }
+}
+
+static void *stbi__load_gif_main_outofmem(stbi__gif *g, stbi_uc *out, int **delays)
+{
+   STBI_FREE(g->out);
+   STBI_FREE(g->history);
+   STBI_FREE(g->background);
+
+   if (out) STBI_FREE(out);
+   if (delays && *delays) STBI_FREE(*delays);
+   return stbi__errpuc("outofmem", "Out of memory");
+}
+
+static void *stbi__load_gif_main(stbi__context *s, int **delays, int *x, int *y, int *z, int *comp, int req_comp)
+{
+   if (stbi__gif_test(s)) {
+      int layers = 0;
+      stbi_uc *u = 0;
+      stbi_uc *out = 0;
+      stbi_uc *two_back = 0;
+      stbi__gif g;
+      int stride;
+      int out_size = 0;
+      int delays_size = 0;
+
+      STBI_NOTUSED(out_size);
+      STBI_NOTUSED(delays_size);
+
+      memset(&g, 0, sizeof(g));
+      if (delays) {
+         *delays = 0;
+      }
+
+      do {
+         u = stbi__gif_load_next(s, &g, comp, req_comp, two_back);
+         if (u == (stbi_uc *) s) u = 0;  // end of animated gif marker
+
+         if (u) {
+            *x = g.w;
+            *y = g.h;
+            ++layers;
+            stride = g.w * g.h * 4;
+
+            if (out) {
+               void *tmp = (stbi_uc*) STBI_REALLOC_SIZED( out, out_size, layers * stride );
+               if (!tmp)
+                  return stbi__load_gif_main_outofmem(&g, out, delays);
+               else {
+                   out = (stbi_uc*) tmp;
+                   out_size = layers * stride;
+               }
+
+               if (delays) {
+                  int *new_delays = (int*) STBI_REALLOC_SIZED( *delays, delays_size, sizeof(int) * layers );
+                  if (!new_delays)
+                     return stbi__load_gif_main_outofmem(&g, out, delays);
+                  *delays = new_delays;
+                  delays_size = layers * sizeof(int);
+               }
+            } else {
+               out = (stbi_uc*)stbi__malloc( layers * stride );
+               if (!out)
+                  return stbi__load_gif_main_outofmem(&g, out, delays);
+               out_size = layers * stride;
+               if (delays) {
+                  *delays = (int*) stbi__malloc( layers * sizeof(int) );
+                  if (!*delays)
+                     return stbi__load_gif_main_outofmem(&g, out, delays);
+                  delays_size = layers * sizeof(int);
+               }
+            }
+            memcpy( out + ((layers - 1) * stride), u, stride );
+            if (layers >= 2) {
+               two_back = out - 2 * stride;
+            }
+
+            if (delays) {
+               (*delays)[layers - 1U] = g.delay;
+            }
+         }
+      } while (u != 0);
+
+      // free temp buffer;
+      STBI_FREE(g.out);
+      STBI_FREE(g.history);
+      STBI_FREE(g.background);
+
+      // do the final conversion after loading everything;
+      if (req_comp && req_comp != 4)
+         out = stbi__convert_format(out, 4, req_comp, layers * g.w, g.h);
+
+      *z = layers;
+      return out;
+   } else {
+      return stbi__errpuc("not GIF", "Image was not as a gif type.");
+   }
+}
+
+static void *stbi__gif_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri)
+{
+   stbi_uc *u = 0;
+   stbi__gif g;
+   memset(&g, 0, sizeof(g));
+   STBI_NOTUSED(ri);
+
+   u = stbi__gif_load_next(s, &g, comp, req_comp, 0);
+   if (u == (stbi_uc *) s) u = 0;  // end of animated gif marker
+   if (u) {
+      *x = g.w;
+      *y = g.h;
+
+      // moved conversion to after successful load so that the same
+      // can be done for multiple frames.
+      if (req_comp && req_comp != 4)
+         u = stbi__convert_format(u, 4, req_comp, g.w, g.h);
+   } else if (g.out) {
+      // if there was an error and we allocated an image buffer, free it!
+      STBI_FREE(g.out);
+   }
+
+   // free buffers needed for multiple frame loading;
+   STBI_FREE(g.history);
+   STBI_FREE(g.background);
+
+   return u;
+}
+
+static int stbi__gif_info(stbi__context *s, int *x, int *y, int *comp)
+{
+   return stbi__gif_info_raw(s,x,y,comp);
+}
+#endif
+
+// *************************************************************************************************
+// Radiance RGBE HDR loader
+// originally by Nicolas Schulz
+#ifndef STBI_NO_HDR
+static int stbi__hdr_test_core(stbi__context *s, const char *signature)
+{
+   int i;
+   for (i=0; signature[i]; ++i)
+      if (stbi__get8(s) != signature[i])
+          return 0;
+   stbi__rewind(s);
+   return 1;
+}
+
+static int stbi__hdr_test(stbi__context* s)
+{
+   int r = stbi__hdr_test_core(s, "#?RADIANCE\n");
+   stbi__rewind(s);
+   if(!r) {
+       r = stbi__hdr_test_core(s, "#?RGBE\n");
+       stbi__rewind(s);
+   }
+   return r;
+}
+
+#define STBI__HDR_BUFLEN  1024
+static char *stbi__hdr_gettoken(stbi__context *z, char *buffer)
+{
+   int len=0;
+   char c = '\0';
+
+   c = (char) stbi__get8(z);
+
+   while (!stbi__at_eof(z) && c != '\n') {
+      buffer[len++] = c;
+      if (len == STBI__HDR_BUFLEN-1) {
+         // flush to end of line
+         while (!stbi__at_eof(z) && stbi__get8(z) != '\n')
+            ;
+         break;
+      }
+      c = (char) stbi__get8(z);
+   }
+
+   buffer[len] = 0;
+   return buffer;
+}
+
+static void stbi__hdr_convert(float *output, stbi_uc *input, int req_comp)
+{
+   if ( input[3] != 0 ) {
+      float f1;
+      // Exponent
+      f1 = (float) ldexp(1.0f, input[3] - (int)(128 + 8));
+      if (req_comp <= 2)
+         output[0] = (input[0] + input[1] + input[2]) * f1 / 3;
+      else {
+         output[0] = input[0] * f1;
+         output[1] = input[1] * f1;
+         output[2] = input[2] * f1;
+      }
+      if (req_comp == 2) output[1] = 1;
+      if (req_comp == 4) output[3] = 1;
+   } else {
+      switch (req_comp) {
+         case 4: output[3] = 1; /* fallthrough */
+         case 3: output[0] = output[1] = output[2] = 0;
+                 break;
+         case 2: output[1] = 1; /* fallthrough */
+         case 1: output[0] = 0;
+                 break;
+      }
+   }
+}
+
+static float *stbi__hdr_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri)
+{
+   char buffer[STBI__HDR_BUFLEN];
+   char *token;
+   int valid = 0;
+   int width, height;
+   stbi_uc *scanline;
+   float *hdr_data;
+   int len;
+   unsigned char count, value;
+   int i, j, k, c1,c2, z;
+   const char *headerToken;
+   STBI_NOTUSED(ri);
+
+   // Check identifier
+   headerToken = stbi__hdr_gettoken(s,buffer);
+   if (strcmp(headerToken, "#?RADIANCE") != 0 && strcmp(headerToken, "#?RGBE") != 0)
+      return stbi__errpf("not HDR", "Corrupt HDR image");
+
+   // Parse header
+   for(;;) {
+      token = stbi__hdr_gettoken(s,buffer);
+      if (token[0] == 0) break;
+      if (strcmp(token, "FORMAT=32-bit_rle_rgbe") == 0) valid = 1;
+   }
+
+   if (!valid)    return stbi__errpf("unsupported format", "Unsupported HDR format");
+
+   // Parse width and height
+   // can't use sscanf() if we're not using stdio!
+   token = stbi__hdr_gettoken(s,buffer);
+   if (strncmp(token, "-Y ", 3))  return stbi__errpf("unsupported data layout", "Unsupported HDR format");
+   token += 3;
+   height = (int) strtol(token, &token, 10);
+   while (*token == ' ') ++token;
+   if (strncmp(token, "+X ", 3))  return stbi__errpf("unsupported data layout", "Unsupported HDR format");
+   token += 3;
+   width = (int) strtol(token, NULL, 10);
+
+   if (height > STBI_MAX_DIMENSIONS) return stbi__errpf("too large","Very large image (corrupt?)");
+   if (width > STBI_MAX_DIMENSIONS) return stbi__errpf("too large","Very large image (corrupt?)");
+
+   *x = width;
+   *y = height;
+
+   if (comp) *comp = 3;
+   if (req_comp == 0) req_comp = 3;
+
+   if (!stbi__mad4sizes_valid(width, height, req_comp, sizeof(float), 0))
+      return stbi__errpf("too large", "HDR image is too large");
+
+   // Read data
+   hdr_data = (float *) stbi__malloc_mad4(width, height, req_comp, sizeof(float), 0);
+   if (!hdr_data)
+      return stbi__errpf("outofmem", "Out of memory");
+
+   // Load image data
+   // image data is stored as some number of sca
+   if ( width < 8 || width >= 32768) {
+      // Read flat data
+      for (j=0; j < height; ++j) {
+         for (i=0; i < width; ++i) {
+            stbi_uc rgbe[4];
+           main_decode_loop:
+            stbi__getn(s, rgbe, 4);
+            stbi__hdr_convert(hdr_data + j * width * req_comp + i * req_comp, rgbe, req_comp);
+         }
+      }
+   } else {
+      // Read RLE-encoded data
+      scanline = NULL;
+
+      for (j = 0; j < height; ++j) {
+         c1 = stbi__get8(s);
+         c2 = stbi__get8(s);
+         len = stbi__get8(s);
+         if (c1 != 2 || c2 != 2 || (len & 0x80)) {
+            // not run-length encoded, so we have to actually use THIS data as a decoded
+            // pixel (note this can't be a valid pixel--one of RGB must be >= 128)
+            stbi_uc rgbe[4];
+            rgbe[0] = (stbi_uc) c1;
+            rgbe[1] = (stbi_uc) c2;
+            rgbe[2] = (stbi_uc) len;
+            rgbe[3] = (stbi_uc) stbi__get8(s);
+            stbi__hdr_convert(hdr_data, rgbe, req_comp);
+            i = 1;
+            j = 0;
+            STBI_FREE(scanline);
+            goto main_decode_loop; // yes, this makes no sense
+         }
+         len <<= 8;
+         len |= stbi__get8(s);
+         if (len != width) { STBI_FREE(hdr_data); STBI_FREE(scanline); return stbi__errpf("invalid decoded scanline length", "corrupt HDR"); }
+         if (scanline == NULL) {
+            scanline = (stbi_uc *) stbi__malloc_mad2(width, 4, 0);
+            if (!scanline) {
+               STBI_FREE(hdr_data);
+               return stbi__errpf("outofmem", "Out of memory");
+            }
+         }
+
+         for (k = 0; k < 4; ++k) {
+            int nleft;
+            i = 0;
+            while ((nleft = width - i) > 0) {
+               count = stbi__get8(s);
+               if (count > 128) {
+                  // Run
+                  value = stbi__get8(s);
+                  count -= 128;
+                  if ((count == 0) || (count > nleft)) { STBI_FREE(hdr_data); STBI_FREE(scanline); return stbi__errpf("corrupt", "bad RLE data in HDR"); }
+                  for (z = 0; z < count; ++z)
+                     scanline[i++ * 4 + k] = value;
+               } else {
+                  // Dump
+                  if ((count == 0) || (count > nleft)) { STBI_FREE(hdr_data); STBI_FREE(scanline); return stbi__errpf("corrupt", "bad RLE data in HDR"); }
+                  for (z = 0; z < count; ++z)
+                     scanline[i++ * 4 + k] = stbi__get8(s);
+               }
+            }
+         }
+         for (i=0; i < width; ++i)
+            stbi__hdr_convert(hdr_data+(j*width + i)*req_comp, scanline + i*4, req_comp);
+      }
+      if (scanline)
+         STBI_FREE(scanline);
+   }
+
+   return hdr_data;
+}
+
+static int stbi__hdr_info(stbi__context *s, int *x, int *y, int *comp)
+{
+   char buffer[STBI__HDR_BUFLEN];
+   char *token;
+   int valid = 0;
+   int dummy;
+
+   if (!x) x = &dummy;
+   if (!y) y = &dummy;
+   if (!comp) comp = &dummy;
+
+   if (stbi__hdr_test(s) == 0) {
+       stbi__rewind( s );
+       return 0;
+   }
+
+   for(;;) {
+      token = stbi__hdr_gettoken(s,buffer);
+      if (token[0] == 0) break;
+      if (strcmp(token, "FORMAT=32-bit_rle_rgbe") == 0) valid = 1;
+   }
+
+   if (!valid) {
+       stbi__rewind( s );
+       return 0;
+   }
+   token = stbi__hdr_gettoken(s,buffer);
+   if (strncmp(token, "-Y ", 3)) {
+       stbi__rewind( s );
+       return 0;
+   }
+   token += 3;
+   *y = (int) strtol(token, &token, 10);
+   while (*token == ' ') ++token;
+   if (strncmp(token, "+X ", 3)) {
+       stbi__rewind( s );
+       return 0;
+   }
+   token += 3;
+   *x = (int) strtol(token, NULL, 10);
+   *comp = 3;
+   return 1;
+}
+#endif // STBI_NO_HDR
+
+#ifndef STBI_NO_BMP
+static int stbi__bmp_info(stbi__context *s, int *x, int *y, int *comp)
+{
+   void *p;
+   stbi__bmp_data info;
+
+   info.all_a = 255;
+   p = stbi__bmp_parse_header(s, &info);
+   if (p == NULL) {
+      stbi__rewind( s );
+      return 0;
+   }
+   if (x) *x = s->img_x;
+   if (y) *y = s->img_y;
+   if (comp) {
+      if (info.bpp == 24 && info.ma == 0xff000000)
+         *comp = 3;
+      else
+         *comp = info.ma ? 4 : 3;
+   }
+   return 1;
+}
+#endif
+
+#ifndef STBI_NO_PSD
+static int stbi__psd_info(stbi__context *s, int *x, int *y, int *comp)
+{
+   int channelCount, dummy, depth;
+   if (!x) x = &dummy;
+   if (!y) y = &dummy;
+   if (!comp) comp = &dummy;
+   if (stbi__get32be(s) != 0x38425053) {
+       stbi__rewind( s );
+       return 0;
+   }
+   if (stbi__get16be(s) != 1) {
+       stbi__rewind( s );
+       return 0;
+   }
+   stbi__skip(s, 6);
+   channelCount = stbi__get16be(s);
+   if (channelCount < 0 || channelCount > 16) {
+       stbi__rewind( s );
+       return 0;
+   }
+   *y = stbi__get32be(s);
+   *x = stbi__get32be(s);
+   depth = stbi__get16be(s);
+   if (depth != 8 && depth != 16) {
+       stbi__rewind( s );
+       return 0;
+   }
+   if (stbi__get16be(s) != 3) {
+       stbi__rewind( s );
+       return 0;
+   }
+   *comp = 4;
+   return 1;
+}
+
+static int stbi__psd_is16(stbi__context *s)
+{
+   int channelCount, depth;
+   if (stbi__get32be(s) != 0x38425053) {
+       stbi__rewind( s );
+       return 0;
+   }
+   if (stbi__get16be(s) != 1) {
+       stbi__rewind( s );
+       return 0;
+   }
+   stbi__skip(s, 6);
+   channelCount = stbi__get16be(s);
+   if (channelCount < 0 || channelCount > 16) {
+       stbi__rewind( s );
+       return 0;
+   }
+   STBI_NOTUSED(stbi__get32be(s));
+   STBI_NOTUSED(stbi__get32be(s));
+   depth = stbi__get16be(s);
+   if (depth != 16) {
+       stbi__rewind( s );
+       return 0;
+   }
+   return 1;
+}
+#endif
+
+#ifndef STBI_NO_PIC
+static int stbi__pic_info(stbi__context *s, int *x, int *y, int *comp)
+{
+   int act_comp=0,num_packets=0,chained,dummy;
+   stbi__pic_packet packets[10];
+
+   if (!x) x = &dummy;
+   if (!y) y = &dummy;
+   if (!comp) comp = &dummy;
+
+   if (!stbi__pic_is4(s,"\x53\x80\xF6\x34")) {
+      stbi__rewind(s);
+      return 0;
+   }
+
+   stbi__skip(s, 88);
+
+   *x = stbi__get16be(s);
+   *y = stbi__get16be(s);
+   if (stbi__at_eof(s)) {
+      stbi__rewind( s);
+      return 0;
+   }
+   if ( (*x) != 0 && (1 << 28) / (*x) < (*y)) {
+      stbi__rewind( s );
+      return 0;
+   }
+
+   stbi__skip(s, 8);
+
+   do {
+      stbi__pic_packet *packet;
+
+      if (num_packets==sizeof(packets)/sizeof(packets[0]))
+         return 0;
+
+      packet = &packets[num_packets++];
+      chained = stbi__get8(s);
+      packet->size    = stbi__get8(s);
+      packet->type    = stbi__get8(s);
+      packet->channel = stbi__get8(s);
+      act_comp |= packet->channel;
+
+      if (stbi__at_eof(s)) {
+          stbi__rewind( s );
+          return 0;
+      }
+      if (packet->size != 8) {
+          stbi__rewind( s );
+          return 0;
+      }
+   } while (chained);
+
+   *comp = (act_comp & 0x10 ? 4 : 3);
+
+   return 1;
+}
+#endif
+
+// *************************************************************************************************
+// Portable Gray Map and Portable Pixel Map loader
+// by Ken Miller
+//
+// PGM: http://netpbm.sourceforge.net/doc/pgm.html
+// PPM: http://netpbm.sourceforge.net/doc/ppm.html
+//
+// Known limitations:
+//    Does not support comments in the header section
+//    Does not support ASCII image data (formats P2 and P3)
+
+#ifndef STBI_NO_PNM
+
+static int      stbi__pnm_test(stbi__context *s)
+{
+   char p, t;
+   p = (char) stbi__get8(s);
+   t = (char) stbi__get8(s);
+   if (p != 'P' || (t != '5' && t != '6')) {
+       stbi__rewind( s );
+       return 0;
