Skip to content

llama : improve batched CPU perf with BLAS #4240

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 3 commits into from
Nov 28, 2023
Merged

llama : improve batched CPU perf with BLAS #4240

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
f815fe4
ggml : use blas even if src0 is not F32
ggerganov Nov 27, 2023
e9b7a5c
llama : use n_threads_batch only when n_tokens >= 32
ggerganov Nov 27, 2023
87f4102
llama : revert n_threads_batch logic
ggerganov Nov 27, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion ggml.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -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)) {

Expand Down
12 changes: 1 addition & 11 deletions llama.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -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;
}

Expand Down