Xenith/xenith/converter.cpp

#include "converter.hpp"

#include <cstring>
#include <iostream>

bool save_to_bfr(const std::string& filename, const Model& model)
{
    std::ofstream out(filename, std::ios::binary);
    if (!out.is_open()) {
        std::cerr << "Критическая ошибка! Не удалось создать файл: " << filename << std::endl;
        std::cerr << "Причина от ОС: " << std::strerror(errno) << " (код ошибки: " << errno << ")" << std::endl;
        return false;
    }
    out.write(reinterpret_cast<const char*>(&model.config), sizeof(ModelConfig));
    uint32_t map_size = static_cast<uint32_t>(model.vocab.size());
    out.write(reinterpret_cast<const char*>(&map_size), sizeof(map_size));
    for (const auto& [token, id] : model.vocab)
    {
        out.write(reinterpret_cast<const char*>(&id), sizeof(id));
        uint32_t token_len = static_cast<uint32_t>(token.length());
        out.write(reinterpret_cast<const char*>(&token_len), sizeof(token_len));
        out.write(token.data(), token_len);
    }
    size_t w_emb_bytes = model.w_emb.size() * sizeof(float);
    out.write(reinterpret_cast<const char*>(model.w_emb.data()), w_emb_bytes);
    size_t w_qkv_bytes = model.w_qkv.size() * sizeof(float);
    out.write(reinterpret_cast<const char*>(model.w_qkv.data()), w_qkv_bytes);
    size_t w_o_bytes = model.w_o.size() * sizeof(float);
    out.write(reinterpret_cast<const char*>(model.w_emb.data()), w_o_bytes);
    return true;
}

Model load_from_bfr(const std::string& filename)
{
    Model mdl;

    mdl.fd = fopen(filename.c_str(), "rb");
    if (!mdl.fd) {
        perror("fopen failed");
        return mdl;
    }

    size_t items_read = fread(&mdl.config, sizeof(ModelConfig), 1, mdl.fd);

    if (items_read != 1) {
        std::cerr << "Ошибка: прочитано " << items_read
                  << " структур вместо 1\n";
        fclose(mdl.fd);
        mdl.fd = nullptr;
        return mdl;
    }

    uint32_t map_size;
    if (fread(&map_size, sizeof(map_size), 1, mdl.fd))
    {

    } else
    {
        std::cerr << "Ошибка чтения файла" << std::endl;
        return mdl;
    }

    if (map_size != mdl.config.vocab_size) {
        std::cerr << "Warning: vocab size mismatch! "
                  << "Header says " << mdl.config.vocab_size
                  << ", file has " << map_size << "\n";
    }

    mdl.vocab.reserve(map_size);

    for (uint32_t i = 0; i < map_size; ++i) {
        uint32_t id, token_len;
        fread(&id, sizeof(id), 1, mdl.fd);
        fread(&token_len, sizeof(token_len), 1, mdl.fd);
        std::string token(token_len, '\0');
        fread(&token[0], 1, token_len, mdl.fd);
        mdl.vocab.emplace(std::move(token), id);
    }

    mdl.offset_emb = ftell(mdl.fd);
    mdl.offset_qkv = mdl.offset_emb + (mdl.config.vocab_size * mdl.config.embedding_dim * sizeof(float));
    mdl.offset_o = mdl.offset_qkv + (mdl.config.num_layers * mdl.config.embedding_dim * 3 * mdl.config.num_heads * sizeof(float));

    return mdl;
}

void Model::close_fd()
{
    if (fd != nullptr)
    {
        fclose(fd);
        fd = nullptr;
    }
}

void Model::load(size_t num, std::vector<float>& target, uint64_t offset)
{
    target.resize(num);

    if (fseek(fd, offset, SEEK_SET) != 0) {
        perror("fseek failed");
        target.clear();
    }

    size_t bt_read = fread(target.data(), 1, num * sizeof(float), fd);

    if (bt_read != num * sizeof(float)) {
        std::cerr << "Read error: expected " << num * sizeof(float)
                  << " bytes, got " << bt_read << "\n";
        target.clear();
    }
}

void Model::load_edm()
{
    load(
        config.vocab_size * config.embedding_dim,
        w_emb,
        offset_emb
    );
}
void Model::load_qkv()
{
    load(
        config.num_layers * config.embedding_dim * 3 * config.num_heads,
        w_qkv,
        offset_qkv
    );
}
void Model::load_o()
{
    load(
        config.num_layers * (config.num_heads * config.head_dim) * config.embedding_dim,
        w_o,
        offset_o
    );
}

void Model::free_edm() { w_emb.resize(0); w_emb.shrink_to_fit(); }
void Model::free_qkv() { w_qkv.resize(0); w_qkv.shrink_to_fit(); }
void Model::free_o() { w_o.resize(0); w_o.shrink_to_fit(); }
void Model::allocate_kv_cache() {
    size_t total_elements =
        (size_t)config.max_ctx *
        config.num_layers *
        2 *
        config.num_heads *
        config.head_dim;

    kv_cache = (float*)malloc(total_elements * sizeof(float));

    if (!kv_cache) {
        perror("Allocation failed");
        return;
    }

    memset(kv_cache, 0, total_elements);

    ctx_len = 0;
}
void Model::free_kv_cache()
{
    free(kv_cache);
    ctx_len = 0;
}
float* Model::get_kv(uint64_t token, uint32_t layer, cache_type kv, uint32_t head, uint32_t dim)
{
    size_t head_stride = config.head_dim;
    size_t kv_stride = config.num_heads * config.head_dim;
    size_t layer_stride = 2 * kv_stride;
    size_t token_stride = config.num_layers * layer_stride;

    size_t offset = (token * token_stride) +
                     (layer * layer_stride) +
                     (static_cast<int>(kv) * kv_stride) +
                     (head * head_stride) +
                     dim;
    return kv_cache + offset;
}