CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-03 09:22:01 +00:00
Code Issues Packages Projects Releases Wiki Activity

llama : separate compute buffer for metal

Browse Source
This commit is contained in:
Georgi Gerganov
2023-07-18 19:19:59 +03:00
parent 652c849643
commit ed960fa1ab
1 changed files with 6 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
llama.cpp
View File

@@ -1616,7 +1616,6 @@ static bool llama_eval_internal(
LLAMA_ASSERT(lctx.graph_logits != nullptr);
// for big prompts, if BLAS is enabled, it is better to use only one thread
// otherwise, the threads are spin-lock waiting for the BLAS calls and are degrading the performance
n_threads = N >= 32 && ggml_cpu_has_blas() ? 1 : n_threads;
@@ -2719,11 +2718,17 @@ struct llama_context * llama_new_context_with_model(
// TODO: size the buffers more accurately - depends on improved memory management
ctx->buf_compute_cpu = ggml_backend_alloc_buffer(&model->backend_cpu, MEM_REQ_EVAL().at(ctx->model.type), 2048);
#ifdef GGML_USE_CUDA
if (params.n_gpu_layers > 0) {
ctx->buf_compute_cuda = ggml_backend_alloc_buffer(&model->backend_cuda, MEM_REQ_EVAL().at(ctx->model.type), 2048);
}
#endif
#ifdef GGML_USE_METAL
if (params.n_gpu_layers > 0) {
ctx->buf_compute_metal = ggml_backend_alloc_buffer(&model->backend_metal, MEM_REQ_EVAL().at(ctx->model.type), 2048);
}
#endif
// initialize the graph input/output buffers
// input buffer