+   }
+   return 1;
+}
+
+static void *stbi__pnm_load(stbi__context *s, int *x, int *y, int *comp, int req_comp, stbi__result_info *ri)
+{
+   stbi_uc *out;
+   STBI_NOTUSED(ri);
+
+   ri->bits_per_channel = stbi__pnm_info(s, (int *)&s->img_x, (int *)&s->img_y, (int *)&s->img_n);
+   if (ri->bits_per_channel == 0)
+      return 0;
+
+   if (s->img_y > STBI_MAX_DIMENSIONS) return stbi__errpuc("too large","Very large image (corrupt?)");
+   if (s->img_x > STBI_MAX_DIMENSIONS) return stbi__errpuc("too large","Very large image (corrupt?)");
+
+   *x = s->img_x;
+   *y = s->img_y;
+   if (comp) *comp = s->img_n;
+
+   if (!stbi__mad4sizes_valid(s->img_n, s->img_x, s->img_y, ri->bits_per_channel / 8, 0))
+      return stbi__errpuc("too large", "PNM too large");
+
+   out = (stbi_uc *) stbi__malloc_mad4(s->img_n, s->img_x, s->img_y, ri->bits_per_channel / 8, 0);
+   if (!out) return stbi__errpuc("outofmem", "Out of memory");
+   if (!stbi__getn(s, out, s->img_n * s->img_x * s->img_y * (ri->bits_per_channel / 8))) {
+      STBI_FREE(out);
+      return stbi__errpuc("bad PNM", "PNM file truncated");
+   }
+
+   if (req_comp && req_comp != s->img_n) {
+      if (ri->bits_per_channel == 16) {
+         out = (stbi_uc *) stbi__convert_format16((stbi__uint16 *) out, s->img_n, req_comp, s->img_x, s->img_y);
+      } else {
+         out = stbi__convert_format(out, s->img_n, req_comp, s->img_x, s->img_y);
+      }
+      if (out == NULL) return out; // stbi__convert_format frees input on failure
+   }
+   return out;
+}
+
+static int      stbi__pnm_isspace(char c)
+{
+   return c == ' ' || c == '\t' || c == '\n' || c == '\v' || c == '\f' || c == '\r';
+}
+
+static void     stbi__pnm_skip_whitespace(stbi__context *s, char *c)
+{
+   for (;;) {
+      while (!stbi__at_eof(s) && stbi__pnm_isspace(*c))
+         *c = (char) stbi__get8(s);
+
+      if (stbi__at_eof(s) || *c != '#')
+         break;
+
+      while (!stbi__at_eof(s) && *c != '\n' && *c != '\r' )
+         *c = (char) stbi__get8(s);
+   }
+}
+
+static int      stbi__pnm_isdigit(char c)
+{
+   return c >= '0' && c <= '9';
+}
+
+static int      stbi__pnm_getinteger(stbi__context *s, char *c)
+{
+   int value = 0;
+
+   while (!stbi__at_eof(s) && stbi__pnm_isdigit(*c)) {
+      value = value*10 + (*c - '0');
+      *c = (char) stbi__get8(s);
+      if((value > 214748364) || (value == 214748364 && *c > '7'))
+          return stbi__err("integer parse overflow", "Parsing an integer in the PPM header overflowed a 32-bit int");
+   }
+
+   return value;
+}
+
+static int      stbi__pnm_info(stbi__context *s, int *x, int *y, int *comp)
+{
+   int maxv, dummy;
+   char c, p, t;
+
+   if (!x) x = &dummy;
+   if (!y) y = &dummy;
+   if (!comp) comp = &dummy;
+
+   stbi__rewind(s);
+
+   // Get identifier
+   p = (char) stbi__get8(s);
+   t = (char) stbi__get8(s);
+   if (p != 'P' || (t != '5' && t != '6')) {
+       stbi__rewind(s);
+       return 0;
+   }
+
+   *comp = (t == '6') ? 3 : 1;  // '5' is 1-component .pgm; '6' is 3-component .ppm
+
+   c = (char) stbi__get8(s);
+   stbi__pnm_skip_whitespace(s, &c);
+
+   *x = stbi__pnm_getinteger(s, &c); // read width
+   if(*x == 0)
+       return stbi__err("invalid width", "PPM image header had zero or overflowing width");
+   stbi__pnm_skip_whitespace(s, &c);
+
+   *y = stbi__pnm_getinteger(s, &c); // read height
+   if (*y == 0)
+       return stbi__err("invalid width", "PPM image header had zero or overflowing width");
+   stbi__pnm_skip_whitespace(s, &c);
+
+   maxv = stbi__pnm_getinteger(s, &c);  // read max value
+   if (maxv > 65535)
+      return stbi__err("max value > 65535", "PPM image supports only 8-bit and 16-bit images");
+   else if (maxv > 255)
+      return 16;
+   else
+      return 8;
+}
+
+static int stbi__pnm_is16(stbi__context *s)
+{
+   if (stbi__pnm_info(s, NULL, NULL, NULL) == 16)
+	   return 1;
+   return 0;
+}
+#endif
+
+static int stbi__info_main(stbi__context *s, int *x, int *y, int *comp)
+{
+   #ifndef STBI_NO_JPEG
+   if (stbi__jpeg_info(s, x, y, comp)) return 1;
+   #endif
+
+   #ifndef STBI_NO_PNG
+   if (stbi__png_info(s, x, y, comp))  return 1;
+   #endif
+
+   #ifndef STBI_NO_GIF
+   if (stbi__gif_info(s, x, y, comp))  return 1;
+   #endif
+
+   #ifndef STBI_NO_BMP
+   if (stbi__bmp_info(s, x, y, comp))  return 1;
+   #endif
+
+   #ifndef STBI_NO_PSD
+   if (stbi__psd_info(s, x, y, comp))  return 1;
+   #endif
+
+   #ifndef STBI_NO_PIC
+   if (stbi__pic_info(s, x, y, comp))  return 1;
+   #endif
+
+   #ifndef STBI_NO_PNM
+   if (stbi__pnm_info(s, x, y, comp))  return 1;
+   #endif
+
+   #ifndef STBI_NO_HDR
+   if (stbi__hdr_info(s, x, y, comp))  return 1;
+   #endif
+
+   // test tga last because it's a crappy test!
+   #ifndef STBI_NO_TGA
+   if (stbi__tga_info(s, x, y, comp))
+       return 1;
+   #endif
+   return stbi__err("unknown image type", "Image not of any known type, or corrupt");
+}
+
+static int stbi__is_16_main(stbi__context *s)
+{
+   #ifndef STBI_NO_PNG
+   if (stbi__png_is16(s))  return 1;
+   #endif
+
+   #ifndef STBI_NO_PSD
+   if (stbi__psd_is16(s))  return 1;
+   #endif
+
+   #ifndef STBI_NO_PNM
+   if (stbi__pnm_is16(s))  return 1;
+   #endif
+   return 0;
+}
+
+#ifndef STBI_NO_STDIO
+STBIDEF int stbi_info(char const *filename, int *x, int *y, int *comp)
+{
+    FILE *f = stbi__fopen(filename, "rb");
+    int result;
+    if (!f) return stbi__err("can't fopen", "Unable to open file");
+    result = stbi_info_from_file(f, x, y, comp);
+    fclose(f);
+    return result;
+}
+
+STBIDEF int stbi_info_from_file(FILE *f, int *x, int *y, int *comp)
+{
+   int r;
+   stbi__context s;
+   long pos = ftell(f);
+   stbi__start_file(&s, f);
+   r = stbi__info_main(&s,x,y,comp);
+   fseek(f,pos,SEEK_SET);
+   return r;
+}
+
+STBIDEF int stbi_is_16_bit(char const *filename)
+{
+    FILE *f = stbi__fopen(filename, "rb");
+    int result;
+    if (!f) return stbi__err("can't fopen", "Unable to open file");
+    result = stbi_is_16_bit_from_file(f);
+    fclose(f);
+    return result;
+}
+
+STBIDEF int stbi_is_16_bit_from_file(FILE *f)
+{
+   int r;
+   stbi__context s;
+   long pos = ftell(f);
+   stbi__start_file(&s, f);
+   r = stbi__is_16_main(&s);
+   fseek(f,pos,SEEK_SET);
+   return r;
+}
+#endif // !STBI_NO_STDIO
+
+STBIDEF int stbi_info_from_memory(stbi_uc const *buffer, int len, int *x, int *y, int *comp)
+{
+   stbi__context s;
+   stbi__start_mem(&s,buffer,len);
+   return stbi__info_main(&s,x,y,comp);
+}
+
+STBIDEF int stbi_info_from_callbacks(stbi_io_callbacks const *c, void *user, int *x, int *y, int *comp)
+{
+   stbi__context s;
+   stbi__start_callbacks(&s, (stbi_io_callbacks *) c, user);
+   return stbi__info_main(&s,x,y,comp);
+}
+
+STBIDEF int stbi_is_16_bit_from_memory(stbi_uc const *buffer, int len)
+{
+   stbi__context s;
+   stbi__start_mem(&s,buffer,len);
+   return stbi__is_16_main(&s);
+}
+
+STBIDEF int stbi_is_16_bit_from_callbacks(stbi_io_callbacks const *c, void *user)
+{
+   stbi__context s;
+   stbi__start_callbacks(&s, (stbi_io_callbacks *) c, user);
+   return stbi__is_16_main(&s);
+}
+
+#endif // STB_IMAGE_IMPLEMENTATION
+
+/*
+   revision history:
+      2.20  (2019-02-07) support utf8 filenames in Windows; fix warnings and platform ifdefs
+      2.19  (2018-02-11) fix warning
+      2.18  (2018-01-30) fix warnings
+      2.17  (2018-01-29) change sbti__shiftsigned to avoid clang -O2 bug
+                         1-bit BMP
+                         *_is_16_bit api
+                         avoid warnings
+      2.16  (2017-07-23) all functions have 16-bit variants;
+                         STBI_NO_STDIO works again;
+                         compilation fixes;
+                         fix rounding in unpremultiply;
+                         optimize vertical flip;
+                         disable raw_len validation;
+                         documentation fixes
+      2.15  (2017-03-18) fix png-1,2,4 bug; now all Imagenet JPGs decode;
+                         warning fixes; disable run-time SSE detection on gcc;
+                         uniform handling of optional "return" values;
+                         thread-safe initialization of zlib tables
+      2.14  (2017-03-03) remove deprecated STBI_JPEG_OLD; fixes for Imagenet JPGs
+      2.13  (2016-11-29) add 16-bit API, only supported for PNG right now
+      2.12  (2016-04-02) fix typo in 2.11 PSD fix that caused crashes
+      2.11  (2016-04-02) allocate large structures on the stack
+                         remove white matting for transparent PSD
+                         fix reported channel count for PNG & BMP
+                         re-enable SSE2 in non-gcc 64-bit
+                         support RGB-formatted JPEG
+                         read 16-bit PNGs (only as 8-bit)
+      2.10  (2016-01-22) avoid warning introduced in 2.09 by STBI_REALLOC_SIZED
+      2.09  (2016-01-16) allow comments in PNM files
+                         16-bit-per-pixel TGA (not bit-per-component)
+                         info() for TGA could break due to .hdr handling
+                         info() for BMP to shares code instead of sloppy parse
+                         can use STBI_REALLOC_SIZED if allocator doesn't support realloc
+                         code cleanup
+      2.08  (2015-09-13) fix to 2.07 cleanup, reading RGB PSD as RGBA
+      2.07  (2015-09-13) fix compiler warnings
+                         partial animated GIF support
+                         limited 16-bpc PSD support
+                         #ifdef unused functions
+                         bug with < 92 byte PIC,PNM,HDR,TGA
+      2.06  (2015-04-19) fix bug where PSD returns wrong '*comp' value
+      2.05  (2015-04-19) fix bug in progressive JPEG handling, fix warning
+      2.04  (2015-04-15) try to re-enable SIMD on MinGW 64-bit
+      2.03  (2015-04-12) extra corruption checking (mmozeiko)
+                         stbi_set_flip_vertically_on_load (nguillemot)
+                         fix NEON support; fix mingw support
+      2.02  (2015-01-19) fix incorrect assert, fix warning
+      2.01  (2015-01-17) fix various warnings; suppress SIMD on gcc 32-bit without -msse2
+      2.00b (2014-12-25) fix STBI_MALLOC in progressive JPEG
+      2.00  (2014-12-25) optimize JPG, including x86 SSE2 & NEON SIMD (ryg)
+                         progressive JPEG (stb)
+                         PGM/PPM support (Ken Miller)
+                         STBI_MALLOC,STBI_REALLOC,STBI_FREE
+                         GIF bugfix -- seemingly never worked
+                         STBI_NO_*, STBI_ONLY_*
+      1.48  (2014-12-14) fix incorrectly-named assert()
+      1.47  (2014-12-14) 1/2/4-bit PNG support, both direct and paletted (Omar Cornut & stb)
+                         optimize PNG (ryg)
+                         fix bug in interlaced PNG with user-specified channel count (stb)
+      1.46  (2014-08-26)
+              fix broken tRNS chunk (colorkey-style transparency) in non-paletted PNG
+      1.45  (2014-08-16)
+              fix MSVC-ARM internal compiler error by wrapping malloc
+      1.44  (2014-08-07)
+              various warning fixes from Ronny Chevalier
+      1.43  (2014-07-15)
+              fix MSVC-only compiler problem in code changed in 1.42
+      1.42  (2014-07-09)
+              don't define _CRT_SECURE_NO_WARNINGS (affects user code)
+              fixes to stbi__cleanup_jpeg path
+              added STBI_ASSERT to avoid requiring assert.h
+      1.41  (2014-06-25)
+              fix search&replace from 1.36 that messed up comments/error messages
+      1.40  (2014-06-22)
+              fix gcc struct-initialization warning
+      1.39  (2014-06-15)
+              fix to TGA optimization when req_comp != number of components in TGA;
+              fix to GIF loading because BMP wasn't rewinding (whoops, no GIFs in my test suite)
+              add support for BMP version 5 (more ignored fields)
+      1.38  (2014-06-06)
+              suppress MSVC warnings on integer casts truncating values
+              fix accidental rename of 'skip' field of I/O
+      1.37  (2014-06-04)
+              remove duplicate typedef
+      1.36  (2014-06-03)
+              convert to header file single-file library
+              if de-iphone isn't set, load iphone images color-swapped instead of returning NULL
+      1.35  (2014-05-27)
+              various warnings
+              fix broken STBI_SIMD path
+              fix bug where stbi_load_from_file no longer left file pointer in correct place
+              fix broken non-easy path for 32-bit BMP (possibly never used)
+              TGA optimization by Arseny Kapoulkine
+      1.34  (unknown)
+              use STBI_NOTUSED in stbi__resample_row_generic(), fix one more leak in tga failure case
+      1.33  (2011-07-14)
+              make stbi_is_hdr work in STBI_NO_HDR (as specified), minor compiler-friendly improvements
+      1.32  (2011-07-13)
+              support for "info" function for all supported filetypes (SpartanJ)
+      1.31  (2011-06-20)
+              a few more leak fixes, bug in PNG handling (SpartanJ)
+      1.30  (2011-06-11)
+              added ability to load files via callbacks to accomidate custom input streams (Ben Wenger)
+              removed deprecated format-specific test/load functions
+              removed support for installable file formats (stbi_loader) -- would have been broken for IO callbacks anyway
+              error cases in bmp and tga give messages and don't leak (Raymond Barbiero, grisha)
+              fix inefficiency in decoding 32-bit BMP (David Woo)
+      1.29  (2010-08-16)
+              various warning fixes from Aurelien Pocheville
+      1.28  (2010-08-01)
+              fix bug in GIF palette transparency (SpartanJ)
+      1.27  (2010-08-01)
+              cast-to-stbi_uc to fix warnings
+      1.26  (2010-07-24)
+              fix bug in file buffering for PNG reported by SpartanJ
+      1.25  (2010-07-17)
+              refix trans_data warning (Won Chun)
+      1.24  (2010-07-12)
+              perf improvements reading from files on platforms with lock-heavy fgetc()
+              minor perf improvements for jpeg
+              deprecated type-specific functions so we'll get feedback if they're needed
+              attempt to fix trans_data warning (Won Chun)
+      1.23    fixed bug in iPhone support
+      1.22  (2010-07-10)
+              removed image *writing* support
+              stbi_info support from Jetro Lauha
+              GIF support from Jean-Marc Lienher
+              iPhone PNG-extensions from James Brown
+              warning-fixes from Nicolas Schulz and Janez Zemva (i.stbi__err. Janez (U+017D)emva)
+      1.21    fix use of 'stbi_uc' in header (reported by jon blow)
+      1.20    added support for Softimage PIC, by Tom Seddon
+      1.19    bug in interlaced PNG corruption check (found by ryg)
+      1.18  (2008-08-02)
+              fix a threading bug (local mutable static)
+      1.17    support interlaced PNG
+      1.16    major bugfix - stbi__convert_format converted one too many pixels
+      1.15    initialize some fields for thread safety
+      1.14    fix threadsafe conversion bug
+              header-file-only version (#define STBI_HEADER_FILE_ONLY before including)
+      1.13    threadsafe
+      1.12    const qualifiers in the API
+      1.11    Support installable IDCT, colorspace conversion routines
+      1.10    Fixes for 64-bit (don't use "unsigned long")
+              optimized upsampling by Fabian "ryg" Giesen
+      1.09    Fix format-conversion for PSD code (bad global variables!)
