seamless_communication/ggml/examples/unity/unity_model_loader.h

// Copyright (c) Meta Platforms, Inc. and affiliates.
// All rights reserved.
//
// This source code is licensed under the license found in the
// LICENSE file in the root directory of this source tree.

#pragma once

#include <vector>
#include "model_loader.h"


// TODO Merge with Ning implementation
struct unity_hparams {
    std::int64_t model_dim;
    std::int64_t w2v2_encoder_config__model_dim;
    std::int64_t w2v2_encoder_config__max_seq_len;
    std::int64_t w2v2_encoder_config__feature_dim;
    std::int64_t w2v2_encoder_config__use_fbank;
    float w2v2_encoder_config__first_pass_dropout_p;
    std::int64_t w2v2_encoder_config__layer_norm_features;
    // Error: Unsupported type <class 'list'> w2v2_encoder_config__feature_extractor_layer_descs;
    std::int64_t w2v2_encoder_config__feature_extractor_bias;
    std::int64_t w2v2_encoder_config__feature_extractor_layer_norm_convs;
    std::int64_t w2v2_encoder_config__feature_grad_scale;
    std::int64_t w2v2_encoder_config__num_fbank_channels;
    std::int64_t w2v2_encoder_config__fbank_stride;
    std::int64_t w2v2_encoder_config__sample_fbank_every_k;
    // Error: Unsupported type <class 'str'> w2v2_encoder_config__pos_encoder_type;
    std::int64_t w2v2_encoder_config__pos_encoder_depth;
    std::int64_t w2v2_encoder_config__pos_conv_kernel_size;
    std::int64_t w2v2_encoder_config__num_pos_conv_groups;
    std::int64_t w2v2_encoder_config__use_conformer;
    std::int64_t w2v2_encoder_config__num_encoder_layers;
    std::int64_t w2v2_encoder_config__num_encoder_attn_heads;
    std::int64_t w2v2_encoder_config__ffn_inner_dim;
    float w2v2_encoder_config__dropout_p;
    float w2v2_encoder_config__attn_dropout_p;
    float w2v2_encoder_config__layer_drop_p;
    std::int32_t w2v2_encoder_config__norm_order;
    std::int64_t w2v2_encoder_config__depthwise_conv_kernel_size;
    std::int64_t nllb_config__model_dim;
    std::int64_t nllb_config__max_seq_len;
    std::int64_t nllb_config__vocabulary_size;
    std::int64_t nllb_config__pad_idx;
    std::int64_t nllb_config__num_encoder_layers;
    std::int64_t nllb_config__num_decoder_layers;
    std::int64_t nllb_config__num_encoder_attn_heads;
    std::int64_t nllb_config__num_decoder_attn_heads;
    std::int64_t nllb_config__ffn_inner_dim;
    float nllb_config__dropout_p;
    std::int64_t t2u_config__model_dim;
    std::int64_t t2u_config__unit_max_seq_len;
    std::int64_t t2u_config__unit_vocabulary_size;
    std::int64_t t2u_config__unit_pad_idx;
    std::int64_t t2u_config__num_encoder_layers;
    std::int64_t t2u_config__num_decoder_layers;
    std::int64_t t2u_config__num_encoder_attn_heads;
    std::int64_t t2u_config__num_decoder_attn_heads;
    std::int64_t t2u_config__ffn_inner_dim;
    float t2u_config__dropout_p;
    std::int64_t use_text_encoder;
    std::int64_t use_conformer_adaptor;
    std::int64_t num_adaptor_layers;
    std::int64_t adaptor_kernel_size;
    std::int64_t adaptor_stride;
    std::int64_t adaptor_layer_norm;
    float adaptor_dropout_p;
    std::int64_t model_byte_size;
    std::int64_t __end_of_hparams__;

