Compare commits

...
Sign in to create a new pull request.

3 commits

Author SHA1 Message Date
Georgi Gerganov
87f4102a70
llama : revert n_threads_batch logic
ggml-ci
2023-11-27 21:47:35 +02:00
Georgi Gerganov
e9b7a5cbd0
llama : use n_threads_batch only when n_tokens >= 32
ggml-ci
2023-11-27 20:49:17 +02:00
Georgi Gerganov
f815fe43d3
ggml : use blas even if src0 is not F32 2023-11-27 20:48:27 +02:00
2 changed files with 2 additions and 12 deletions

2
ggml.c
View file

@ -9373,7 +9373,7 @@ static bool ggml_compute_forward_mul_mat_use_blas(
// TODO: find the optimal values for these
if (ggml_is_contiguous(src0) &&
ggml_is_contiguous(src1) &&
src0->type == GGML_TYPE_F32 &&
//src0->type == GGML_TYPE_F32 &&
src1->type == GGML_TYPE_F32 &&
(ne0 >= 32 && ne1 >= 32 && ne10 >= 32)) {

View file

@ -5550,18 +5550,8 @@ static int llama_decode_internal(
n_threads = std::min(4, n_threads);
}
// If all tensors can be run on the GPU then using more than 1 thread is detrimental.
const bool full_offload_supported =
model.arch == LLM_ARCH_LLAMA ||
model.arch == LLM_ARCH_BAICHUAN ||
model.arch == LLM_ARCH_FALCON ||
model.arch == LLM_ARCH_REFACT ||
model.arch == LLM_ARCH_MPT ||
model.arch == LLM_ARCH_STARCODER ||
model.arch == LLM_ARCH_STABLELM;
const bool fully_offloaded = model.n_gpu_layers >= (int) hparams.n_layer + 3;
if (ggml_cpu_has_cublas() && full_offload_supported && fully_offloaded) {
if (ggml_cpu_has_cublas() && fully_offloaded) {
n_threads = 1;
}