+      1.08    Thatcher Ulrich's PSD code integrated by Nicolas Schulz
+      1.07    attempt to fix C++ warning/errors again
+      1.06    attempt to fix C++ warning/errors again
+      1.05    fix TGA loading to return correct *comp and use good luminance calc
+      1.04    default float alpha is 1, not 255; use 'void *' for stbi_image_free
+      1.03    bugfixes to STBI_NO_STDIO, STBI_NO_HDR
+      1.02    support for (subset of) HDR files, float interface for preferred access to them
+      1.01    fix bug: possible bug in handling right-side up bmps... not sure
+              fix bug: the stbi__bmp_load() and stbi__tga_load() functions didn't work at all
+      1.00    interface to zlib that skips zlib header
+      0.99    correct handling of alpha in palette
+      0.98    TGA loader by lonesock; dynamically add loaders (untested)
+      0.97    jpeg errors on too large a file; also catch another malloc failure
+      0.96    fix detection of invalid v value - particleman@mollyrocket forum
+      0.95    during header scan, seek to markers in case of padding
+      0.94    STBI_NO_STDIO to disable stdio usage; rename all #defines the same
+      0.93    handle jpegtran output; verbose errors
+      0.92    read 4,8,16,24,32-bit BMP files of several formats
+      0.91    output 24-bit Windows 3.0 BMP files
+      0.90    fix a few more warnings; bump version number to approach 1.0
+      0.61    bugfixes due to Marc LeBlanc, Christopher Lloyd
+      0.60    fix compiling as c++
+      0.59    fix warnings: merge Dave Moore's -Wall fixes
+      0.58    fix bug: zlib uncompressed mode len/nlen was wrong endian
+      0.57    fix bug: jpg last huffman symbol before marker was >9 bits but less than 16 available
+      0.56    fix bug: zlib uncompressed mode len vs. nlen
+      0.55    fix bug: restart_interval not initialized to 0
+      0.54    allow NULL for 'int *comp'
+      0.53    fix bug in png 3->4; speedup png decoding
+      0.52    png handles req_comp=3,4 directly; minor cleanup; jpeg comments
+      0.51    obey req_comp requests, 1-component jpegs return as 1-component,
+              on 'test' only check type, not whether we support this variant
+      0.50  (2006-11-19)
+              first released version
+*/
+
+
+/*
+------------------------------------------------------------------------------
+This software is available under 2 licenses -- choose whichever you prefer.
+------------------------------------------------------------------------------
+ALTERNATIVE A - MIT License
+Copyright (c) 2017 Sean Barrett
+Permission is hereby granted, free of charge, to any person obtaining a copy of
+this software and associated documentation files (the "Software"), to deal in
+the Software without restriction, including without limitation the rights to
+use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies
+of the Software, and to permit persons to whom the Software is furnished to do
+so, subject to the following conditions:
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
+------------------------------------------------------------------------------
+ALTERNATIVE B - Public Domain (www.unlicense.org)
+This is free and unencumbered software released into the public domain.
+Anyone is free to copy, modify, publish, use, compile, sell, or distribute this
+software, either in source code form or as a compiled binary, for any purpose,
+commercial or non-commercial, and by any means.
+In jurisdictions that recognize copyright laws, the author or authors of this
+software dedicate any and all copyright interest in the software to the public
+domain. We make this dedication for the benefit of the public at large and to
+the detriment of our heirs and successors. We intend this dedication to be an
+overt act of relinquishment in perpetuity of all present and future rights to
+this software under copyright law.
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
+WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+------------------------------------------------------------------------------
+*/

ggml/examples/stb_image_write.h

@@ -0,0 +1,1724 @@
+/* stb_image_write - v1.16 - public domain - http://nothings.org/stb
+   writes out PNG/BMP/TGA/JPEG/HDR images to C stdio - Sean Barrett 2010-2015
+                                     no warranty implied; use at your own risk
+
+   Before #including,
+
+       #define STB_IMAGE_WRITE_IMPLEMENTATION
+
+   in the file that you want to have the implementation.
+
+   Will probably not work correctly with strict-aliasing optimizations.
+
+ABOUT:
+
+   This header file is a library for writing images to C stdio or a callback.
+
+   The PNG output is not optimal; it is 20-50% larger than the file
+   written by a decent optimizing implementation; though providing a custom
+   zlib compress function (see STBIW_ZLIB_COMPRESS) can mitigate that.
+   This library is designed for source code compactness and simplicity,
+   not optimal image file size or run-time performance.
+
+BUILDING:
+
+   You can #define STBIW_ASSERT(x) before the #include to avoid using assert.h.
+   You can #define STBIW_MALLOC(), STBIW_REALLOC(), and STBIW_FREE() to replace
+   malloc,realloc,free.
+   You can #define STBIW_MEMMOVE() to replace memmove()
+   You can #define STBIW_ZLIB_COMPRESS to use a custom zlib-style compress function
+   for PNG compression (instead of the builtin one), it must have the following signature:
+   unsigned char * my_compress(unsigned char *data, int data_len, int *out_len, int quality);
+   The returned data will be freed with STBIW_FREE() (free() by default),
+   so it must be heap allocated with STBIW_MALLOC() (malloc() by default),
+
+UNICODE:
+
+   If compiling for Windows and you wish to use Unicode filenames, compile
+   with
+       #define STBIW_WINDOWS_UTF8
+   and pass utf8-encoded filenames. Call stbiw_convert_wchar_to_utf8 to convert
+   Windows wchar_t filenames to utf8.
+
+USAGE:
+
+   There are five functions, one for each image file format:
+
+     int stbi_write_png(char const *filename, int w, int h, int comp, const void *data, int stride_in_bytes);
+     int stbi_write_bmp(char const *filename, int w, int h, int comp, const void *data);
+     int stbi_write_tga(char const *filename, int w, int h, int comp, const void *data);
+     int stbi_write_jpg(char const *filename, int w, int h, int comp, const void *data, int quality);
+     int stbi_write_hdr(char const *filename, int w, int h, int comp, const float *data);
+
+     void stbi_flip_vertically_on_write(int flag); // flag is non-zero to flip data vertically
+
+   There are also five equivalent functions that use an arbitrary write function. You are
+   expected to open/close your file-equivalent before and after calling these:
+
+     int stbi_write_png_to_func(stbi_write_func *func, void *context, int w, int h, int comp, const void  *data, int stride_in_bytes);
+     int stbi_write_bmp_to_func(stbi_write_func *func, void *context, int w, int h, int comp, const void  *data);
+     int stbi_write_tga_to_func(stbi_write_func *func, void *context, int w, int h, int comp, const void  *data);
+     int stbi_write_hdr_to_func(stbi_write_func *func, void *context, int w, int h, int comp, const float *data);
+     int stbi_write_jpg_to_func(stbi_write_func *func, void *context, int x, int y, int comp, const void *data, int quality);
+
+   where the callback is:
+      void stbi_write_func(void *context, void *data, int size);
+
+   You can configure it with these global variables:
+      int stbi_write_tga_with_rle;             // defaults to true; set to 0 to disable RLE
+      int stbi_write_png_compression_level;    // defaults to 8; set to higher for more compression
+      int stbi_write_force_png_filter;         // defaults to -1; set to 0..5 to force a filter mode
+
+
+   You can define STBI_WRITE_NO_STDIO to disable the file variant of these
+   functions, so the library will not use stdio.h at all. However, this will
+   also disable HDR writing, because it requires stdio for formatted output.
+
+   Each function returns 0 on failure and non-0 on success.
+
+   The functions create an image file defined by the parameters. The image
+   is a rectangle of pixels stored from left-to-right, top-to-bottom.
+   Each pixel contains 'comp' channels of data stored interleaved with 8-bits
+   per channel, in the following order: 1=Y, 2=YA, 3=RGB, 4=RGBA. (Y is
+   monochrome color.) The rectangle is 'w' pixels wide and 'h' pixels tall.
+   The *data pointer points to the first byte of the top-left-most pixel.
+   For PNG, "stride_in_bytes" is the distance in bytes from the first byte of
+   a row of pixels to the first byte of the next row of pixels.
+
+   PNG creates output files with the same number of components as the input.
+   The BMP format expands Y to RGB in the file format and does not
+   output alpha.
+
+   PNG supports writing rectangles of data even when the bytes storing rows of
+   data are not consecutive in memory (e.g. sub-rectangles of a larger image),
+   by supplying the stride between the beginning of adjacent rows. The other
+   formats do not. (Thus you cannot write a native-format BMP through the BMP
+   writer, both because it is in BGR order and because it may have padding
+   at the end of the line.)
+
+   PNG allows you to set the deflate compression level by setting the global
+   variable 'stbi_write_png_compression_level' (it defaults to 8).
+
+   HDR expects linear float data. Since the format is always 32-bit rgb(e)
+   data, alpha (if provided) is discarded, and for monochrome data it is
+   replicated across all three channels.
+
+   TGA supports RLE or non-RLE compressed data. To use non-RLE-compressed
+   data, set the global variable 'stbi_write_tga_with_rle' to 0.
+
+   JPEG does ignore alpha channels in input data; quality is between 1 and 100.
+   Higher quality looks better but results in a bigger image.
+   JPEG baseline (no JPEG progressive).
+
+CREDITS:
+
+
+   Sean Barrett           -    PNG/BMP/TGA
+   Baldur Karlsson        -    HDR
+   Jean-Sebastien Guay    -    TGA monochrome
+   Tim Kelsey             -    misc enhancements
+   Alan Hickman           -    TGA RLE
+   Emmanuel Julien        -    initial file IO callback implementation
+   Jon Olick              -    original jo_jpeg.cpp code
+   Daniel Gibson          -    integrate JPEG, allow external zlib
+   Aarni Koskela          -    allow choosing PNG filter
+
+   bugfixes:
+      github:Chribba
+      Guillaume Chereau
+      github:jry2
+      github:romigrou
+      Sergio Gonzalez
+      Jonas Karlsson
+      Filip Wasil
+      Thatcher Ulrich
+      github:poppolopoppo
+      Patrick Boettcher
+      github:xeekworx
+      Cap Petschulat
+      Simon Rodriguez
+      Ivan Tikhonov
+      github:ignotion
+      Adam Schackart
+      Andrew Kensler
+
+LICENSE
+
+  See end of file for license information.
+
+*/
+
+#ifndef INCLUDE_STB_IMAGE_WRITE_H
+#define INCLUDE_STB_IMAGE_WRITE_H
+
+#include <stdlib.h>
+
+// if STB_IMAGE_WRITE_STATIC causes problems, try defining STBIWDEF to 'inline' or 'static inline'
+#ifndef STBIWDEF
+#ifdef STB_IMAGE_WRITE_STATIC
+#define STBIWDEF  static
+#else
+#ifdef __cplusplus
+#define STBIWDEF  extern "C"
+#else
+#define STBIWDEF  extern
+#endif
+#endif
+#endif
+
+#ifndef STB_IMAGE_WRITE_STATIC  // C++ forbids static forward declarations
+STBIWDEF int stbi_write_tga_with_rle;
+STBIWDEF int stbi_write_png_compression_level;
+STBIWDEF int stbi_write_force_png_filter;
+#endif
+
+#ifndef STBI_WRITE_NO_STDIO
+STBIWDEF int stbi_write_png(char const *filename, int w, int h, int comp, const void  *data, int stride_in_bytes);
+STBIWDEF int stbi_write_bmp(char const *filename, int w, int h, int comp, const void  *data);
+STBIWDEF int stbi_write_tga(char const *filename, int w, int h, int comp, const void  *data);
+STBIWDEF int stbi_write_hdr(char const *filename, int w, int h, int comp, const float *data);
+STBIWDEF int stbi_write_jpg(char const *filename, int x, int y, int comp, const void  *data, int quality);
+
+#ifdef STBIW_WINDOWS_UTF8
+STBIWDEF int stbiw_convert_wchar_to_utf8(char *buffer, size_t bufferlen, const wchar_t* input);
+#endif
+#endif
+
+typedef void stbi_write_func(void *context, void *data, int size);
+
+STBIWDEF int stbi_write_png_to_func(stbi_write_func *func, void *context, int w, int h, int comp, const void  *data, int stride_in_bytes);
+STBIWDEF int stbi_write_bmp_to_func(stbi_write_func *func, void *context, int w, int h, int comp, const void  *data);
+STBIWDEF int stbi_write_tga_to_func(stbi_write_func *func, void *context, int w, int h, int comp, const void  *data);
+STBIWDEF int stbi_write_hdr_to_func(stbi_write_func *func, void *context, int w, int h, int comp, const float *data);
+STBIWDEF int stbi_write_jpg_to_func(stbi_write_func *func, void *context, int x, int y, int comp, const void  *data, int quality);
+
+STBIWDEF void stbi_flip_vertically_on_write(int flip_boolean);
+
+#endif//INCLUDE_STB_IMAGE_WRITE_H
+
+#ifdef STB_IMAGE_WRITE_IMPLEMENTATION
+
+#ifdef _WIN32
+   #ifndef _CRT_SECURE_NO_WARNINGS
+   #define _CRT_SECURE_NO_WARNINGS
+   #endif
+   #ifndef _CRT_NONSTDC_NO_DEPRECATE
+   #define _CRT_NONSTDC_NO_DEPRECATE
+   #endif
+#endif
+
+#ifndef STBI_WRITE_NO_STDIO
+#include <stdio.h>
+#endif // STBI_WRITE_NO_STDIO
+
+#include <stdarg.h>
+#include <stdlib.h>
+#include <string.h>
+#include <math.h>
+
+#if defined(STBIW_MALLOC) && defined(STBIW_FREE) && (defined(STBIW_REALLOC) || defined(STBIW_REALLOC_SIZED))
+// ok
+#elif !defined(STBIW_MALLOC) && !defined(STBIW_FREE) && !defined(STBIW_REALLOC) && !defined(STBIW_REALLOC_SIZED)
+// ok
+#else
+#error "Must define all or none of STBIW_MALLOC, STBIW_FREE, and STBIW_REALLOC (or STBIW_REALLOC_SIZED)."
+#endif
+
+#ifndef STBIW_MALLOC
+#define STBIW_MALLOC(sz)        malloc(sz)
+#define STBIW_REALLOC(p,newsz)  realloc(p,newsz)
+#define STBIW_FREE(p)           free(p)
+#endif
+
+#ifndef STBIW_REALLOC_SIZED
+#define STBIW_REALLOC_SIZED(p,oldsz,newsz) STBIW_REALLOC(p,newsz)
+#endif
+
+
+#ifndef STBIW_MEMMOVE
+#define STBIW_MEMMOVE(a,b,sz) memmove(a,b,sz)
+#endif
+
+
+#ifndef STBIW_ASSERT
+#include <assert.h>
+#define STBIW_ASSERT(x) assert(x)
+#endif
+
+#define STBIW_UCHAR(x) (unsigned char) ((x) & 0xff)
+
+#ifdef STB_IMAGE_WRITE_STATIC
+static int stbi_write_png_compression_level = 8;
+static int stbi_write_tga_with_rle = 1;
+static int stbi_write_force_png_filter = -1;
+#else
+int stbi_write_png_compression_level = 8;
+int stbi_write_tga_with_rle = 1;
+int stbi_write_force_png_filter = -1;
+#endif
+
+static int stbi__flip_vertically_on_write = 0;
+
+STBIWDEF void stbi_flip_vertically_on_write(int flag)
+{
+   stbi__flip_vertically_on_write = flag;
+}
+
+typedef struct
+{
+   stbi_write_func *func;
+   void *context;
+   unsigned char buffer[64];
+   int buf_used;
+} stbi__write_context;
+
+// initialize a callback-based context
+static void stbi__start_write_callbacks(stbi__write_context *s, stbi_write_func *c, void *context)
+{
+   s->func    = c;
+   s->context = context;
+}
+
+#ifndef STBI_WRITE_NO_STDIO
+
+static void stbi__stdio_write(void *context, void *data, int size)
+{
+   fwrite(data,1,size,(FILE*) context);
+}
+
+#if defined(_WIN32) && defined(STBIW_WINDOWS_UTF8)
+#ifdef __cplusplus
+#define STBIW_EXTERN extern "C"
+#else
+#define STBIW_EXTERN extern
+#endif
+STBIW_EXTERN __declspec(dllimport) int __stdcall MultiByteToWideChar(unsigned int cp, unsigned long flags, const char *str, int cbmb, wchar_t *widestr, int cchwide);
+STBIW_EXTERN __declspec(dllimport) int __stdcall WideCharToMultiByte(unsigned int cp, unsigned long flags, const wchar_t *widestr, int cchwide, char *str, int cbmb, const char *defchar, int *used_default);
+
+STBIWDEF int stbiw_convert_wchar_to_utf8(char *buffer, size_t bufferlen, const wchar_t* input)
+{
+   return WideCharToMultiByte(65001 /* UTF8 */, 0, input, -1, buffer, (int) bufferlen, NULL, NULL);
+}
+#endif
+
+static FILE *stbiw__fopen(char const *filename, char const *mode)
+{
+   FILE *f;
+#if defined(_WIN32) && defined(STBIW_WINDOWS_UTF8)
+   wchar_t wMode[64];
+   wchar_t wFilename[1024];
+   if (0 == MultiByteToWideChar(65001 /* UTF8 */, 0, filename, -1, wFilename, sizeof(wFilename)/sizeof(*wFilename)))
+      return 0;
+
+   if (0 == MultiByteToWideChar(65001 /* UTF8 */, 0, mode, -1, wMode, sizeof(wMode)/sizeof(*wMode)))
+      return 0;
+
+#if defined(_MSC_VER) && _MSC_VER >= 1400
+   if (0 != _wfopen_s(&f, wFilename, wMode))
+      f = 0;
+#else
+   f = _wfopen(wFilename, wMode);
+#endif
+
+#elif defined(_MSC_VER) && _MSC_VER >= 1400
+   if (0 != fopen_s(&f, filename, mode))
+      f=0;
+#else
+   f = fopen(filename, mode);
+#endif
+   return f;
+}
+
+static int stbi__start_write_file(stbi__write_context *s, const char *filename)
+{
+   FILE *f = stbiw__fopen(filename, "wb");
+   stbi__start_write_callbacks(s, stbi__stdio_write, (void *) f);
+   return f != NULL;
+}
+
+static void stbi__end_write_file(stbi__write_context *s)
+{
+   fclose((FILE *)s->context);
+}
+
+#endif // !STBI_WRITE_NO_STDIO
+
+typedef unsigned int stbiw_uint32;
+typedef int stb_image_write_test[sizeof(stbiw_uint32)==4 ? 1 : -1];
+
+static void stbiw__writefv(stbi__write_context *s, const char *fmt, va_list v)
+{
+   while (*fmt) {
+      switch (*fmt++) {
+         case ' ': break;
+         case '1': { unsigned char x = STBIW_UCHAR(va_arg(v, int));
+                     s->func(s->context,&x,1);
+                     break; }
+         case '2': { int x = va_arg(v,int);
+                     unsigned char b[2];
+                     b[0] = STBIW_UCHAR(x);
+                     b[1] = STBIW_UCHAR(x>>8);
+                     s->func(s->context,b,2);
+                     break; }
+         case '4': { stbiw_uint32 x = va_arg(v,int);
+                     unsigned char b[4];
+                     b[0]=STBIW_UCHAR(x);
+                     b[1]=STBIW_UCHAR(x>>8);
+                     b[2]=STBIW_UCHAR(x>>16);
+                     b[3]=STBIW_UCHAR(x>>24);
+                     s->func(s->context,b,4);
+                     break; }
+         default:
+            STBIW_ASSERT(0);
+            return;
+      }
+   }
+}
+
+static void stbiw__writef(stbi__write_context *s, const char *fmt, ...)