};

void read_unity_hparams(unity_hparams* out, std::ifstream &fin) {
    fin.read((char*) &out->model_dim, sizeof(out->model_dim));
    fin.read((char*) &out->w2v2_encoder_config__model_dim, sizeof(out->w2v2_encoder_config__model_dim));
    fin.read((char*) &out->w2v2_encoder_config__max_seq_len, sizeof(out->w2v2_encoder_config__max_seq_len));
    fin.read((char*) &out->w2v2_encoder_config__feature_dim, sizeof(out->w2v2_encoder_config__feature_dim));
    fin.read((char*) &out->w2v2_encoder_config__use_fbank, sizeof(out->w2v2_encoder_config__use_fbank));
    fin.read((char*) &out->w2v2_encoder_config__first_pass_dropout_p, sizeof(out->w2v2_encoder_config__first_pass_dropout_p));
    fin.read((char*) &out->w2v2_encoder_config__layer_norm_features, sizeof(out->w2v2_encoder_config__layer_norm_features));
    fin.read((char*) &out->w2v2_encoder_config__feature_extractor_bias, sizeof(out->w2v2_encoder_config__feature_extractor_bias));
    fin.read((char*) &out->w2v2_encoder_config__feature_extractor_layer_norm_convs, sizeof(out->w2v2_encoder_config__feature_extractor_layer_norm_convs));
    fin.read((char*) &out->w2v2_encoder_config__feature_grad_scale, sizeof(out->w2v2_encoder_config__feature_grad_scale));
    fin.read((char*) &out->w2v2_encoder_config__num_fbank_channels, sizeof(out->w2v2_encoder_config__num_fbank_channels));
    fin.read((char*) &out->w2v2_encoder_config__fbank_stride, sizeof(out->w2v2_encoder_config__fbank_stride));
    fin.read((char*) &out->w2v2_encoder_config__sample_fbank_every_k, sizeof(out->w2v2_encoder_config__sample_fbank_every_k));
    fin.read((char*) &out->w2v2_encoder_config__pos_encoder_depth, sizeof(out->w2v2_encoder_config__pos_encoder_depth));
    fin.read((char*) &out->w2v2_encoder_config__pos_conv_kernel_size, sizeof(out->w2v2_encoder_config__pos_conv_kernel_size));
    fin.read((char*) &out->w2v2_encoder_config__num_pos_conv_groups, sizeof(out->w2v2_encoder_config__num_pos_conv_groups));
    fin.read((char*) &out->w2v2_encoder_config__use_conformer, sizeof(out->w2v2_encoder_config__use_conformer));
    fin.read((char*) &out->w2v2_encoder_config__num_encoder_layers, sizeof(out->w2v2_encoder_config__num_encoder_layers));
    fin.read((char*) &out->w2v2_encoder_config__num_encoder_attn_heads, sizeof(out->w2v2_encoder_config__num_encoder_attn_heads));
    fin.read((char*) &out->w2v2_encoder_config__ffn_inner_dim, sizeof(out->w2v2_encoder_config__ffn_inner_dim));
    fin.read((char*) &out->w2v2_encoder_config__dropout_p, sizeof(out->w2v2_encoder_config__dropout_p));
    fin.read((char*) &out->w2v2_encoder_config__attn_dropout_p, sizeof(out->w2v2_encoder_config__attn_dropout_p));
    fin.read((char*) &out->w2v2_encoder_config__layer_drop_p, sizeof(out->w2v2_encoder_config__layer_drop_p));
    fin.read((char*) &out->w2v2_encoder_config__norm_order, sizeof(out->w2v2_encoder_config__norm_order));
