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

fix for yr-rocm, large gpu scratch

Browse source
This commit is contained in:
Concedo 2023-06-30 12:40:08 +08:00
parent 1347d3acc0
commit 86469d15c4
6 changed files with 16 additions and 7 deletions
Download patch file Download diff file Expand all files Collapse all files

5
llama.cpp
View file

@ -12,7 +12,8 @@
#include "ggml.h" #include "ggml.h"
#ifdef GGML_USE_CUBLAS #ifdef GGML_USE_CUBLAS
#include "ggml-cuda.h" #include "ggml-cuda.h"
#elif defined(GGML_USE_CLBLAST) #endif
#if defined(GGML_USE_CLBLAST)
#include "ggml-opencl.h" #include "ggml-opencl.h"
#endif #endif
@ -1113,7 +1114,7 @@ static void llama_model_load_internal(
fprintf(stderr, "%s: not allocating a VRAM scratch buffer due to low VRAM option\n", __func__); fprintf(stderr, "%s: not allocating a VRAM scratch buffer due to low VRAM option\n", __func__);
ggml_cuda_set_scratch_size(0); // disable scratch ggml_cuda_set_scratch_size(0); // disable scratch
} else { } else {
vram_scratch = n_batch * MB; vram_scratch = n_batch * MB * bigctxmul;
ggml_cuda_set_scratch_size(vram_scratch); ggml_cuda_set_scratch_size(vram_scratch);
if (n_gpu_layers > 0) { if (n_gpu_layers > 0) {
fprintf(stderr, "%s: allocating batch_size x 1 MB = %zd MB VRAM for the scratch buffer\n", fprintf(stderr, "%s: allocating batch_size x 1 MB = %zd MB VRAM for the scratch buffer\n",

4
otherarch/gpt2_v3.cpp
View file

@ -18,10 +18,12 @@
#ifdef GGML_USE_CUBLAS #ifdef GGML_USE_CUBLAS
#include "ggml-cuda.h" #include "ggml-cuda.h"
#elif defined(GGML_USE_CLBLAST) #endif
#if defined(GGML_USE_CLBLAST)
#include "ggml-opencl.h" #include "ggml-opencl.h"
#endif #endif
// load the model's weights from a file // load the model's weights from a file
ModelLoadResult gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab & vocab, FileFormat file_format, int gpulayers) { ModelLoadResult gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab & vocab, FileFormat file_format, int gpulayers) {
printf("%s: loading model from '%s'\n", __func__, fname.c_str()); printf("%s: loading model from '%s'\n", __func__, fname.c_str());

3
otherarch/gptj_v3.cpp
View file

@ -18,7 +18,8 @@
#ifdef GGML_USE_CUBLAS #ifdef GGML_USE_CUBLAS
#include "ggml-cuda.h" #include "ggml-cuda.h"
#elif defined(GGML_USE_CLBLAST) #endif
#if defined(GGML_USE_CLBLAST)
#include "ggml-opencl.h" #include "ggml-opencl.h"
#endif #endif

5
otherarch/llama_v2.cpp
View file

@ -9,12 +9,15 @@
#include "llama_v2.h" #include "llama_v2.h"
#include "ggml_v2.h" #include "ggml_v2.h"
#ifdef GGML_USE_CUBLAS #ifdef GGML_USE_CUBLAS
#include "ggml_v2-cuda.h" #include "ggml_v2-cuda.h"
#elif defined(GGML_USE_CLBLAST) #endif
#if defined(GGML_USE_CLBLAST)
#include "ggml_v2-opencl.h" #include "ggml_v2-opencl.h"
#endif #endif
#include <array> #include <array>
#include <ctime> #include <ctime>
#include <cinttypes> #include <cinttypes>

3
otherarch/mpt_v3.cpp
View file

@ -18,7 +18,8 @@
#ifdef GGML_USE_CUBLAS #ifdef GGML_USE_CUBLAS
#include "ggml-cuda.h" #include "ggml-cuda.h"
#elif defined(GGML_USE_CLBLAST) #endif
#if defined(GGML_USE_CLBLAST)
#include "ggml-opencl.h" #include "ggml-opencl.h"
#endif #endif

3
otherarch/neox_v3.cpp
View file

@ -16,7 +16,8 @@
#ifdef GGML_USE_CUBLAS #ifdef GGML_USE_CUBLAS
#include "ggml-cuda.h" #include "ggml-cuda.h"
#elif defined(GGML_USE_CLBLAST) #endif
#if defined(GGML_USE_CLBLAST)
#include "ggml-opencl.h" #include "ggml-opencl.h"
#endif #endif