+{
+   va_list v;
+   va_start(v, fmt);
+   stbiw__writefv(s, fmt, v);
+   va_end(v);
+}
+
+static void stbiw__write_flush(stbi__write_context *s)
+{
+   if (s->buf_used) {
+      s->func(s->context, &s->buffer, s->buf_used);
+      s->buf_used = 0;
+   }
+}
+
+static void stbiw__putc(stbi__write_context *s, unsigned char c)
+{
+   s->func(s->context, &c, 1);
+}
+
+static void stbiw__write1(stbi__write_context *s, unsigned char a)
+{
+   if ((size_t)s->buf_used + 1 > sizeof(s->buffer))
+      stbiw__write_flush(s);
+   s->buffer[s->buf_used++] = a;
+}
+
+static void stbiw__write3(stbi__write_context *s, unsigned char a, unsigned char b, unsigned char c)
+{
+   int n;
+   if ((size_t)s->buf_used + 3 > sizeof(s->buffer))
+      stbiw__write_flush(s);
+   n = s->buf_used;
+   s->buf_used = n+3;
+   s->buffer[n+0] = a;
+   s->buffer[n+1] = b;
+   s->buffer[n+2] = c;
+}
+
+static void stbiw__write_pixel(stbi__write_context *s, int rgb_dir, int comp, int write_alpha, int expand_mono, unsigned char *d)
+{
+   unsigned char bg[3] = { 255, 0, 255}, px[3];
+   int k;
+
+   if (write_alpha < 0)
+      stbiw__write1(s, d[comp - 1]);
+
+   switch (comp) {
+      case 2: // 2 pixels = mono + alpha, alpha is written separately, so same as 1-channel case
+      case 1:
+         if (expand_mono)
+            stbiw__write3(s, d[0], d[0], d[0]); // monochrome bmp
+         else
+            stbiw__write1(s, d[0]);  // monochrome TGA
+         break;
+      case 4:
+         if (!write_alpha) {
+            // composite against pink background
+            for (k = 0; k < 3; ++k)
+               px[k] = bg[k] + ((d[k] - bg[k]) * d[3]) / 255;
+            stbiw__write3(s, px[1 - rgb_dir], px[1], px[1 + rgb_dir]);
+            break;
+         }
+         /* FALLTHROUGH */
+      case 3:
+         stbiw__write3(s, d[1 - rgb_dir], d[1], d[1 + rgb_dir]);
+         break;
+   }
+   if (write_alpha > 0)
+      stbiw__write1(s, d[comp - 1]);
+}
+
+static void stbiw__write_pixels(stbi__write_context *s, int rgb_dir, int vdir, int x, int y, int comp, void *data, int write_alpha, int scanline_pad, int expand_mono)
+{
+   stbiw_uint32 zero = 0;
+   int i,j, j_end;
+
+   if (y <= 0)
+      return;
+
+   if (stbi__flip_vertically_on_write)
+      vdir *= -1;
+
+   if (vdir < 0) {
+      j_end = -1; j = y-1;
+   } else {
+      j_end =  y; j = 0;
+   }
+
+   for (; j != j_end; j += vdir) {
+      for (i=0; i < x; ++i) {
+         unsigned char *d = (unsigned char *) data + (j*x+i)*comp;
+         stbiw__write_pixel(s, rgb_dir, comp, write_alpha, expand_mono, d);
+      }
+      stbiw__write_flush(s);
+      s->func(s->context, &zero, scanline_pad);
+   }
+}
+
+static int stbiw__outfile(stbi__write_context *s, int rgb_dir, int vdir, int x, int y, int comp, int expand_mono, void *data, int alpha, int pad, const char *fmt, ...)
+{
+   if (y < 0 || x < 0) {
+      return 0;
+   } else {
+      va_list v;
+      va_start(v, fmt);
+      stbiw__writefv(s, fmt, v);
+      va_end(v);
+      stbiw__write_pixels(s,rgb_dir,vdir,x,y,comp,data,alpha,pad, expand_mono);
+      return 1;
+   }
+}
+
+static int stbi_write_bmp_core(stbi__write_context *s, int x, int y, int comp, const void *data)
+{
+   if (comp != 4) {
+      // write RGB bitmap
+      int pad = (-x*3) & 3;
+      return stbiw__outfile(s,-1,-1,x,y,comp,1,(void *) data,0,pad,
+              "11 4 22 4" "4 44 22 444444",
+              'B', 'M', 14+40+(x*3+pad)*y, 0,0, 14+40,  // file header
+               40, x,y, 1,24, 0,0,0,0,0,0);             // bitmap header
+   } else {
+      // RGBA bitmaps need a v4 header
+      // use BI_BITFIELDS mode with 32bpp and alpha mask
+      // (straight BI_RGB with alpha mask doesn't work in most readers)
+      return stbiw__outfile(s,-1,-1,x,y,comp,1,(void *)data,1,0,
+         "11 4 22 4" "4 44 22 444444 4444 4 444 444 444 444",
+         'B', 'M', 14+108+x*y*4, 0, 0, 14+108, // file header
+         108, x,y, 1,32, 3,0,0,0,0,0, 0xff0000,0xff00,0xff,0xff000000u, 0, 0,0,0, 0,0,0, 0,0,0, 0,0,0); // bitmap V4 header
+   }
+}
+
+STBIWDEF int stbi_write_bmp_to_func(stbi_write_func *func, void *context, int x, int y, int comp, const void *data)
+{
+   stbi__write_context s = { 0 };
+   stbi__start_write_callbacks(&s, func, context);
+   return stbi_write_bmp_core(&s, x, y, comp, data);
+}
+
+#ifndef STBI_WRITE_NO_STDIO
+STBIWDEF int stbi_write_bmp(char const *filename, int x, int y, int comp, const void *data)
+{
+   stbi__write_context s = { 0 };
+   if (stbi__start_write_file(&s,filename)) {
+      int r = stbi_write_bmp_core(&s, x, y, comp, data);
+      stbi__end_write_file(&s);
+      return r;
+   } else
+      return 0;
+}
+#endif //!STBI_WRITE_NO_STDIO
+
+static int stbi_write_tga_core(stbi__write_context *s, int x, int y, int comp, void *data)
+{
+   int has_alpha = (comp == 2 || comp == 4);
+   int colorbytes = has_alpha ? comp-1 : comp;
+   int format = colorbytes < 2 ? 3 : 2; // 3 color channels (RGB/RGBA) = 2, 1 color channel (Y/YA) = 3
+
+   if (y < 0 || x < 0)
+      return 0;
+
+   if (!stbi_write_tga_with_rle) {
+      return stbiw__outfile(s, -1, -1, x, y, comp, 0, (void *) data, has_alpha, 0,
+         "111 221 2222 11", 0, 0, format, 0, 0, 0, 0, 0, x, y, (colorbytes + has_alpha) * 8, has_alpha * 8);
+   } else {
+      int i,j,k;
+      int jend, jdir;
+
+      stbiw__writef(s, "111 221 2222 11", 0,0,format+8, 0,0,0, 0,0,x,y, (colorbytes + has_alpha) * 8, has_alpha * 8);
+
+      if (stbi__flip_vertically_on_write) {
+         j = 0;
+         jend = y;
+         jdir = 1;
+      } else {
+         j = y-1;
+         jend = -1;
+         jdir = -1;
+      }
+      for (; j != jend; j += jdir) {
+         unsigned char *row = (unsigned char *) data + j * x * comp;
+         int len;
+
+         for (i = 0; i < x; i += len) {
+            unsigned char *begin = row + i * comp;
+            int diff = 1;
+            len = 1;
+
+            if (i < x - 1) {
+               ++len;
+               diff = memcmp(begin, row + (i + 1) * comp, comp);
+               if (diff) {
+                  const unsigned char *prev = begin;
+                  for (k = i + 2; k < x && len < 128; ++k) {
+                     if (memcmp(prev, row + k * comp, comp)) {
+                        prev += comp;
+                        ++len;
+                     } else {
+                        --len;
+                        break;
+                     }
+                  }
+               } else {
+                  for (k = i + 2; k < x && len < 128; ++k) {
+                     if (!memcmp(begin, row + k * comp, comp)) {
+                        ++len;
+                     } else {
+                        break;
+                     }
+                  }
+               }
+            }
+
+            if (diff) {
+               unsigned char header = STBIW_UCHAR(len - 1);
+               stbiw__write1(s, header);
+               for (k = 0; k < len; ++k) {
+                  stbiw__write_pixel(s, -1, comp, has_alpha, 0, begin + k * comp);
+               }
+            } else {
+               unsigned char header = STBIW_UCHAR(len - 129);
+               stbiw__write1(s, header);
+               stbiw__write_pixel(s, -1, comp, has_alpha, 0, begin);
+            }
+         }
+      }
+      stbiw__write_flush(s);
+   }
+   return 1;
+}
+
+STBIWDEF int stbi_write_tga_to_func(stbi_write_func *func, void *context, int x, int y, int comp, const void *data)
+{
+   stbi__write_context s = { 0 };
+   stbi__start_write_callbacks(&s, func, context);
+   return stbi_write_tga_core(&s, x, y, comp, (void *) data);
+}
+
+#ifndef STBI_WRITE_NO_STDIO
+STBIWDEF int stbi_write_tga(char const *filename, int x, int y, int comp, const void *data)
+{
+   stbi__write_context s = { 0 };
+   if (stbi__start_write_file(&s,filename)) {
+      int r = stbi_write_tga_core(&s, x, y, comp, (void *) data);
+      stbi__end_write_file(&s);
+      return r;
+   } else
+      return 0;
+}
+#endif
+
+// *************************************************************************************************
+// Radiance RGBE HDR writer
+// by Baldur Karlsson
+
+#define stbiw__max(a, b)  ((a) > (b) ? (a) : (b))
+
+#ifndef STBI_WRITE_NO_STDIO
+
+static void stbiw__linear_to_rgbe(unsigned char *rgbe, float *linear)
+{
+   int exponent;
+   float maxcomp = stbiw__max(linear[0], stbiw__max(linear[1], linear[2]));
+
+   if (maxcomp < 1e-32f) {
+      rgbe[0] = rgbe[1] = rgbe[2] = rgbe[3] = 0;
+   } else {
+      float normalize = (float) frexp(maxcomp, &exponent) * 256.0f/maxcomp;
+
+      rgbe[0] = (unsigned char)(linear[0] * normalize);
+      rgbe[1] = (unsigned char)(linear[1] * normalize);
+      rgbe[2] = (unsigned char)(linear[2] * normalize);
+      rgbe[3] = (unsigned char)(exponent + 128);
+   }
+}
+
+static void stbiw__write_run_data(stbi__write_context *s, int length, unsigned char databyte)
+{
+   unsigned char lengthbyte = STBIW_UCHAR(length+128);
+   STBIW_ASSERT(length+128 <= 255);
+   s->func(s->context, &lengthbyte, 1);
+   s->func(s->context, &databyte, 1);
+}
+
+static void stbiw__write_dump_data(stbi__write_context *s, int length, unsigned char *data)
+{
+   unsigned char lengthbyte = STBIW_UCHAR(length);
+   STBIW_ASSERT(length <= 128); // inconsistent with spec but consistent with official code
+   s->func(s->context, &lengthbyte, 1);
+   s->func(s->context, data, length);
+}
+
+static void stbiw__write_hdr_scanline(stbi__write_context *s, int width, int ncomp, unsigned char *scratch, float *scanline)
+{
+   unsigned char scanlineheader[4] = { 2, 2, 0, 0 };
+   unsigned char rgbe[4];
+   float linear[3];
+   int x;
+
+   scanlineheader[2] = (width&0xff00)>>8;
+   scanlineheader[3] = (width&0x00ff);
+
+   /* skip RLE for images too small or large */
+   if (width < 8 || width >= 32768) {
+      for (x=0; x < width; x++) {
+         switch (ncomp) {
+            case 4: /* fallthrough */
+            case 3: linear[2] = scanline[x*ncomp + 2];
+                    linear[1] = scanline[x*ncomp + 1];
+                    linear[0] = scanline[x*ncomp + 0];
+                    break;
+            default:
+                    linear[0] = linear[1] = linear[2] = scanline[x*ncomp + 0];
+                    break;
+         }
+         stbiw__linear_to_rgbe(rgbe, linear);
+         s->func(s->context, rgbe, 4);
+      }
+   } else {
+      int c,r;
+      /* encode into scratch buffer */
+      for (x=0; x < width; x++) {
+         switch(ncomp) {
+            case 4: /* fallthrough */
+            case 3: linear[2] = scanline[x*ncomp + 2];
+                    linear[1] = scanline[x*ncomp + 1];
+                    linear[0] = scanline[x*ncomp + 0];
+                    break;
+            default:
+                    linear[0] = linear[1] = linear[2] = scanline[x*ncomp + 0];
+                    break;
+         }
+         stbiw__linear_to_rgbe(rgbe, linear);
+         scratch[x + width*0] = rgbe[0];
+         scratch[x + width*1] = rgbe[1];
+         scratch[x + width*2] = rgbe[2];
+         scratch[x + width*3] = rgbe[3];
+      }
+
+      s->func(s->context, scanlineheader, 4);
+
+      /* RLE each component separately */
+      for (c=0; c < 4; c++) {
+         unsigned char *comp = &scratch[width*c];
+
+         x = 0;
+         while (x < width) {
+            // find first run
+            r = x;
+            while (r+2 < width) {
+               if (comp[r] == comp[r+1] && comp[r] == comp[r+2])
+                  break;
+               ++r;
+            }
+            if (r+2 >= width)
+               r = width;
+            // dump up to first run
+            while (x < r) {
+               int len = r-x;
+               if (len > 128) len = 128;
+               stbiw__write_dump_data(s, len, &comp[x]);
+               x += len;
+            }
+            // if there's a run, output it
+            if (r+2 < width) { // same test as what we break out of in search loop, so only true if we break'd
+               // find next byte after run
+               while (r < width && comp[r] == comp[x])
+                  ++r;
+               // output run up to r
+               while (x < r) {
+                  int len = r-x;
+                  if (len > 127) len = 127;
+                  stbiw__write_run_data(s, len, comp[x]);
+                  x += len;
+               }
+            }
+         }
+      }
+   }
+}
+
+static int stbi_write_hdr_core(stbi__write_context *s, int x, int y, int comp, float *data)
+{
+   if (y <= 0 || x <= 0 || data == NULL)
+      return 0;
+   else {
+      // Each component is stored separately. Allocate scratch space for full output scanline.
