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

enable CPU HBM (#2603)

Browse source 
* add cpu hbm support

* add memalign 0 byte check

* Update ggml.c

* Update llama.cpp

* ggml : allow ggml_init with 0 size

* retrigger ci

* fix code style

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
This commit is contained in:
Kunshang Ji 2023-09-08 09:46:56 +08:00 • committed by GitHub
parent 6336d834ec
commit 7f412dab9c
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
3 changed files with 38 additions and 2 deletions
Download patch file Download diff file Expand all files Collapse all files

12
llama.cpp
View file

@ -126,6 +126,9 @@ void replace_all(std::string & s, const std::string & search, const std::string
}
s = std::move(result);
}
#ifdef GGML_USE_CPU_HBM
#include <hbwmalloc.h>
#endif
static void zeros(std::ofstream & file, size_t n) {
char zero = 0;
@ -450,6 +453,9 @@ static void ggml_graph_compute_helper(std::vector<uint8_t> & buf, ggml_cgraph *
#elif GGML_USE_METAL
# define llama_host_malloc(n) ggml_metal_host_malloc(n)
# define llama_host_free(data) ggml_metal_host_free(data)
#elif GGML_USE_CPU_HBM
# define llama_host_malloc(n) hbw_malloc(n)
# define llama_host_free(data) if (data != NULL) hbw_free(data)
#else
# define llama_host_malloc(n) malloc(n)
# define llama_host_free(data) free(data)
@ -1489,7 +1495,11 @@ struct llama_model_loader {
// allocate temp buffer if not using mmap
if (!use_mmap && cur->data == NULL) {
GGML_ASSERT(cur->backend != GGML_BACKEND_CPU);
cur->data = malloc(ggml_nbytes(cur));
#ifdef GGML_USE_CPU_HBM
cur->data = (uint8_t*)hbw_malloc(ggml_nbytes(cur));
#else
cur->data = (uint8_t*)malloc(ggml_nbytes(cur));
#endif
}
load_data_for(cur);