    fin.read((char*) &out->w2v2_encoder_config__depthwise_conv_kernel_size, sizeof(out->w2v2_encoder_config__depthwise_conv_kernel_size));
    fin.read((char*) &out->nllb_config__model_dim, sizeof(out->nllb_config__model_dim));
    fin.read((char*) &out->nllb_config__max_seq_len, sizeof(out->nllb_config__max_seq_len));
    fin.read((char*) &out->nllb_config__vocabulary_size, sizeof(out->nllb_config__vocabulary_size));
    fin.read((char*) &out->nllb_config__pad_idx, sizeof(out->nllb_config__pad_idx));
    fin.read((char*) &out->nllb_config__num_encoder_layers, sizeof(out->nllb_config__num_encoder_layers));
    fin.read((char*) &out->nllb_config__num_decoder_layers, sizeof(out->nllb_config__num_decoder_layers));
    fin.read((char*) &out->nllb_config__num_encoder_attn_heads, sizeof(out->nllb_config__num_encoder_attn_heads));
    fin.read((char*) &out->nllb_config__num_decoder_attn_heads, sizeof(out->nllb_config__num_decoder_attn_heads));
    fin.read((char*) &out->nllb_config__ffn_inner_dim, sizeof(out->nllb_config__ffn_inner_dim));
    fin.read((char*) &out->nllb_config__dropout_p, sizeof(out->nllb_config__dropout_p));
    fin.read((char*) &out->t2u_config__model_dim, sizeof(out->t2u_config__model_dim));
    fin.read((char*) &out->t2u_config__unit_max_seq_len, sizeof(out->t2u_config__unit_max_seq_len));
    fin.read((char*) &out->t2u_config__unit_vocabulary_size, sizeof(out->t2u_config__unit_vocabulary_size));
    fin.read((char*) &out->t2u_config__unit_pad_idx, sizeof(out->t2u_config__unit_pad_idx));
    fin.read((char*) &out->t2u_config__num_encoder_layers, sizeof(out->t2u_config__num_encoder_layers));
    fin.read((char*) &out->t2u_config__num_decoder_layers, sizeof(out->t2u_config__num_decoder_layers));
    fin.read((char*) &out->t2u_config__num_encoder_attn_heads, sizeof(out->t2u_config__num_encoder_attn_heads));
    fin.read((char*) &out->t2u_config__num_decoder_attn_heads, sizeof(out->t2u_config__num_decoder_attn_heads));
    fin.read((char*) &out->t2u_config__ffn_inner_dim, sizeof(out->t2u_config__ffn_inner_dim));
    fin.read((char*) &out->t2u_config__dropout_p, sizeof(out->t2u_config__dropout_p));
    fin.read((char*) &out->use_text_encoder, sizeof(out->use_text_encoder));
    fin.read((char*) &out->use_conformer_adaptor, sizeof(out->use_conformer_adaptor));
    fin.read((char*) &out->num_adaptor_layers, sizeof(out->num_adaptor_layers));
    fin.read((char*) &out->adaptor_kernel_size, sizeof(out->adaptor_kernel_size));
    fin.read((char*) &out->adaptor_stride, sizeof(out->adaptor_stride));
    fin.read((char*) &out->adaptor_layer_norm, sizeof(out->adaptor_layer_norm));
    fin.read((char*) &out->adaptor_dropout_p, sizeof(out->adaptor_dropout_p));
    fin.read((char*) &out->model_byte_size, sizeof(out->model_byte_size));
    fin.read((char*) &out->__end_of_hparams__, sizeof(out->__end_of_hparams__));