+      unsigned char *scratch = (unsigned char *) STBIW_MALLOC(x*4);
+      int i, len;
+      char buffer[128];
+      char header[] = "#?RADIANCE\n# Written by stb_image_write.h\nFORMAT=32-bit_rle_rgbe\n";
+      s->func(s->context, header, sizeof(header)-1);
+
+#ifdef __STDC_LIB_EXT1__
+      len = sprintf_s(buffer, sizeof(buffer), "EXPOSURE=          1.0000000000000\n\n-Y %d +X %d\n", y, x);
+#else
+      len = sprintf(buffer, "EXPOSURE=          1.0000000000000\n\n-Y %d +X %d\n", y, x);
+#endif
+      s->func(s->context, buffer, len);
+
+      for(i=0; i < y; i++)
+         stbiw__write_hdr_scanline(s, x, comp, scratch, data + comp*x*(stbi__flip_vertically_on_write ? y-1-i : i));
+      STBIW_FREE(scratch);
+      return 1;
+   }
+}
+
+STBIWDEF int stbi_write_hdr_to_func(stbi_write_func *func, void *context, int x, int y, int comp, const float *data)
+{
+   stbi__write_context s = { 0 };
+   stbi__start_write_callbacks(&s, func, context);
+   return stbi_write_hdr_core(&s, x, y, comp, (float *) data);
+}
+
+STBIWDEF int stbi_write_hdr(char const *filename, int x, int y, int comp, const float *data)
+{
+   stbi__write_context s = { 0 };
+   if (stbi__start_write_file(&s,filename)) {
+      int r = stbi_write_hdr_core(&s, x, y, comp, (float *) data);
+      stbi__end_write_file(&s);
+      return r;
+   } else
+      return 0;
+}
+#endif // STBI_WRITE_NO_STDIO
+
+
+//////////////////////////////////////////////////////////////////////////////
+//
+// PNG writer
+//
+
+#ifndef STBIW_ZLIB_COMPRESS
+// stretchy buffer; stbiw__sbpush() == vector<>::push_back() -- stbiw__sbcount() == vector<>::size()
+#define stbiw__sbraw(a) ((int *) (void *) (a) - 2)
+#define stbiw__sbm(a)   stbiw__sbraw(a)[0]
+#define stbiw__sbn(a)   stbiw__sbraw(a)[1]
+
+#define stbiw__sbneedgrow(a,n)  ((a)==0 || stbiw__sbn(a)+n >= stbiw__sbm(a))
+#define stbiw__sbmaybegrow(a,n) (stbiw__sbneedgrow(a,(n)) ? stbiw__sbgrow(a,n) : 0)
+#define stbiw__sbgrow(a,n)  stbiw__sbgrowf((void **) &(a), (n), sizeof(*(a)))
+
+#define stbiw__sbpush(a, v)      (stbiw__sbmaybegrow(a,1), (a)[stbiw__sbn(a)++] = (v))
+#define stbiw__sbcount(a)        ((a) ? stbiw__sbn(a) : 0)
+#define stbiw__sbfree(a)         ((a) ? STBIW_FREE(stbiw__sbraw(a)),0 : 0)
+
+static void *stbiw__sbgrowf(void **arr, int increment, int itemsize)
+{
+   int m = *arr ? 2*stbiw__sbm(*arr)+increment : increment+1;
+   void *p = STBIW_REALLOC_SIZED(*arr ? stbiw__sbraw(*arr) : 0, *arr ? (stbiw__sbm(*arr)*itemsize + sizeof(int)*2) : 0, itemsize * m + sizeof(int)*2);
+   STBIW_ASSERT(p);
+   if (p) {
+      if (!*arr) ((int *) p)[1] = 0;
+      *arr = (void *) ((int *) p + 2);
+      stbiw__sbm(*arr) = m;
+   }
+   return *arr;
+}
+
+static unsigned char *stbiw__zlib_flushf(unsigned char *data, unsigned int *bitbuffer, int *bitcount)
+{
+   while (*bitcount >= 8) {
+      stbiw__sbpush(data, STBIW_UCHAR(*bitbuffer));
+      *bitbuffer >>= 8;
+      *bitcount -= 8;
+   }
+   return data;
+}
+
+static int stbiw__zlib_bitrev(int code, int codebits)
+{
+   int res=0;
+   while (codebits--) {
+      res = (res << 1) | (code & 1);
+      code >>= 1;
+   }
+   return res;
+}
+
+static unsigned int stbiw__zlib_countm(unsigned char *a, unsigned char *b, int limit)
+{
+   int i;
+   for (i=0; i < limit && i < 258; ++i)
+      if (a[i] != b[i]) break;
+   return i;
+}
+
+static unsigned int stbiw__zhash(unsigned char *data)
+{
+   stbiw_uint32 hash = data[0] + (data[1] << 8) + (data[2] << 16);
+   hash ^= hash << 3;
+   hash += hash >> 5;
+   hash ^= hash << 4;
+   hash += hash >> 17;
+   hash ^= hash << 25;
+   hash += hash >> 6;
+   return hash;
+}
+
+#define stbiw__zlib_flush() (out = stbiw__zlib_flushf(out, &bitbuf, &bitcount))
+#define stbiw__zlib_add(code,codebits) \
+      (bitbuf |= (code) << bitcount, bitcount += (codebits), stbiw__zlib_flush())
+#define stbiw__zlib_huffa(b,c)  stbiw__zlib_add(stbiw__zlib_bitrev(b,c),c)
+// default huffman tables
+#define stbiw__zlib_huff1(n)  stbiw__zlib_huffa(0x30 + (n), 8)
+#define stbiw__zlib_huff2(n)  stbiw__zlib_huffa(0x190 + (n)-144, 9)
+#define stbiw__zlib_huff3(n)  stbiw__zlib_huffa(0 + (n)-256,7)
+#define stbiw__zlib_huff4(n)  stbiw__zlib_huffa(0xc0 + (n)-280,8)
+#define stbiw__zlib_huff(n)  ((n) <= 143 ? stbiw__zlib_huff1(n) : (n) <= 255 ? stbiw__zlib_huff2(n) : (n) <= 279 ? stbiw__zlib_huff3(n) : stbiw__zlib_huff4(n))
+#define stbiw__zlib_huffb(n) ((n) <= 143 ? stbiw__zlib_huff1(n) : stbiw__zlib_huff2(n))
+
+#define stbiw__ZHASH   16384
+
+#endif // STBIW_ZLIB_COMPRESS
+
+STBIWDEF unsigned char * stbi_zlib_compress(unsigned char *data, int data_len, int *out_len, int quality)
+{
+#ifdef STBIW_ZLIB_COMPRESS
+   // user provided a zlib compress implementation, use that
+   return STBIW_ZLIB_COMPRESS(data, data_len, out_len, quality);
+#else // use builtin
+   static unsigned short lengthc[] = { 3,4,5,6,7,8,9,10,11,13,15,17,19,23,27,31,35,43,51,59,67,83,99,115,131,163,195,227,258, 259 };
+   static unsigned char  lengtheb[]= { 0,0,0,0,0,0,0, 0, 1, 1, 1, 1, 2, 2, 2, 2, 3, 3, 3, 3, 4, 4, 4,  4,  5,  5,  5,  5,  0 };
+   static unsigned short distc[]   = { 1,2,3,4,5,7,9,13,17,25,33,49,65,97,129,193,257,385,513,769,1025,1537,2049,3073,4097,6145,8193,12289,16385,24577, 32768 };
+   static unsigned char  disteb[]  = { 0,0,0,0,1,1,2,2,3,3,4,4,5,5,6,6,7,7,8,8,9,9,10,10,11,11,12,12,13,13 };
+   unsigned int bitbuf=0;
+   int i,j, bitcount=0;
+   unsigned char *out = NULL;
+   unsigned char ***hash_table = (unsigned char***) STBIW_MALLOC(stbiw__ZHASH * sizeof(unsigned char**));
+   if (hash_table == NULL)
+      return NULL;
+   if (quality < 5) quality = 5;
+
+   stbiw__sbpush(out, 0x78);   // DEFLATE 32K window
+   stbiw__sbpush(out, 0x5e);   // FLEVEL = 1
+   stbiw__zlib_add(1,1);  // BFINAL = 1
+   stbiw__zlib_add(1,2);  // BTYPE = 1 -- fixed huffman
+
+   for (i=0; i < stbiw__ZHASH; ++i)
+      hash_table[i] = NULL;
+
+   i=0;
+   while (i < data_len-3) {
+      // hash next 3 bytes of data to be compressed
+      int h = stbiw__zhash(data+i)&(stbiw__ZHASH-1), best=3;
+      unsigned char *bestloc = 0;
+      unsigned char **hlist = hash_table[h];
+      int n = stbiw__sbcount(hlist);
+      for (j=0; j < n; ++j) {
+         if (hlist[j]-data > i-32768) { // if entry lies within window
+            int d = stbiw__zlib_countm(hlist[j], data+i, data_len-i);
+            if (d >= best) { best=d; bestloc=hlist[j]; }
+         }
+      }
+      // when hash table entry is too long, delete half the entries
+      if (hash_table[h] && stbiw__sbn(hash_table[h]) == 2*quality) {
+         STBIW_MEMMOVE(hash_table[h], hash_table[h]+quality, sizeof(hash_table[h][0])*quality);
+         stbiw__sbn(hash_table[h]) = quality;
+      }
+      stbiw__sbpush(hash_table[h],data+i);
+
+      if (bestloc) {
+         // "lazy matching" - check match at *next* byte, and if it's better, do cur byte as literal
+         h = stbiw__zhash(data+i+1)&(stbiw__ZHASH-1);
+         hlist = hash_table[h];
+         n = stbiw__sbcount(hlist);
+         for (j=0; j < n; ++j) {
+            if (hlist[j]-data > i-32767) {
+               int e = stbiw__zlib_countm(hlist[j], data+i+1, data_len-i-1);
+               if (e > best) { // if next match is better, bail on current match
+                  bestloc = NULL;
+                  break;
+               }
+            }
+         }
+      }
+
+      if (bestloc) {
+         int d = (int) (data+i - bestloc); // distance back
+         STBIW_ASSERT(d <= 32767 && best <= 258);
+         for (j=0; best > lengthc[j+1]-1; ++j);
+         stbiw__zlib_huff(j+257);
+         if (lengtheb[j]) stbiw__zlib_add(best - lengthc[j], lengtheb[j]);
+         for (j=0; d > distc[j+1]-1; ++j);
+         stbiw__zlib_add(stbiw__zlib_bitrev(j,5),5);
+         if (disteb[j]) stbiw__zlib_add(d - distc[j], disteb[j]);
+         i += best;
+      } else {
+         stbiw__zlib_huffb(data[i]);
+         ++i;
+      }
+   }
+   // write out final bytes
+   for (;i < data_len; ++i)
+      stbiw__zlib_huffb(data[i]);
+   stbiw__zlib_huff(256); // end of block
+   // pad with 0 bits to byte boundary
+   while (bitcount)
+      stbiw__zlib_add(0,1);
+
+   for (i=0; i < stbiw__ZHASH; ++i)
+      (void) stbiw__sbfree(hash_table[i]);
+   STBIW_FREE(hash_table);
+
+   // store uncompressed instead if compression was worse
+   if (stbiw__sbn(out) > data_len + 2 + ((data_len+32766)/32767)*5) {
+      stbiw__sbn(out) = 2;  // truncate to DEFLATE 32K window and FLEVEL = 1
+      for (j = 0; j < data_len;) {
+         int blocklen = data_len - j;
+         if (blocklen > 32767) blocklen = 32767;
+         stbiw__sbpush(out, data_len - j == blocklen); // BFINAL = ?, BTYPE = 0 -- no compression
+         stbiw__sbpush(out, STBIW_UCHAR(blocklen)); // LEN
+         stbiw__sbpush(out, STBIW_UCHAR(blocklen >> 8));
+         stbiw__sbpush(out, STBIW_UCHAR(~blocklen)); // NLEN
+         stbiw__sbpush(out, STBIW_UCHAR(~blocklen >> 8));
+         memcpy(out+stbiw__sbn(out), data+j, blocklen);
+         stbiw__sbn(out) += blocklen;
+         j += blocklen;
+      }
+   }
+
+   {
+      // compute adler32 on input
+      unsigned int s1=1, s2=0;
+      int blocklen = (int) (data_len % 5552);
+      j=0;
+      while (j < data_len) {
+         for (i=0; i < blocklen; ++i) { s1 += data[j+i]; s2 += s1; }
+         s1 %= 65521; s2 %= 65521;
+         j += blocklen;
+         blocklen = 5552;
+      }
+      stbiw__sbpush(out, STBIW_UCHAR(s2 >> 8));
+      stbiw__sbpush(out, STBIW_UCHAR(s2));
+      stbiw__sbpush(out, STBIW_UCHAR(s1 >> 8));
+      stbiw__sbpush(out, STBIW_UCHAR(s1));
+   }
+   *out_len = stbiw__sbn(out);
+   // make returned pointer freeable
+   STBIW_MEMMOVE(stbiw__sbraw(out), out, *out_len);
+   return (unsigned char *) stbiw__sbraw(out);
+#endif // STBIW_ZLIB_COMPRESS
+}
+
+static unsigned int stbiw__crc32(unsigned char *buffer, int len)
+{
+#ifdef STBIW_CRC32
+    return STBIW_CRC32(buffer, len);
+#else
+   static unsigned int crc_table[256] =
+   {
+      0x00000000, 0x77073096, 0xEE0E612C, 0x990951BA, 0x076DC419, 0x706AF48F, 0xE963A535, 0x9E6495A3,
+      0x0eDB8832, 0x79DCB8A4, 0xE0D5E91E, 0x97D2D988, 0x09B64C2B, 0x7EB17CBD, 0xE7B82D07, 0x90BF1D91,
+      0x1DB71064, 0x6AB020F2, 0xF3B97148, 0x84BE41DE, 0x1ADAD47D, 0x6DDDE4EB, 0xF4D4B551, 0x83D385C7,
+      0x136C9856, 0x646BA8C0, 0xFD62F97A, 0x8A65C9EC, 0x14015C4F, 0x63066CD9, 0xFA0F3D63, 0x8D080DF5,
+      0x3B6E20C8, 0x4C69105E, 0xD56041E4, 0xA2677172, 0x3C03E4D1, 0x4B04D447, 0xD20D85FD, 0xA50AB56B,
+      0x35B5A8FA, 0x42B2986C, 0xDBBBC9D6, 0xACBCF940, 0x32D86CE3, 0x45DF5C75, 0xDCD60DCF, 0xABD13D59,
+      0x26D930AC, 0x51DE003A, 0xC8D75180, 0xBFD06116, 0x21B4F4B5, 0x56B3C423, 0xCFBA9599, 0xB8BDA50F,
+      0x2802B89E, 0x5F058808, 0xC60CD9B2, 0xB10BE924, 0x2F6F7C87, 0x58684C11, 0xC1611DAB, 0xB6662D3D,
+      0x76DC4190, 0x01DB7106, 0x98D220BC, 0xEFD5102A, 0x71B18589, 0x06B6B51F, 0x9FBFE4A5, 0xE8B8D433,
+      0x7807C9A2, 0x0F00F934, 0x9609A88E, 0xE10E9818, 0x7F6A0DBB, 0x086D3D2D, 0x91646C97, 0xE6635C01,
+      0x6B6B51F4, 0x1C6C6162, 0x856530D8, 0xF262004E, 0x6C0695ED, 0x1B01A57B, 0x8208F4C1, 0xF50FC457,
+      0x65B0D9C6, 0x12B7E950, 0x8BBEB8EA, 0xFCB9887C, 0x62DD1DDF, 0x15DA2D49, 0x8CD37CF3, 0xFBD44C65,
+      0x4DB26158, 0x3AB551CE, 0xA3BC0074, 0xD4BB30E2, 0x4ADFA541, 0x3DD895D7, 0xA4D1C46D, 0xD3D6F4FB,
+      0x4369E96A, 0x346ED9FC, 0xAD678846, 0xDA60B8D0, 0x44042D73, 0x33031DE5, 0xAA0A4C5F, 0xDD0D7CC9,
+      0x5005713C, 0x270241AA, 0xBE0B1010, 0xC90C2086, 0x5768B525, 0x206F85B3, 0xB966D409, 0xCE61E49F,
+      0x5EDEF90E, 0x29D9C998, 0xB0D09822, 0xC7D7A8B4, 0x59B33D17, 0x2EB40D81, 0xB7BD5C3B, 0xC0BA6CAD,
+      0xEDB88320, 0x9ABFB3B6, 0x03B6E20C, 0x74B1D29A, 0xEAD54739, 0x9DD277AF, 0x04DB2615, 0x73DC1683,
+      0xE3630B12, 0x94643B84, 0x0D6D6A3E, 0x7A6A5AA8, 0xE40ECF0B, 0x9309FF9D, 0x0A00AE27, 0x7D079EB1,
+      0xF00F9344, 0x8708A3D2, 0x1E01F268, 0x6906C2FE, 0xF762575D, 0x806567CB, 0x196C3671, 0x6E6B06E7,
+      0xFED41B76, 0x89D32BE0, 0x10DA7A5A, 0x67DD4ACC, 0xF9B9DF6F, 0x8EBEEFF9, 0x17B7BE43, 0x60B08ED5,
+      0xD6D6A3E8, 0xA1D1937E, 0x38D8C2C4, 0x4FDFF252, 0xD1BB67F1, 0xA6BC5767, 0x3FB506DD, 0x48B2364B,
+      0xD80D2BDA, 0xAF0A1B4C, 0x36034AF6, 0x41047A60, 0xDF60EFC3, 0xA867DF55, 0x316E8EEF, 0x4669BE79,
+      0xCB61B38C, 0xBC66831A, 0x256FD2A0, 0x5268E236, 0xCC0C7795, 0xBB0B4703, 0x220216B9, 0x5505262F,
+      0xC5BA3BBE, 0xB2BD0B28, 0x2BB45A92, 0x5CB36A04, 0xC2D7FFA7, 0xB5D0CF31, 0x2CD99E8B, 0x5BDEAE1D,
+      0x9B64C2B0, 0xEC63F226, 0x756AA39C, 0x026D930A, 0x9C0906A9, 0xEB0E363F, 0x72076785, 0x05005713,
+      0x95BF4A82, 0xE2B87A14, 0x7BB12BAE, 0x0CB61B38, 0x92D28E9B, 0xE5D5BE0D, 0x7CDCEFB7, 0x0BDBDF21,
+      0x86D3D2D4, 0xF1D4E242, 0x68DDB3F8, 0x1FDA836E, 0x81BE16CD, 0xF6B9265B, 0x6FB077E1, 0x18B74777,
+      0x88085AE6, 0xFF0F6A70, 0x66063BCA, 0x11010B5C, 0x8F659EFF, 0xF862AE69, 0x616BFFD3, 0x166CCF45,
+      0xA00AE278, 0xD70DD2EE, 0x4E048354, 0x3903B3C2, 0xA7672661, 0xD06016F7, 0x4969474D, 0x3E6E77DB,
+      0xAED16A4A, 0xD9D65ADC, 0x40DF0B66, 0x37D83BF0, 0xA9BCAE53, 0xDEBB9EC5, 0x47B2CF7F, 0x30B5FFE9,
+      0xBDBDF21C, 0xCABAC28A, 0x53B39330, 0x24B4A3A6, 0xBAD03605, 0xCDD70693, 0x54DE5729, 0x23D967BF,
+      0xB3667A2E, 0xC4614AB8, 0x5D681B02, 0x2A6F2B94, 0xB40BBE37, 0xC30C8EA1, 0x5A05DF1B, 0x2D02EF8D
+   };
+
+   unsigned int crc = ~0u;
+   int i;
+   for (i=0; i < len; ++i)
+      crc = (crc >> 8) ^ crc_table[buffer[i] ^ (crc & 0xff)];
+   return ~crc;
+#endif
+}
+
+#define stbiw__wpng4(o,a,b,c,d) ((o)[0]=STBIW_UCHAR(a),(o)[1]=STBIW_UCHAR(b),(o)[2]=STBIW_UCHAR(c),(o)[3]=STBIW_UCHAR(d),(o)+=4)
+#define stbiw__wp32(data,v) stbiw__wpng4(data, (v)>>24,(v)>>16,(v)>>8,(v));
+#define stbiw__wptag(data,s) stbiw__wpng4(data, s[0],s[1],s[2],s[3])
+
+static void stbiw__wpcrc(unsigned char **data, int len)
+{
+   unsigned int crc = stbiw__crc32(*data - len - 4, len+4);
+   stbiw__wp32(*data, crc);
+}
+
+static unsigned char stbiw__paeth(int a, int b, int c)
+{
+   int p = a + b - c, pa = abs(p-a), pb = abs(p-b), pc = abs(p-c);
+   if (pa <= pb && pa <= pc) return STBIW_UCHAR(a);
+   if (pb <= pc) return STBIW_UCHAR(b);
+   return STBIW_UCHAR(c);
+}
+
+// @OPTIMIZE: provide an option that always forces left-predict or paeth predict
+static void stbiw__encode_png_line(unsigned char *pixels, int stride_bytes, int width, int height, int y, int n, int filter_type, signed char *line_buffer)
+{
+   static int mapping[] = { 0,1,2,3,4 };
+   static int firstmap[] = { 0,1,0,5,6 };
+   int *mymap = (y != 0) ? mapping : firstmap;
+   int i;
+   int type = mymap[filter_type];
+   unsigned char *z = pixels + stride_bytes * (stbi__flip_vertically_on_write ? height-1-y : y);
+   int signed_stride = stbi__flip_vertically_on_write ? -stride_bytes : stride_bytes;
+
+   if (type==0) {
+      memcpy(line_buffer, z, width*n);
+      return;
+   }
+
+   // first loop isn't optimized since it's just one pixel
+   for (i = 0; i < n; ++i) {
+      switch (type) {
+         case 1: line_buffer[i] = z[i]; break;
+         case 2: line_buffer[i] = z[i] - z[i-signed_stride]; break;
+         case 3: line_buffer[i] = z[i] - (z[i-signed_stride]>>1); break;
+         case 4: line_buffer[i] = (signed char) (z[i] - stbiw__paeth(0,z[i-signed_stride],0)); break;
+         case 5: line_buffer[i] = z[i]; break;
+         case 6: line_buffer[i] = z[i]; break;
+      }
+   }
+   switch (type) {
+      case 1: for (i=n; i < width*n; ++i) line_buffer[i] = z[i] - z[i-n]; break;
+      case 2: for (i=n; i < width*n; ++i) line_buffer[i] = z[i] - z[i-signed_stride]; break;
+      case 3: for (i=n; i < width*n; ++i) line_buffer[i] = z[i] - ((z[i-n] + z[i-signed_stride])>>1); break;
+      case 4: for (i=n; i < width*n; ++i) line_buffer[i] = z[i] - stbiw__paeth(z[i-n], z[i-signed_stride], z[i-signed_stride-n]); break;
+      case 5: for (i=n; i < width*n; ++i) line_buffer[i] = z[i] - (z[i-n]>>1); break;
+      case 6: for (i=n; i < width*n; ++i) line_buffer[i] = z[i] - stbiw__paeth(z[i-n], 0,0); break;
+   }
+}
+
+STBIWDEF unsigned char *stbi_write_png_to_mem(const unsigned char *pixels, int stride_bytes, int x, int y, int n, int *out_len)
+{
+   int force_filter = stbi_write_force_png_filter;
+   int ctype[5] = { -1, 0, 4, 2, 6 };
+   unsigned char sig[8] = { 137,80,78,71,13,10,26,10 };
+   unsigned char *out,*o, *filt, *zlib;
+   signed char *line_buffer;
+   int j,zlen;
+
+   if (stride_bytes == 0)
+      stride_bytes = x * n;
+
+   if (force_filter >= 5) {
+      force_filter = -1;
+   }
+
+   filt = (unsigned char *) STBIW_MALLOC((x*n+1) * y); if (!filt) return 0;
+   line_buffer = (signed char *) STBIW_MALLOC(x * n); if (!line_buffer) { STBIW_FREE(filt); return 0; }
+   for (j=0; j < y; ++j) {
+      int filter_type;
+      if (force_filter > -1) {
+         filter_type = force_filter;
+         stbiw__encode_png_line((unsigned char*)(pixels), stride_bytes, x, y, j, n, force_filter, line_buffer);
+      } else { // Estimate the best filter by running through all of them:
+         int best_filter = 0, best_filter_val = 0x7fffffff, est, i;
+         for (filter_type = 0; filter_type < 5; filter_type++) {
+            stbiw__encode_png_line((unsigned char*)(pixels), stride_bytes, x, y, j, n, filter_type, line_buffer);
+
+            // Estimate the entropy of the line using this filter; the less, the better.
