Inference with FasterTransformer

FasterTransformer provides a script and recipe to run the highly optimized transformer-based encoder and decoder component, and it is tested and maintained by NVIDIA.

We adapted the GLM-130B based on Fastertransformer for fast inference, with details in benchmark section.

Setup

Requirements

CMake >= 3.13 for PyTorch
CUDA 11.0 or newer version
NCCL 2.10 or newer version
Python 3 is recommended because some features are not supported in python 2
PyTorch: Verify on 1.11.0, >= 1.8.0 should work.

All the packages can be installed using conda.

conda install -y cmake numpy pybind11 pytorch torchvision cudatoolkit-dev cudnn
cp -r $CONDA_PREFIX/lib/libcudnn* /usr/local/cuda/lib64/
cp -r $CONDA_PREFIX/include/cudnn*.h /usr/local/cuda/include/

GLM-130B is trained with FP16 precision, a total of 260G of GPU memory is required to store model weights. The model is tested with 8 * 40G A100s.

Build

Get the code and install all dependencies:

git clone https://github.com/THUDM/FasterTransformer.git
mkdir -p FasterTransformer/build
cd FasterTransformer/build
git submodule init && git submodule update
pip3 install fire jax jaxlib icetk

Note: the xx of -DSM=xx in following scripts means the compute capability of your GPU. For example, 60 (P40) or 61 (P4) or 70 (V100) or 75(T4) or 80 (A100). Default setting is including 70, 75, 80 and 86.

cmake -DSM=80 -DCMAKE_BUILD_TYPE=Release -DBUILD_PYT=ON -DBUILD_MULTI_GPU=ON ..
make -j

Download the Model

See Get Model.

The original checkpoint compatible with SAT, but each time the model is initialized it needs to be extracted, which costs time. So we provide a script FasterTransformer/examples/pytorch/glm/utils/glm_ckpt_convert.py to extract the downloaded checkpoint.

For example:

# convert SAT checkpoint to FT checkpoint
python3 ../examples/pytorch/glm/utils/glm_ckpt_convert.py -i global_step20000/iter_0020000 -o ft_output -i_g 8

Run GLM-130B

Generate the gemm_config.in file.

# ./bin/gpt_gemm <batch_size> <beam_width> <max_input_len> <head_number> <size_per_head> <inter_size> <vocab_size> <data_type> <tensor_para_size>
./bin/gpt_gemm 1 1 128 96 128 49152 150528 1 8

Running GLM_130B in Pytorch.

bash ../examples/pytorch/glm/benchmark-generation.sh

You need to check and edit this file to set arguments such as the checkpoint's load path.

When running GLM_130B, pay special attention to the following arguments:

--sat-ckpt-dir is the path to the original downloaded checkpoint, compatible with SwissArmyTransformer.
--ft-ckpt-dir is the path to the extracted checkpoint. It is faster to load, but you have to run examples/pytorch/glm/utils/glm_ckpt_convert.py to convert the downloaded checkpoint.
--n-inference-gpus number of GPUs used for inference, defaults to 8. The binary model parameters are saved to ${output-dir}/${n-inference-gpus}-gpu/
--sample-input-file everyline is a batch, you can set max_batch_size to get multiple generations at one time, however, you need to ensure that all inputs are of the same length after being converted to tokens, otherwise only the longest sentence will get a right output.

Optimization methods

Optimization in GLM_130B are similar to optimization in GPT and GPT-J, describing in the FasterTransformer/gpt_guide.md. Meanwhile, some of the operators are differ from GPT, such as the implementation of RotaryEmbedding, and the use of GeGLU, so we add them additionally into FasterTransformer.

Benchmark

Hardware: DGX-A100(8 * 40G)

Encode

Sequence Len	512	1024	2048
Megatron	145 ms	250 ms	453 ms
FasterTransformer	120 ms	220 ms	OOM

Decode

Sequence Len	512	1024	2048
Megatron	45.21 s	89.00 s	179.22 s
FasterTransformer	18.77 s	39.81 s	89.88 s

inference-with-fastertransformer.md 4.2 KB Riwayat Mentahan