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

fixed blasbatchmul multiplier

Browse source
This commit is contained in:
Concedo 2023-08-08 00:41:02 +08:00
parent 28ad80b6e4
commit 3554080502
1 changed files with 7 additions and 2 deletions
Download patch file Download diff file Expand all files Collapse all files

9
llama.cpp
View file

@ -937,6 +937,11 @@ bool llama_mlock_supported() {
return llama_mlock::SUPPORTED;
}
int get_blas_batch_mul(int batch)
{
return (batch>512?(batch>1024?4:2):1);
}
void llama_backend_init(bool numa) {
ggml_time_init();
@ -1042,7 +1047,7 @@ static void llama_model_load_internal(
void * progress_callback_user_data) {
model.t_start_us = ggml_time_us();
size_t blasbatchmul = (n_batch>512?(n_batch>1024?4:2):1);
size_t blasbatchmul = get_blas_batch_mul(n_batch);
std::unique_ptr<llama_model_loader> ml(new llama_model_loader(fname, use_mmap));
@ -3234,7 +3239,7 @@ struct llama_context * llama_new_context_with_model(
params.seed = time(NULL);
}
size_t blasbatchmul = (n_batch>512?(n_batch>1024?4:2):1);
size_t blasbatchmul = get_blas_batch_mul(params.n_batch);
unsigned cur_percentage = 0;
if (params.progress_callback == NULL) {