+            est = 0;
+            for (i = 0; i < x*n; ++i) {
+               est += abs((signed char) line_buffer[i]);
+            }
+            if (est < best_filter_val) {
+               best_filter_val = est;
+               best_filter = filter_type;
+            }
+         }
+         if (filter_type != best_filter) {  // If the last iteration already got us the best filter, don't redo it
+            stbiw__encode_png_line((unsigned char*)(pixels), stride_bytes, x, y, j, n, best_filter, line_buffer);
+            filter_type = best_filter;
+         }
+      }
+      // when we get here, filter_type contains the filter type, and line_buffer contains the data
+      filt[j*(x*n+1)] = (unsigned char) filter_type;
+      STBIW_MEMMOVE(filt+j*(x*n+1)+1, line_buffer, x*n);
+   }
+   STBIW_FREE(line_buffer);
+   zlib = stbi_zlib_compress(filt, y*( x*n+1), &zlen, stbi_write_png_compression_level);
+   STBIW_FREE(filt);
+   if (!zlib) return 0;
+
+   // each tag requires 12 bytes of overhead
+   out = (unsigned char *) STBIW_MALLOC(8 + 12+13 + 12+zlen + 12);
+   if (!out) return 0;
+   *out_len = 8 + 12+13 + 12+zlen + 12;
+
+   o=out;
+   STBIW_MEMMOVE(o,sig,8); o+= 8;
+   stbiw__wp32(o, 13); // header length
+   stbiw__wptag(o, "IHDR");
+   stbiw__wp32(o, x);
+   stbiw__wp32(o, y);
+   *o++ = 8;
+   *o++ = STBIW_UCHAR(ctype[n]);
+   *o++ = 0;
+   *o++ = 0;
+   *o++ = 0;
+   stbiw__wpcrc(&o,13);
+
+   stbiw__wp32(o, zlen);
+   stbiw__wptag(o, "IDAT");
+   STBIW_MEMMOVE(o, zlib, zlen);
+   o += zlen;
+   STBIW_FREE(zlib);
+   stbiw__wpcrc(&o, zlen);
+
+   stbiw__wp32(o,0);
+   stbiw__wptag(o, "IEND");
+   stbiw__wpcrc(&o,0);
+
+   STBIW_ASSERT(o == out + *out_len);
+
+   return out;
+}
+
+#ifndef STBI_WRITE_NO_STDIO
+STBIWDEF int stbi_write_png(char const *filename, int x, int y, int comp, const void *data, int stride_bytes)
+{
+   FILE *f;
+   int len;
+   unsigned char *png = stbi_write_png_to_mem((const unsigned char *) data, stride_bytes, x, y, comp, &len);
+   if (png == NULL) return 0;
+
+   f = stbiw__fopen(filename, "wb");
+   if (!f) { STBIW_FREE(png); return 0; }
+   fwrite(png, 1, len, f);
+   fclose(f);
+   STBIW_FREE(png);
+   return 1;
+}
+#endif
+
+STBIWDEF int stbi_write_png_to_func(stbi_write_func *func, void *context, int x, int y, int comp, const void *data, int stride_bytes)
+{
+   int len;
+   unsigned char *png = stbi_write_png_to_mem((const unsigned char *) data, stride_bytes, x, y, comp, &len);
+   if (png == NULL) return 0;
+   func(context, png, len);
+   STBIW_FREE(png);
+   return 1;
+}
+
+
+/* ***************************************************************************
+ *
+ * JPEG writer
+ *
+ * This is based on Jon Olick's jo_jpeg.cpp:
+ * public domain Simple, Minimalistic JPEG writer - http://www.jonolick.com/code.html
+ */
+
+static const unsigned char stbiw__jpg_ZigZag[] = { 0,1,5,6,14,15,27,28,2,4,7,13,16,26,29,42,3,8,12,17,25,30,41,43,9,11,18,
+      24,31,40,44,53,10,19,23,32,39,45,52,54,20,22,33,38,46,51,55,60,21,34,37,47,50,56,59,61,35,36,48,49,57,58,62,63 };
+
+static void stbiw__jpg_writeBits(stbi__write_context *s, int *bitBufP, int *bitCntP, const unsigned short *bs) {
+   int bitBuf = *bitBufP, bitCnt = *bitCntP;
+   bitCnt += bs[1];
+   bitBuf |= bs[0] << (24 - bitCnt);
+   while(bitCnt >= 8) {
+      unsigned char c = (bitBuf >> 16) & 255;
+      stbiw__putc(s, c);
+      if(c == 255) {
+         stbiw__putc(s, 0);
+      }
+      bitBuf <<= 8;
+      bitCnt -= 8;
+   }
+   *bitBufP = bitBuf;
+   *bitCntP = bitCnt;
+}
+
+static void stbiw__jpg_DCT(float *d0p, float *d1p, float *d2p, float *d3p, float *d4p, float *d5p, float *d6p, float *d7p) {
+   float d0 = *d0p, d1 = *d1p, d2 = *d2p, d3 = *d3p, d4 = *d4p, d5 = *d5p, d6 = *d6p, d7 = *d7p;
+   float z1, z2, z3, z4, z5, z11, z13;
+
+   float tmp0 = d0 + d7;
+   float tmp7 = d0 - d7;
+   float tmp1 = d1 + d6;
+   float tmp6 = d1 - d6;
+   float tmp2 = d2 + d5;
+   float tmp5 = d2 - d5;
+   float tmp3 = d3 + d4;
+   float tmp4 = d3 - d4;
+
+   // Even part
+   float tmp10 = tmp0 + tmp3;   // phase 2
+   float tmp13 = tmp0 - tmp3;
+   float tmp11 = tmp1 + tmp2;
+   float tmp12 = tmp1 - tmp2;
+
+   d0 = tmp10 + tmp11;       // phase 3
+   d4 = tmp10 - tmp11;
+
+   z1 = (tmp12 + tmp13) * 0.707106781f; // c4
+   d2 = tmp13 + z1;       // phase 5
+   d6 = tmp13 - z1;
+
+   // Odd part
+   tmp10 = tmp4 + tmp5;       // phase 2
+   tmp11 = tmp5 + tmp6;
+   tmp12 = tmp6 + tmp7;
+
+   // The rotator is modified from fig 4-8 to avoid extra negations.
+   z5 = (tmp10 - tmp12) * 0.382683433f; // c6
+   z2 = tmp10 * 0.541196100f + z5; // c2-c6
+   z4 = tmp12 * 1.306562965f + z5; // c2+c6
+   z3 = tmp11 * 0.707106781f; // c4
+
+   z11 = tmp7 + z3;      // phase 5
+   z13 = tmp7 - z3;
+
+   *d5p = z13 + z2;         // phase 6
+   *d3p = z13 - z2;
+   *d1p = z11 + z4;
+   *d7p = z11 - z4;
+
+   *d0p = d0;  *d2p = d2;  *d4p = d4;  *d6p = d6;
+}
+
+static void stbiw__jpg_calcBits(int val, unsigned short bits[2]) {
+   int tmp1 = val < 0 ? -val : val;
+   val = val < 0 ? val-1 : val;
+   bits[1] = 1;
+   while(tmp1 >>= 1) {
+      ++bits[1];
+   }
+   bits[0] = val & ((1<<bits[1])-1);
+}
+
+static int stbiw__jpg_processDU(stbi__write_context *s, int *bitBuf, int *bitCnt, float *CDU, int du_stride, float *fdtbl, int DC, const unsigned short HTDC[256][2], const unsigned short HTAC[256][2]) {
+   const unsigned short EOB[2] = { HTAC[0x00][0], HTAC[0x00][1] };
+   const unsigned short M16zeroes[2] = { HTAC[0xF0][0], HTAC[0xF0][1] };
+   int dataOff, i, j, n, diff, end0pos, x, y;
+   int DU[64];
+
+   // DCT rows
+   for(dataOff=0, n=du_stride*8; dataOff<n; dataOff+=du_stride) {
+      stbiw__jpg_DCT(&CDU[dataOff], &CDU[dataOff+1], &CDU[dataOff+2], &CDU[dataOff+3], &CDU[dataOff+4], &CDU[dataOff+5], &CDU[dataOff+6], &CDU[dataOff+7]);
+   }
+   // DCT columns
+   for(dataOff=0; dataOff<8; ++dataOff) {
+      stbiw__jpg_DCT(&CDU[dataOff], &CDU[dataOff+du_stride], &CDU[dataOff+du_stride*2], &CDU[dataOff+du_stride*3], &CDU[dataOff+du_stride*4],
+                     &CDU[dataOff+du_stride*5], &CDU[dataOff+du_stride*6], &CDU[dataOff+du_stride*7]);
+   }
+   // Quantize/descale/zigzag the coefficients
+   for(y = 0, j=0; y < 8; ++y) {
+      for(x = 0; x < 8; ++x,++j) {
+         float v;
+         i = y*du_stride+x;
+         v = CDU[i]*fdtbl[j];
+         // DU[stbiw__jpg_ZigZag[j]] = (int)(v < 0 ? ceilf(v - 0.5f) : floorf(v + 0.5f));
+         // ceilf() and floorf() are C99, not C89, but I /think/ they're not needed here anyway?
