Function llama_convert_tensor_internal

llama.cpp:8959–9015

};

static void llama_convert_tensor_internal(
    struct ggml_tensor * tensor, std::vector<no_init<float>> & output, std::vector<std::thread> & workers,
    const size_t nelements, const int nthread
) {
    if (output.size() < nelements) {
        output.resize(nelements);
    }
    float * f32_output = (float *) output.data();

    ggml_type_traits_t qtype;
    if (ggml_is_quantized(tensor->type)) {
        qtype = ggml_internal_get_type_traits(tensor->type);
        if (qtype.to_float == NULL) {
            throw std::runtime_error(format("type %s unsupported for integer quantization: no dequantization available", ggml_type_name(tensor->type)));
        }
    } else if (tensor->type != GGML_TYPE_F16) {
        throw std::runtime_error(format("cannot dequantize/convert tensor type %s", ggml_type_name(tensor->type)));
    }

    if (nthread < 2) {
        if (tensor->type == GGML_TYPE_F16) {
            ggml_fp16_to_fp32_row((ggml_fp16_t *)tensor->data, f32_output, nelements);
        } else if (ggml_is_quantized(tensor->type)) {
            qtype.to_float(tensor->data, f32_output, nelements);
        } else {
            GGML_ASSERT(false); // unreachable
        }
        return;
    }

    auto block_size = tensor->type == GGML_TYPE_F16 ? 1 : (size_t)ggml_blck_size(tensor->type);
    auto block_size_bytes = ggml_type_size(tensor->type);

    GGML_ASSERT(nelements % block_size == 0);
    auto nblocks = nelements / block_size;
    auto blocks_per_thread = nblocks / nthread;
    auto spare_blocks = nblocks - (blocks_per_thread * nthread); // if blocks aren't divisible by thread count

    for (auto tnum = 0, in_buff_offs = 0, out_buff_offs = 0; tnum < nthread; tnum++) {
        auto thr_blocks = blocks_per_thread + (tnum == nthread - 1 ? spare_blocks : 0); // num blocks for this thread
        auto thr_elems = thr_blocks * block_size; // number of elements for this thread
        auto thr_block_bytes = thr_blocks * block_size_bytes; // number of input bytes for this thread

        auto compute = [qtype] (ggml_type typ, uint8_t * inbuf, float * outbuf, int nels) {
            if (typ == GGML_TYPE_F16) {
                ggml_fp16_to_fp32_row((ggml_fp16_t *)inbuf, outbuf, nels);
            } else {
                qtype.to_float(inbuf, outbuf, nels);
            }
        };
        workers.emplace_back(compute, tensor->type, (uint8_t *) tensor->data + in_buff_offs, f32_output + out_buff_offs, thr_elems);
        in_buff_offs += thr_block_bytes;
        out_buff_offs += thr_elems;
    }
    for (auto & w : workers) { w.join(); }
    workers.clear();
}