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

metal : reduce command encoding overhead (#9698)

Browse source 
* metal : reduce command encoding overhead

ggml-ci

* metal : add comments
This commit is contained in:
Georgi Gerganov 2024-10-01 16:00:25 +03:00 • committed by GitHub
parent a90484c6d9
commit cad341d889
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
5 changed files with 2000 additions and 1912 deletions
Download patch file Download diff file Expand all files Collapse all files

7
examples/cvector-generator/pca.hpp
View file

@ -204,13 +204,6 @@ static ggml_status compute_piter(
ggml_backend_cpu_set_n_threads(model.backend, params.n_threads);
}
// TODO: enable GPU support when support for GGML_OP_SQRT is added
//#ifdef GGML_USE_METAL
// if (ggml_backend_is_metal(model.backend)) {
// ggml_backend_metal_set_n_cb(model.backend, params.n_threads);
// }
//#endif
ggml_status res = ggml_backend_graph_compute(model.backend, gf);
if (res == GGML_STATUS_SUCCESS) {
auto extract_i = [](std::string prefix, std::string str) -> int {

6
examples/llava/clip.cpp
View file

@ -2444,12 +2444,6 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
ggml_backend_cpu_set_n_threads(ctx->backend, n_threads);
}
#ifdef GGML_USE_METAL
if (ggml_backend_is_metal(ctx->backend)) {
ggml_backend_metal_set_n_cb(ctx->backend, n_threads);
}
#endif
ggml_backend_graph_compute(ctx->backend, gf);
// the last node is the embedding tensor