+         DU[stbiw__jpg_ZigZag[j]] = (int)(v < 0 ? v - 0.5f : v + 0.5f);
+      }
+   }
+
+   // Encode DC
+   diff = DU[0] - DC;
+   if (diff == 0) {
+      stbiw__jpg_writeBits(s, bitBuf, bitCnt, HTDC[0]);
+   } else {
+      unsigned short bits[2];
+      stbiw__jpg_calcBits(diff, bits);
+      stbiw__jpg_writeBits(s, bitBuf, bitCnt, HTDC[bits[1]]);
+      stbiw__jpg_writeBits(s, bitBuf, bitCnt, bits);
+   }
+   // Encode ACs
+   end0pos = 63;
+   for(; (end0pos>0)&&(DU[end0pos]==0); --end0pos) {
+   }
+   // end0pos = first element in reverse order !=0
+   if(end0pos == 0) {
+      stbiw__jpg_writeBits(s, bitBuf, bitCnt, EOB);
+      return DU[0];
+   }
+   for(i = 1; i <= end0pos; ++i) {
+      int startpos = i;
+      int nrzeroes;
+      unsigned short bits[2];
+      for (; DU[i]==0 && i<=end0pos; ++i) {
+      }
+      nrzeroes = i-startpos;
+      if ( nrzeroes >= 16 ) {
+         int lng = nrzeroes>>4;
+         int nrmarker;
+         for (nrmarker=1; nrmarker <= lng; ++nrmarker)
+            stbiw__jpg_writeBits(s, bitBuf, bitCnt, M16zeroes);
+         nrzeroes &= 15;
+      }
+      stbiw__jpg_calcBits(DU[i], bits);
+      stbiw__jpg_writeBits(s, bitBuf, bitCnt, HTAC[(nrzeroes<<4)+bits[1]]);
+      stbiw__jpg_writeBits(s, bitBuf, bitCnt, bits);
+   }
+   if(end0pos != 63) {
+      stbiw__jpg_writeBits(s, bitBuf, bitCnt, EOB);
+   }
+   return DU[0];
+}
+
+static int stbi_write_jpg_core(stbi__write_context *s, int width, int height, int comp, const void* data, int quality) {
+   // Constants that don't pollute global namespace
+   static const unsigned char std_dc_luminance_nrcodes[] = {0,0,1,5,1,1,1,1,1,1,0,0,0,0,0,0,0};
+   static const unsigned char std_dc_luminance_values[] = {0,1,2,3,4,5,6,7,8,9,10,11};
+   static const unsigned char std_ac_luminance_nrcodes[] = {0,0,2,1,3,3,2,4,3,5,5,4,4,0,0,1,0x7d};
+   static const unsigned char std_ac_luminance_values[] = {
+      0x01,0x02,0x03,0x00,0x04,0x11,0x05,0x12,0x21,0x31,0x41,0x06,0x13,0x51,0x61,0x07,0x22,0x71,0x14,0x32,0x81,0x91,0xa1,0x08,
+      0x23,0x42,0xb1,0xc1,0x15,0x52,0xd1,0xf0,0x24,0x33,0x62,0x72,0x82,0x09,0x0a,0x16,0x17,0x18,0x19,0x1a,0x25,0x26,0x27,0x28,
+      0x29,0x2a,0x34,0x35,0x36,0x37,0x38,0x39,0x3a,0x43,0x44,0x45,0x46,0x47,0x48,0x49,0x4a,0x53,0x54,0x55,0x56,0x57,0x58,0x59,
+      0x5a,0x63,0x64,0x65,0x66,0x67,0x68,0x69,0x6a,0x73,0x74,0x75,0x76,0x77,0x78,0x79,0x7a,0x83,0x84,0x85,0x86,0x87,0x88,0x89,
+      0x8a,0x92,0x93,0x94,0x95,0x96,0x97,0x98,0x99,0x9a,0xa2,0xa3,0xa4,0xa5,0xa6,0xa7,0xa8,0xa9,0xaa,0xb2,0xb3,0xb4,0xb5,0xb6,
+      0xb7,0xb8,0xb9,0xba,0xc2,0xc3,0xc4,0xc5,0xc6,0xc7,0xc8,0xc9,0xca,0xd2,0xd3,0xd4,0xd5,0xd6,0xd7,0xd8,0xd9,0xda,0xe1,0xe2,
+      0xe3,0xe4,0xe5,0xe6,0xe7,0xe8,0xe9,0xea,0xf1,0xf2,0xf3,0xf4,0xf5,0xf6,0xf7,0xf8,0xf9,0xfa
+   };
+   static const unsigned char std_dc_chrominance_nrcodes[] = {0,0,3,1,1,1,1,1,1,1,1,1,0,0,0,0,0};
+   static const unsigned char std_dc_chrominance_values[] = {0,1,2,3,4,5,6,7,8,9,10,11};
+   static const unsigned char std_ac_chrominance_nrcodes[] = {0,0,2,1,2,4,4,3,4,7,5,4,4,0,1,2,0x77};
+   static const unsigned char std_ac_chrominance_values[] = {
+      0x00,0x01,0x02,0x03,0x11,0x04,0x05,0x21,0x31,0x06,0x12,0x41,0x51,0x07,0x61,0x71,0x13,0x22,0x32,0x81,0x08,0x14,0x42,0x91,
+      0xa1,0xb1,0xc1,0x09,0x23,0x33,0x52,0xf0,0x15,0x62,0x72,0xd1,0x0a,0x16,0x24,0x34,0xe1,0x25,0xf1,0x17,0x18,0x19,0x1a,0x26,
+      0x27,0x28,0x29,0x2a,0x35,0x36,0x37,0x38,0x39,0x3a,0x43,0x44,0x45,0x46,0x47,0x48,0x49,0x4a,0x53,0x54,0x55,0x56,0x57,0x58,
+      0x59,0x5a,0x63,0x64,0x65,0x66,0x67,0x68,0x69,0x6a,0x73,0x74,0x75,0x76,0x77,0x78,0x79,0x7a,0x82,0x83,0x84,0x85,0x86,0x87,
+      0x88,0x89,0x8a,0x92,0x93,0x94,0x95,0x96,0x97,0x98,0x99,0x9a,0xa2,0xa3,0xa4,0xa5,0xa6,0xa7,0xa8,0xa9,0xaa,0xb2,0xb3,0xb4,
+      0xb5,0xb6,0xb7,0xb8,0xb9,0xba,0xc2,0xc3,0xc4,0xc5,0xc6,0xc7,0xc8,0xc9,0xca,0xd2,0xd3,0xd4,0xd5,0xd6,0xd7,0xd8,0xd9,0xda,
+      0xe2,0xe3,0xe4,0xe5,0xe6,0xe7,0xe8,0xe9,0xea,0xf2,0xf3,0xf4,0xf5,0xf6,0xf7,0xf8,0xf9,0xfa
+   };
+   // Huffman tables
+   static const unsigned short YDC_HT[256][2] = { {0,2},{2,3},{3,3},{4,3},{5,3},{6,3},{14,4},{30,5},{62,6},{126,7},{254,8},{510,9}};
+   static const unsigned short UVDC_HT[256][2] = { {0,2},{1,2},{2,2},{6,3},{14,4},{30,5},{62,6},{126,7},{254,8},{510,9},{1022,10},{2046,11}};
+   static const unsigned short YAC_HT[256][2] = {
+      {10,4},{0,2},{1,2},{4,3},{11,4},{26,5},{120,7},{248,8},{1014,10},{65410,16},{65411,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {12,4},{27,5},{121,7},{502,9},{2038,11},{65412,16},{65413,16},{65414,16},{65415,16},{65416,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {28,5},{249,8},{1015,10},{4084,12},{65417,16},{65418,16},{65419,16},{65420,16},{65421,16},{65422,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {58,6},{503,9},{4085,12},{65423,16},{65424,16},{65425,16},{65426,16},{65427,16},{65428,16},{65429,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {59,6},{1016,10},{65430,16},{65431,16},{65432,16},{65433,16},{65434,16},{65435,16},{65436,16},{65437,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {122,7},{2039,11},{65438,16},{65439,16},{65440,16},{65441,16},{65442,16},{65443,16},{65444,16},{65445,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {123,7},{4086,12},{65446,16},{65447,16},{65448,16},{65449,16},{65450,16},{65451,16},{65452,16},{65453,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {250,8},{4087,12},{65454,16},{65455,16},{65456,16},{65457,16},{65458,16},{65459,16},{65460,16},{65461,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {504,9},{32704,15},{65462,16},{65463,16},{65464,16},{65465,16},{65466,16},{65467,16},{65468,16},{65469,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {505,9},{65470,16},{65471,16},{65472,16},{65473,16},{65474,16},{65475,16},{65476,16},{65477,16},{65478,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {506,9},{65479,16},{65480,16},{65481,16},{65482,16},{65483,16},{65484,16},{65485,16},{65486,16},{65487,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {1017,10},{65488,16},{65489,16},{65490,16},{65491,16},{65492,16},{65493,16},{65494,16},{65495,16},{65496,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {1018,10},{65497,16},{65498,16},{65499,16},{65500,16},{65501,16},{65502,16},{65503,16},{65504,16},{65505,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {2040,11},{65506,16},{65507,16},{65508,16},{65509,16},{65510,16},{65511,16},{65512,16},{65513,16},{65514,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {65515,16},{65516,16},{65517,16},{65518,16},{65519,16},{65520,16},{65521,16},{65522,16},{65523,16},{65524,16},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {2041,11},{65525,16},{65526,16},{65527,16},{65528,16},{65529,16},{65530,16},{65531,16},{65532,16},{65533,16},{65534,16},{0,0},{0,0},{0,0},{0,0},{0,0}
+   };
+   static const unsigned short UVAC_HT[256][2] = {
+      {0,2},{1,2},{4,3},{10,4},{24,5},{25,5},{56,6},{120,7},{500,9},{1014,10},{4084,12},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {11,4},{57,6},{246,8},{501,9},{2038,11},{4085,12},{65416,16},{65417,16},{65418,16},{65419,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {26,5},{247,8},{1015,10},{4086,12},{32706,15},{65420,16},{65421,16},{65422,16},{65423,16},{65424,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {27,5},{248,8},{1016,10},{4087,12},{65425,16},{65426,16},{65427,16},{65428,16},{65429,16},{65430,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {58,6},{502,9},{65431,16},{65432,16},{65433,16},{65434,16},{65435,16},{65436,16},{65437,16},{65438,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {59,6},{1017,10},{65439,16},{65440,16},{65441,16},{65442,16},{65443,16},{65444,16},{65445,16},{65446,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {121,7},{2039,11},{65447,16},{65448,16},{65449,16},{65450,16},{65451,16},{65452,16},{65453,16},{65454,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {122,7},{2040,11},{65455,16},{65456,16},{65457,16},{65458,16},{65459,16},{65460,16},{65461,16},{65462,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {249,8},{65463,16},{65464,16},{65465,16},{65466,16},{65467,16},{65468,16},{65469,16},{65470,16},{65471,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {503,9},{65472,16},{65473,16},{65474,16},{65475,16},{65476,16},{65477,16},{65478,16},{65479,16},{65480,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {504,9},{65481,16},{65482,16},{65483,16},{65484,16},{65485,16},{65486,16},{65487,16},{65488,16},{65489,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {505,9},{65490,16},{65491,16},{65492,16},{65493,16},{65494,16},{65495,16},{65496,16},{65497,16},{65498,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {506,9},{65499,16},{65500,16},{65501,16},{65502,16},{65503,16},{65504,16},{65505,16},{65506,16},{65507,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {2041,11},{65508,16},{65509,16},{65510,16},{65511,16},{65512,16},{65513,16},{65514,16},{65515,16},{65516,16},{0,0},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {16352,14},{65517,16},{65518,16},{65519,16},{65520,16},{65521,16},{65522,16},{65523,16},{65524,16},{65525,16},{0,0},{0,0},{0,0},{0,0},{0,0},
+      {1018,10},{32707,15},{65526,16},{65527,16},{65528,16},{65529,16},{65530,16},{65531,16},{65532,16},{65533,16},{65534,16},{0,0},{0,0},{0,0},{0,0},{0,0}
+   };
+   static const int YQT[] = {16,11,10,16,24,40,51,61,12,12,14,19,26,58,60,55,14,13,16,24,40,57,69,56,14,17,22,29,51,87,80,62,18,22,
+                             37,56,68,109,103,77,24,35,55,64,81,104,113,92,49,64,78,87,103,121,120,101,72,92,95,98,112,100,103,99};
+   static const int UVQT[] = {17,18,24,47,99,99,99,99,18,21,26,66,99,99,99,99,24,26,56,99,99,99,99,99,47,66,99,99,99,99,99,99,
+                              99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99,99};
+   static const float aasf[] = { 1.0f * 2.828427125f, 1.387039845f * 2.828427125f, 1.306562965f * 2.828427125f, 1.175875602f * 2.828427125f,
+                                 1.0f * 2.828427125f, 0.785694958f * 2.828427125f, 0.541196100f * 2.828427125f, 0.275899379f * 2.828427125f };
+
+   int row, col, i, k, subsample;
+   float fdtbl_Y[64], fdtbl_UV[64];
+   unsigned char YTable[64], UVTable[64];
+
+   if(!data || !width || !height || comp > 4 || comp < 1) {
+      return 0;
+   }
+
+   quality = quality ? quality : 90;
+   subsample = quality <= 90 ? 1 : 0;
+   quality = quality < 1 ? 1 : quality > 100 ? 100 : quality;
+   quality = quality < 50 ? 5000 / quality : 200 - quality * 2;
+
+   for(i = 0; i < 64; ++i) {
+      int uvti, yti = (YQT[i]*quality+50)/100;
+      YTable[stbiw__jpg_ZigZag[i]] = (unsigned char) (yti < 1 ? 1 : yti > 255 ? 255 : yti);
+      uvti = (UVQT[i]*quality+50)/100;
+      UVTable[stbiw__jpg_ZigZag[i]] = (unsigned char) (uvti < 1 ? 1 : uvti > 255 ? 255 : uvti);
+   }
+
+   for(row = 0, k = 0; row < 8; ++row) {
+      for(col = 0; col < 8; ++col, ++k) {
+         fdtbl_Y[k]  = 1 / (YTable [stbiw__jpg_ZigZag[k]] * aasf[row] * aasf[col]);
+         fdtbl_UV[k] = 1 / (UVTable[stbiw__jpg_ZigZag[k]] * aasf[row] * aasf[col]);
+      }
+   }
+
+   // Write Headers
+   {
+      static const unsigned char head0[] = { 0xFF,0xD8,0xFF,0xE0,0,0x10,'J','F','I','F',0,1,1,0,0,1,0,1,0,0,0xFF,0xDB,0,0x84,0 };
+      static const unsigned char head2[] = { 0xFF,0xDA,0,0xC,3,1,0,2,0x11,3,0x11,0,0x3F,0 };
+      const unsigned char head1[] = { 0xFF,0xC0,0,0x11,8,(unsigned char)(height>>8),STBIW_UCHAR(height),(unsigned char)(width>>8),STBIW_UCHAR(width),
+                                      3,1,(unsigned char)(subsample?0x22:0x11),0,2,0x11,1,3,0x11,1,0xFF,0xC4,0x01,0xA2,0 };
+      s->func(s->context, (void*)head0, sizeof(head0));
+      s->func(s->context, (void*)YTable, sizeof(YTable));
+      stbiw__putc(s, 1);
+      s->func(s->context, UVTable, sizeof(UVTable));
+      s->func(s->context, (void*)head1, sizeof(head1));
+      s->func(s->context, (void*)(std_dc_luminance_nrcodes+1), sizeof(std_dc_luminance_nrcodes)-1);
+      s->func(s->context, (void*)std_dc_luminance_values, sizeof(std_dc_luminance_values));
+      stbiw__putc(s, 0x10); // HTYACinfo
+      s->func(s->context, (void*)(std_ac_luminance_nrcodes+1), sizeof(std_ac_luminance_nrcodes)-1);
+      s->func(s->context, (void*)std_ac_luminance_values, sizeof(std_ac_luminance_values));
+      stbiw__putc(s, 1); // HTUDCinfo
+      s->func(s->context, (void*)(std_dc_chrominance_nrcodes+1), sizeof(std_dc_chrominance_nrcodes)-1);
+      s->func(s->context, (void*)std_dc_chrominance_values, sizeof(std_dc_chrominance_values));
+      stbiw__putc(s, 0x11); // HTUACinfo
+      s->func(s->context, (void*)(std_ac_chrominance_nrcodes+1), sizeof(std_ac_chrominance_nrcodes)-1);
+      s->func(s->context, (void*)std_ac_chrominance_values, sizeof(std_ac_chrominance_values));
+      s->func(s->context, (void*)head2, sizeof(head2));
+   }
+
+   // Encode 8x8 macroblocks
+   {
+      static const unsigned short fillBits[] = {0x7F, 7};
+      int DCY=0, DCU=0, DCV=0;
+      int bitBuf=0, bitCnt=0;
+      // comp == 2 is grey+alpha (alpha is ignored)
+      int ofsG = comp > 2 ? 1 : 0, ofsB = comp > 2 ? 2 : 0;
+      const unsigned char *dataR = (const unsigned char *)data;
+      const unsigned char *dataG = dataR + ofsG;
+      const unsigned char *dataB = dataR + ofsB;
+      int x, y, pos;
+      if(subsample) {
+         for(y = 0; y < height; y += 16) {
+            for(x = 0; x < width; x += 16) {
+               float Y[256], U[256], V[256];
+               for(row = y, pos = 0; row < y+16; ++row) {
+                  // row >= height => use last input row
+                  int clamped_row = (row < height) ? row : height - 1;
+                  int base_p = (stbi__flip_vertically_on_write ? (height-1-clamped_row) : clamped_row)*width*comp;
+                  for(col = x; col < x+16; ++col, ++pos) {
+                     // if col >= width => use pixel from last input column
+                     int p = base_p + ((col < width) ? col : (width-1))*comp;
+                     float r = dataR[p], g = dataG[p], b = dataB[p];
+                     Y[pos]= +0.29900f*r + 0.58700f*g + 0.11400f*b - 128;
+                     U[pos]= -0.16874f*r - 0.33126f*g + 0.50000f*b;
+                     V[pos]= +0.50000f*r - 0.41869f*g - 0.08131f*b;
+                  }
+               }
+               DCY = stbiw__jpg_processDU(s, &bitBuf, &bitCnt, Y+0,   16, fdtbl_Y, DCY, YDC_HT, YAC_HT);
+               DCY = stbiw__jpg_processDU(s, &bitBuf, &bitCnt, Y+8,   16, fdtbl_Y, DCY, YDC_HT, YAC_HT);
+               DCY = stbiw__jpg_processDU(s, &bitBuf, &bitCnt, Y+128, 16, fdtbl_Y, DCY, YDC_HT, YAC_HT);
+               DCY = stbiw__jpg_processDU(s, &bitBuf, &bitCnt, Y+136, 16, fdtbl_Y, DCY, YDC_HT, YAC_HT);
+
+               // subsample U,V
+               {
+                  float subU[64], subV[64];
+                  int yy, xx;
+                  for(yy = 0, pos = 0; yy < 8; ++yy) {
+                     for(xx = 0; xx < 8; ++xx, ++pos) {
+                        int j = yy*32+xx*2;
+                        subU[pos] = (U[j+0] + U[j+1] + U[j+16] + U[j+17]) * 0.25f;
+                        subV[pos] = (V[j+0] + V[j+1] + V[j+16] + V[j+17]) * 0.25f;
+                     }
+                  }
+                  DCU = stbiw__jpg_processDU(s, &bitBuf, &bitCnt, subU, 8, fdtbl_UV, DCU, UVDC_HT, UVAC_HT);
+                  DCV = stbiw__jpg_processDU(s, &bitBuf, &bitCnt, subV, 8, fdtbl_UV, DCV, UVDC_HT, UVAC_HT);
+               }
+            }
+         }
+      } else {
+         for(y = 0; y < height; y += 8) {
+            for(x = 0; x < width; x += 8) {
+               float Y[64], U[64], V[64];
+               for(row = y, pos = 0; row < y+8; ++row) {
+                  // row >= height => use last input row
+                  int clamped_row = (row < height) ? row : height - 1;
+                  int base_p = (stbi__flip_vertically_on_write ? (height-1-clamped_row) : clamped_row)*width*comp;
+                  for(col = x; col < x+8; ++col, ++pos) {
+                     // if col >= width => use pixel from last input column
+                     int p = base_p + ((col < width) ? col : (width-1))*comp;
+                     float r = dataR[p], g = dataG[p], b = dataB[p];
+                     Y[pos]= +0.29900f*r + 0.58700f*g + 0.11400f*b - 128;
+                     U[pos]= -0.16874f*r - 0.33126f*g + 0.50000f*b;
+                     V[pos]= +0.50000f*r - 0.41869f*g - 0.08131f*b;
+                  }
+               }
+
+               DCY = stbiw__jpg_processDU(s, &bitBuf, &bitCnt, Y, 8, fdtbl_Y,  DCY, YDC_HT, YAC_HT);
+               DCU = stbiw__jpg_processDU(s, &bitBuf, &bitCnt, U, 8, fdtbl_UV, DCU, UVDC_HT, UVAC_HT);
+               DCV = stbiw__jpg_processDU(s, &bitBuf, &bitCnt, V, 8, fdtbl_UV, DCV, UVDC_HT, UVAC_HT);
+            }
+         }
+      }
+
+      // Do the bit alignment of the EOI marker
+      stbiw__jpg_writeBits(s, &bitBuf, &bitCnt, fillBits);
+   }
+
+   // EOI
+   stbiw__putc(s, 0xFF);
+   stbiw__putc(s, 0xD9);
+
+   return 1;
+}
+
+STBIWDEF int stbi_write_jpg_to_func(stbi_write_func *func, void *context, int x, int y, int comp, const void *data, int quality)
+{
+   stbi__write_context s = { 0 };
+   stbi__start_write_callbacks(&s, func, context);
+   return stbi_write_jpg_core(&s, x, y, comp, (void *) data, quality);
+}
+
+
+#ifndef STBI_WRITE_NO_STDIO
+STBIWDEF int stbi_write_jpg(char const *filename, int x, int y, int comp, const void *data, int quality)
+{
+   stbi__write_context s = { 0 };
+   if (stbi__start_write_file(&s,filename)) {
+      int r = stbi_write_jpg_core(&s, x, y, comp, data, quality);
+      stbi__end_write_file(&s);
+      return r;
+   } else
+      return 0;
+}
+#endif
+
+#endif // STB_IMAGE_WRITE_IMPLEMENTATION
+
+/* Revision history
+      1.16  (2021-07-11)
+             make Deflate code emit uncompressed blocks when it would otherwise expand
+             support writing BMPs with alpha channel
+      1.15  (2020-07-13) unknown
+      1.14  (2020-02-02) updated JPEG writer to downsample chroma channels
+      1.13
+      1.12
+      1.11  (2019-08-11)
+
+      1.10  (2019-02-07)
+             support utf8 filenames in Windows; fix warnings and platform ifdefs
+      1.09  (2018-02-11)
+             fix typo in zlib quality API, improve STB_I_W_STATIC in C++
+      1.08  (2018-01-29)
+             add stbi__flip_vertically_on_write, external zlib, zlib quality, choose PNG filter
+      1.07  (2017-07-24)
+             doc fix
+      1.06 (2017-07-23)
+             writing JPEG (using Jon Olick's code)
+      1.05   ???