};





// Embedding
std::size_t compute_embed_size(int32_t vocab_size, int32_t dim)
{
    return vocab_size * dim * ggml_type_size(GGML_TYPE_F32);
};

// Attention Layer

struct attention_layer {
    struct ggml_tensor* layer_norm_w; // model_dim
    struct ggml_tensor* layer_norm_b; // model_dim

    struct ggml_tensor* q_proj_w; // model_dim x model_dim
    struct ggml_tensor* q_proj_b; // model_dim
    struct ggml_tensor* k_proj_w; // model_dim x model_dim
    struct ggml_tensor* k_proj_b; // model_dim
    struct ggml_tensor* v_proj_w; // model_dim x model_dim
    struct ggml_tensor* v_proj_b; // model_dim

    struct ggml_tensor* output_proj_w; // model_dim x model_dim
    struct ggml_tensor* output_proj_b; // model_dim
};

std::size_t compute_attention_layer_size(int32_t dim)
{
    return LayerNorm_size(dim)
        + 4 * Linear_size(dim, dim); // q, k, v, and out
};

void init_attention_layer(
    attention_layer *layer,
    fairseq2_model &model_ctx,
    const std::string &prefix)
{
    auto hparams = (unity_hparams&)model_ctx.hparams;
    const auto dim = hparams.nllb_config__model_dim;
    auto ctx = model_ctx.ctx;
    auto &tensor_map = model_ctx.tensors;

    layer->layer_norm_w = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, dim);
    tensor_map[prefix + "_layer_norm.weight"] = layer->layer_norm_w;
    layer->layer_norm_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, dim);
    tensor_map[prefix + "_layer_norm.bias"] = layer->layer_norm_b;

    layer->q_proj_w = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, dim, dim);
    tensor_map[prefix + ".q_proj.weight"] = layer->q_proj_w;
    layer->q_proj_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, dim);
    tensor_map[prefix + ".q_proj.bias"] = layer->q_proj_b;

    layer->k_proj_w = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, dim, dim);
    tensor_map[prefix + ".k_proj.weight"] = layer->k_proj_w;
    layer->k_proj_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, dim);
    tensor_map[prefix + ".k_proj.bias"] = layer->k_proj_b;

    layer->v_proj_w = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, dim, dim);
    tensor_map[prefix + ".v_proj.weight"] = layer->v_proj_w;
    layer->v_proj_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, dim);
    tensor_map[prefix + ".v_proj.bias"] = layer->v_proj_b;

    layer->output_proj_w = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, dim, dim);
    tensor_map[prefix + ".output_proj.weight"] = layer->output_proj_w;
    layer->output_proj_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, dim);
    tensor_map[prefix + ".output_proj.bias"] = layer->output_proj_b;
}


// Attention Head

struct attention_head {
    struct attention_layer* self_attn; // model_dim
    struct attention_layer* encoder_decoder_attn; // model_dim
    struct StandardFeedForwardNetwork* ffn;
};

std::size_t compute_attention_head_size(int32_t dim, int32_t inner_dim)
{
    return 2 * compute_attention_layer_size(dim) + StandardFeedForwardNetwork_size(dim, inner_dim);
};

void init_attention_head(
    attention_head *head,
    fairseq2_model &model_ctx,
    const std::string &prefix)
{
    auto hparams = (unity_hparams&)model_ctx.hparams;
    init_attention_layer(head->self_attn, model_ctx, prefix + ".self_attn");
    init_attention_layer(head->encoder_decoder_attn, model_ctx, prefix + ".encoder_decoder_attn");
    StandardFeedForwardNetwork_init((StandardFeedForwardNetwork&)(head->ffn), model_ctx, prefix + ".ffn", hparams.nllb_config__model_dim, hparams.nllb_config__ffn_inner_dim);
}

// TODO: attention_head_compute_graph

// Text Decoder

struct text_decoder {
    struct ggml_tensor* frontend_embed_w; // vocab_size x model_dim
    std::vector<attention_head*> multi_head;
    struct ggml_tensor* layer_norm_w;
    struct ggml_tensor* layer_norm_b;
};

std::size_t compute_context_size(void* raw_hparams)
{
    auto hparams = (unity_hparams&)raw_hparams;
    const auto vocab_size = hparams.nllb_config__vocabulary_size;
    const auto dim = hparams.nllb_config__model_dim;
    const auto inner_dim = hparams.nllb_config__ffn_inner_dim;
    const auto n_layers = hparams.nllb_config__num_decoder_layers;

    const auto overhead = (6 + 12 * n_layers) * 512; // TODO Find out what this is.

    return compute_embed_size(vocab_size, dim)
        + n_layers * compute_attention_head_size(dim, inner_dim)
        + LayerNorm_size(dim)
        + overhead;
};

class unity_model_loader: public model_loader {
    public:
    void load_hparams(fairseq2_model& model, std::ifstream &fin);

    std::size_t compute_context_size(void* raw_hparams);

    void tensors_alloc(fairseq2_model &model);
};