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

build : on Mac OS enable Metal by default (#2901)

Browse source 
* build : on Mac OS enable Metal by default

* make : try to fix build on Linux

* make : move targets back to the top

* make : fix target clean

* llama : enable GPU inference by default with Metal

* llama : fix vocab_only logic when GPU is enabled

* common : better `n_gpu_layers` assignment

* readme : update Metal instructions

* make : fix merge conflict remnants

* gitignore : metal
This commit is contained in:
Georgi Gerganov 2023-09-04 22:26:24 +03:00 • committed by GitHub
parent bd33e5ab92
commit e36ecdccc8
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
9 changed files with 143 additions and 133 deletions
Download patch file Download diff file Expand all files Collapse all files

54
llama.cpp
View file

@ -5340,7 +5340,7 @@ struct llama_context_params llama_context_default_params() {
/*.seed =*/ LLAMA_DEFAULT_SEED,
/*.n_ctx =*/ 512,
/*.n_batch =*/ 512,
/*.gpu_layers =*/ 0,
/*.n_gpu_layers =*/ 0,
/*.main_gpu =*/ 0,
/*.tensor_split =*/ nullptr,
/*.rope_freq_base =*/ 10000.0f,
@ -5357,6 +5357,10 @@ struct llama_context_params llama_context_default_params() {
/*.embedding =*/ false,
};
#ifdef GGML_USE_METAL
result.n_gpu_layers = 1;
#endif
return result;
}
@ -5549,43 +5553,43 @@ struct llama_context * llama_new_context_with_model(
}
#endif
}
}
#ifdef GGML_USE_METAL
if (params.n_gpu_layers > 0) {
// this allocates all Metal resources and memory buffers
if (params.n_gpu_layers > 0) {
// this allocates all Metal resources and memory buffers
void * data_ptr = NULL;
size_t data_size = 0;
void * data_ptr = NULL;
size_t data_size = 0;
if (params.use_mmap) {
data_ptr = ctx->model.mapping->addr;
data_size = ctx->model.mapping->size;
} else {
data_ptr = ggml_get_mem_buffer(ctx->model.ctx);
data_size = ggml_get_mem_size (ctx->model.ctx);
}
if (params.use_mmap) {
data_ptr = ctx->model.mapping->addr;
data_size = ctx->model.mapping->size;
} else {
data_ptr = ggml_get_mem_buffer(ctx->model.ctx);
data_size = ggml_get_mem_size (ctx->model.ctx);
}
const size_t max_size = ggml_get_max_tensor_size(ctx->model.ctx);
const size_t max_size = ggml_get_max_tensor_size(ctx->model.ctx);
LLAMA_LOG_INFO("%s: max tensor size = %8.2f MB\n", __func__, max_size/1024.0/1024.0);
LLAMA_LOG_INFO("%s: max tensor size = %8.2f MB\n", __func__, max_size/1024.0/1024.0);
#define LLAMA_METAL_CHECK_BUF(result) \
if (!(result)) { \
LLAMA_LOG_ERROR("%s: failed to add buffer\n", __func__); \
llama_free(ctx); \
return NULL; \
}
if (!(result)) { \
LLAMA_LOG_ERROR("%s: failed to add buffer\n", __func__); \
llama_free(ctx); \
return NULL; \
}
LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "data", data_ptr, data_size, max_size));
LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "data", data_ptr, data_size, max_size));
LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "eval", ctx->buf_compute.data, ctx->buf_compute.size, 0));
LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "kv", ctx->kv_self.buf.data, ctx->kv_self.buf.size, 0));
LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "eval", ctx->buf_compute.data, ctx->buf_compute.size, 0));
LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "kv", ctx->kv_self.buf.data, ctx->kv_self.buf.size, 0));
LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "alloc", ctx->buf_alloc.data, ctx->buf_alloc.size, 0));
LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "alloc", ctx->buf_alloc.data, ctx->buf_alloc.size, 0));
#undef LLAMA_METAL_CHECK_BUF
}
}
#endif
}
#ifdef GGML_USE_MPI
ctx->ctx_mpi = ggml_mpi_init();