vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

llama : don't zero-init vectors in quantize -> 5.1% faster

Browse source
This commit is contained in:
Cebtenzzre 2023-09-09 17:30:16 -04:00
parent a95aa21dad
commit f727ad5fc9
1 changed files with 10 additions and 4 deletions
Download patch file Download diff file Expand all files Collapse all files

14
llama.cpp
View file

@ -4639,8 +4639,14 @@ void llama_beam_search(llama_context * ctx,
// quantization // quantization
// //
template <typename T>
struct no_init {
T value;
no_init() { /* do nothing */ }
};
static void llama_convert_tensor_internal( static void llama_convert_tensor_internal(
struct ggml_tensor * tensor, std::vector<float> & output, std::vector<std::thread> & workers, struct ggml_tensor * tensor, std::vector<no_init<float>> & output, std::vector<std::thread> & workers,
const size_t nelements, const int nthread const size_t nelements, const int nthread
) { ) {
if (output.size() < nelements) { if (output.size() < nelements) {
@ -4895,9 +4901,9 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
int idx = 0; int idx = 0;
std::vector<uint8_t> read_data; std::vector<no_init<uint8_t>> read_data;
std::vector<uint8_t> work; std::vector<no_init<uint8_t>> work;
std::vector<float> f32_conv_buf; std::vector<no_init<float>> f32_conv_buf;
// populate the original tensors so we get an initial meta data // populate the original tensors so we get an initial meta data
for (int i = 0; i < ml->n_tensors; ++i) { for (int i = 0; i < ml->n_tensors; ++i) {