+      1.04 (2017-03-03)
+             monochrome BMP expansion
+      1.03   ???
+      1.02 (2016-04-02)
+             avoid allocating large structures on the stack
+      1.01 (2016-01-16)
+             STBIW_REALLOC_SIZED: support allocators with no realloc support
+             avoid race-condition in crc initialization
+             minor compile issues
+      1.00 (2015-09-14)
+             installable file IO function
+      0.99 (2015-09-13)
+             warning fixes; TGA rle support
+      0.98 (2015-04-08)
+             added STBIW_MALLOC, STBIW_ASSERT etc
+      0.97 (2015-01-18)
+             fixed HDR asserts, rewrote HDR rle logic
+      0.96 (2015-01-17)
+             add HDR output
+             fix monochrome BMP
+      0.95 (2014-08-17)
+             add monochrome TGA output
+      0.94 (2014-05-31)
+             rename private functions to avoid conflicts with stb_image.h
+      0.93 (2014-05-27)
+             warning fixes
+      0.92 (2010-08-01)
+             casts to unsigned char to fix warnings
+      0.91 (2010-07-17)
+             first public release
+      0.90   first internal release
+*/
+
+/*
+------------------------------------------------------------------------------
+This software is available under 2 licenses -- choose whichever you prefer.
+------------------------------------------------------------------------------
+ALTERNATIVE A - MIT License
+Copyright (c) 2017 Sean Barrett
+Permission is hereby granted, free of charge, to any person obtaining a copy of
+this software and associated documentation files (the "Software"), to deal in
+the Software without restriction, including without limitation the rights to
+use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies
+of the Software, and to permit persons to whom the Software is furnished to do
+so, subject to the following conditions:
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
+------------------------------------------------------------------------------
+ALTERNATIVE B - Public Domain (www.unlicense.org)
+This is free and unencumbered software released into the public domain.
+Anyone is free to copy, modify, publish, use, compile, sell, or distribute this
+software, either in source code form or as a compiled binary, for any purpose,
+commercial or non-commercial, and by any means.
+In jurisdictions that recognize copyright laws, the author or authors of this
+software dedicate any and all copyright interest in the software to the public
+domain. We make this dedication for the benefit of the public at large and to
+the detriment of our heirs and successors. We intend this dedication to be an
+overt act of relinquishment in perpetuity of all present and future rights to
+this software under copyright law.
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
+WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+------------------------------------------------------------------------------
+*/

ggml/examples/whisper/CMakeLists.txt

@@ -0,0 +1,22 @@
+#
+# whisper
+
+add_library(whisper-cpp
+    whisper.cpp
+    )
+
+target_link_libraries(whisper-cpp PRIVATE
+    ggml
+    )
+
+set(TEST_TARGET whisper)
+add_executable(${TEST_TARGET} main.cpp)
+target_link_libraries(${TEST_TARGET} PRIVATE whisper-cpp common)
+target_include_directories(${TEST_TARGET} PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/..)
+
+#
+# whisper-quantize
+
+set(TEST_TARGET whisper-quantize)
+add_executable(${TEST_TARGET} quantize.cpp)
+target_link_libraries(${TEST_TARGET} PRIVATE ggml common common-ggml)

ggml/examples/whisper/README.md

@@ -0,0 +1,29 @@
+# whisper
+
+Port of [OpenAI's Whisper](https://github.com/openai/whisper) ASR model in C/C++ using
+[ggml](https://github.com/ggerganov/ggml)
+
+## More info
+
+Checkout https://github.com/ggerganov/whisper.cpp
+
+## Memory usage
+
+| Model  | Disk   | Mem     |
+| ---    | ---    | ---     |
+| tiny   |  75 MB | ~280 MB |
+| base   | 142 MB | ~430 MB |
+| small  | 466 MB | ~1.0 GB |
+| medium | 1.5 GB | ~2.6 GB |
+| large  | 2.9 GB | ~4.7 GB |
+
+## ggml format
+
+The original models are converted to a custom binary format. This allows to pack everything needed into a single file:
+
+- model parameters
+- mel filters
+- vocabulary
+- weights
+
+For more details, see the conversion script [convert-pt-to-ggml.py](convert-pt-to-ggml.py)

ggml/examples/whisper/convert-pt-to-ggml.py

@@ -0,0 +1,342 @@
+# Convert Whisper transformer model from PyTorch to ggml format
+#
+# Usage: python convert-pt-to-ggml.py ~/.cache/whisper/medium.pt ~/path/to/repo/whisper/ ./models/whisper-medium
+#
+# You need to clone the original repo in ~/path/to/repo/whisper/
+#
+#  git clone https://github.com/openai/whisper ~/path/to/repo/whisper/
+#
+# It is used to various assets needed by the algorithm:
+#
+#  - tokenizer
+#  - mel filters
+#
+# Also, you need to have the original models in ~/.cache/whisper/
+# See the original repo for more details.
+#
+# This script loads the specified model and whisper assets and saves them in ggml format.
+# The output is a single binary file containing the following information:
+#
+#  - hparams
+#  - mel filters
+#  - tokenizer vocab
+#  - model variables
+#
+# For each variable, write the following:
+#
+#  - Number of dimensions (int)
+#  - Name length (int)
+#  - Dimensions (int[n_dims])
+#  - Name (char[name_length])
+#  - Data (float[n_dims])
+#
+
+import io
+import os
+import sys
+import struct
+import json
+import code
+import torch
+import numpy as np
+import base64
+from pathlib import Path
+#from transformers import GPTJForCausalLM
+#from transformers import GPT2TokenizerFast
+
+# ref: https://github.com/openai/whisper/blob/8cf36f3508c9acd341a45eb2364239a3d81458b9/whisper/tokenizer.py#L10-L110
+#LANGUAGES = {
+#    "en": "english",
+#    "zh": "chinese",
+#    "de": "german",
+#    "es": "spanish",
+#    "ru": "russian",
+#    "ko": "korean",
+#    "fr": "french",
+#    "ja": "japanese",
+#    "pt": "portuguese",
+#    "tr": "turkish",
+#    "pl": "polish",
+#    "ca": "catalan",
+#    "nl": "dutch",
+#    "ar": "arabic",
+#    "sv": "swedish",
+#    "it": "italian",
+#    "id": "indonesian",
+#    "hi": "hindi",
+#    "fi": "finnish",
+#    "vi": "vietnamese",
+#    "iw": "hebrew",
+#    "uk": "ukrainian",
+#    "el": "greek",
+#    "ms": "malay",
+#    "cs": "czech",
+#    "ro": "romanian",
+#    "da": "danish",
+#    "hu": "hungarian",
+#    "ta": "tamil",
+#    "no": "norwegian",
+#    "th": "thai",
+#    "ur": "urdu",
+#    "hr": "croatian",
+#    "bg": "bulgarian",
+#    "lt": "lithuanian",
+#    "la": "latin",
+#    "mi": "maori",
+#    "ml": "malayalam",
+#    "cy": "welsh",
+#    "sk": "slovak",
+#    "te": "telugu",
+#    "fa": "persian",
+#    "lv": "latvian",
+#    "bn": "bengali",
+#    "sr": "serbian",
+#    "az": "azerbaijani",
+#    "sl": "slovenian",
+#    "kn": "kannada",
+#    "et": "estonian",
+#    "mk": "macedonian",
+#    "br": "breton",
+#    "eu": "basque",
+#    "is": "icelandic",
+#    "hy": "armenian",
+#    "ne": "nepali",
+#    "mn": "mongolian",
+#    "bs": "bosnian",
+#    "kk": "kazakh",
+#    "sq": "albanian",
+#    "sw": "swahili",
+#    "gl": "galician",
+#    "mr": "marathi",
+#    "pa": "punjabi",
+#    "si": "sinhala",
+#    "km": "khmer",
+#    "sn": "shona",
+#    "yo": "yoruba",
+#    "so": "somali",
+#    "af": "afrikaans",
+#    "oc": "occitan",
+#    "ka": "georgian",
+#    "be": "belarusian",
+#    "tg": "tajik",
+#    "sd": "sindhi",
+#    "gu": "gujarati",
+#    "am": "amharic",
+#    "yi": "yiddish",
+#    "lo": "lao",
+#    "uz": "uzbek",
+#    "fo": "faroese",
+#    "ht": "haitian creole",
+#    "ps": "pashto",
+#    "tk": "turkmen",
+#    "nn": "nynorsk",
+#    "mt": "maltese",
+#    "sa": "sanskrit",
+#    "lb": "luxembourgish",
+#    "my": "myanmar",
+#    "bo": "tibetan",
+#    "tl": "tagalog",
+#    "mg": "malagasy",
+#    "as": "assamese",
+#    "tt": "tatar",
+#    "haw": "hawaiian",
+#    "ln": "lingala",
+#    "ha": "hausa",
+#    "ba": "bashkir",
+#    "jw": "javanese",
+#    "su": "sundanese",
+#}
+
+## ref: https://github.com/openai/whisper/blob/8cf36f3508c9acd341a45eb2364239a3d81458b9/whisper/tokenizer.py#L273-L292
+#def build_tokenizer(path_to_whisper_repo: str, name: str = "gpt2"):
+#    os.environ["TOKENIZERS_PARALLELISM"] = "false"
+#    path = os.path.join(path_to_whisper_repo, "whisper/assets", name)
+#    tokenizer = GPT2TokenizerFast.from_pretrained(path)
+#
+#    specials = [
+#        "<|startoftranscript|>",
+#        *[f"<|{lang}|>" for lang in LANGUAGES.keys()],
+#        "<|translate|>",
+#        "<|transcribe|>",
+#        "<|startoflm|>",
+#        "<|startofprev|>",
+#        "<|nocaptions|>",
+#        "<|notimestamps|>",
+#    ]
+#
+#    tokenizer.add_special_tokens(dict(additional_special_tokens=specials))
+#    return tokenizer
+
+# ref: https://github.com/openai/gpt-2/blob/master/src/encoder.py
+def bytes_to_unicode():
+    """
+    Returns list of utf-8 byte and a corresponding list of unicode strings.
+    The reversible bpe codes work on unicode strings.
+    This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.
+    When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.
+    This is a signficant percentage of your normal, say, 32K bpe vocab.
+    To avoid that, we want lookup tables between utf-8 bytes and unicode strings.
+    And avoids mapping to whitespace/control characters the bpe code barfs on.
+    """
+    bs = list(range(ord("!"), ord("~")+1))+list(range(ord("¡"), ord("¬")+1))+list(range(ord("®"), ord("ÿ")+1))
+    cs = bs[:]
+    n = 0
+    for b in range(2**8):
+        if b not in bs:
+            bs.append(b)
+            cs.append(2**8+n)
+            n += 1
+    cs = [chr(n) for n in cs]
+    return dict(zip(bs, cs))
+
+
+if len(sys.argv) < 4:
+    print("Usage: convert-pt-to-ggml.py model.pt path-to-whisper-repo dir-output [use-f32]\n")
+    sys.exit(1)
+
+fname_inp   = Path(sys.argv[1])
+dir_whisper = Path(sys.argv[2])
+dir_out     = Path(sys.argv[3])
+
+# try to load PyTorch binary data
+try:
+    model_bytes = open(fname_inp, "rb").read()
+    with io.BytesIO(model_bytes) as fp:
+        checkpoint = torch.load(fp, map_location="cpu")
+except Exception:
+    print("Error: failed to load PyTorch model file:" , fname_inp)
+    sys.exit(1)
+
+hparams = checkpoint["dims"]
+print("hparams:", hparams)
+
+list_vars = checkpoint["model_state_dict"]
+
+#print(list_vars['encoder.positional_embedding'])
+#print(list_vars['encoder.conv1.weight'])
+#print(list_vars['encoder.conv1.weight'].shape)
+
+# load mel filters
+n_mels = hparams["n_mels"]
+with np.load(dir_whisper / "whisper" / "assets" / "mel_filters.npz") as f:
+    filters = torch.from_numpy(f[f"mel_{n_mels}"])
+    #print (filters)
+
+#code.interact(local=locals())
+
+# load tokenizer
+# for backwards compatibility, also check for older hf_transformers format tokenizer files
+# old format: dir_whisper/whisper/assets/[multilingual/gpt2]/vocab.json
+# new format: dir_whisper/whisper/assets/[multilingual/gpt2].tiktoken
+multilingual = hparams["n_vocab"] == 51865
+tokenizer = dir_whisper / "whisper" / "assets" / (multilingual and "multilingual.tiktoken" or "gpt2.tiktoken")
+tokenizer_type = "tiktoken"
+if not tokenizer.is_file():
+    tokenizer = dir_whisper / "whisper" / "assets" / (multilingual and "multilingual" or "gpt2") / "vocab.json"
+    tokenizer_type = "hf_transformers"
+    if not tokenizer.is_file():
+        print("Error: failed to find either tiktoken or hf_transformers tokenizer file:", tokenizer)
+        sys.exit(1)
+
+byte_encoder = bytes_to_unicode()
+byte_decoder = {v:k for k, v in byte_encoder.items()}
+
+if tokenizer_type == "tiktoken":
+    with open(tokenizer, "rb") as f:
+        contents = f.read()
+        tokens = {base64.b64decode(token): int(rank) for token, rank in (line.split() for line in contents.splitlines() if line)}
+elif tokenizer_type == "hf_transformers":
+    with open(tokenizer, "r", encoding="utf8") as f:
+        _tokens_raw = json.load(f)
+        if '<|endoftext|>' in _tokens_raw:
+            # ensures exact same model as tokenizer_type == tiktoken
+            # details: https://github.com/ggerganov/whisper.cpp/pull/725
+            del _tokens_raw['<|endoftext|>']
+        tokens = {bytes([byte_decoder[c] for c in token]): int(idx) for token, idx in _tokens_raw.items()}
+
+# output in the same directory as the model
+fname_out = dir_out / "ggml-model.bin"
+
+# use 16-bit or 32-bit floats
+use_f16 = True
+if len(sys.argv) > 4:
+    use_f16 = False
+    fname_out = dir_out / "ggml-model-f32.bin"
+
+fout = fname_out.open("wb")
+
+fout.write(struct.pack("i", 0x67676d6c)) # magic: ggml in hex
+fout.write(struct.pack("i", hparams["n_vocab"]))
+fout.write(struct.pack("i", hparams["n_audio_ctx"]))
+fout.write(struct.pack("i", hparams["n_audio_state"]))
+fout.write(struct.pack("i", hparams["n_audio_head"]))
+fout.write(struct.pack("i", hparams["n_audio_layer"]))
+fout.write(struct.pack("i", hparams["n_text_ctx"]))
+fout.write(struct.pack("i", hparams["n_text_state"]))
+fout.write(struct.pack("i", hparams["n_text_head"]))
+fout.write(struct.pack("i", hparams["n_text_layer"]))
+fout.write(struct.pack("i", hparams["n_mels"]))
+fout.write(struct.pack("i", use_f16))
+
+# write mel filters
+fout.write(struct.pack("i", filters.shape[0]))
+fout.write(struct.pack("i", filters.shape[1]))
+for i in range(filters.shape[0]):
+    for j in range(filters.shape[1]):
+        fout.write(struct.pack("f", filters[i][j]))
+
+# write tokenizer
+fout.write(struct.pack("i", len(tokens)))
+
+for key in tokens:
+    fout.write(struct.pack("i", len(key)))
+    fout.write(key)
+
+for name in list_vars.keys():
+    data = list_vars[name].squeeze().numpy()
+    print("Processing variable: " , name ,  " with shape: ", data.shape)
+
+    # reshape conv bias from [n] to [n, 1]
+    if name in ["encoder.conv1.bias", "encoder.conv2.bias"]:
+        data = data.reshape(data.shape[0], 1)
+        print(f"  Reshaped variable: {name} to shape: ", data.shape)
+
+    n_dims = len(data.shape)
+
+    # looks like the whisper models are in f16 by default
+    # so we need to convert the small tensors to f32 until we fully support f16 in ggml
+    # ftype == 0 -> float32, ftype == 1 -> float16
+    ftype = 1
+    if use_f16:
+        if n_dims < 2 or \
+                name == "encoder.conv1.bias"   or \
+                name == "encoder.conv2.bias"   or \
+                name == "encoder.positional_embedding" or \
+                name == "decoder.positional_embedding":
+            print("  Converting to float32")
+            data = data.astype(np.float32)
+            ftype = 0
+    else:
+        data = data.astype(np.float32)
+        ftype = 0
+
+    #if name.startswith("encoder"):
+    #    if name.endswith("mlp.0.weight") or \
+    #       name.endswith("mlp.2.weight"):
+    #        print("  Transposing")
+    #        data = data.transpose()
+
+    # header
+    str_ = name.encode('utf-8')
+    fout.write(struct.pack("iii", n_dims, len(str_), ftype))
+    for i in range(n_dims):
+        fout.write(struct.pack("i", data.shape[n_dims - 1 - i]))
+    fout.write(str_)
+
+    # data
+    data.tofile(fout)
+
+fout.close()
+
+print("Done. Output file: " , fname_out)
+print("")