root/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-12-25 02:44:36 +00:00
Code Issues Actions 9 Packages Projects Releases Wiki Activity

Add an option to build without CUDA VMM (#7067)

Browse Source 
Add an option to build ggml cuda without CUDA VMM
resolves
https://github.com/ggerganov/llama.cpp/issues/6889
https://forums.developer.nvidia.com/t/potential-nvshmem-allocated-memory-performance-issue/275416/4
This commit is contained in:
William Tambellini 2024-05-06 11:12:14 -07:00 committed by GitHub
parent b3a995b416
commit 858f6b73f6
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 13 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

11
CMakeLists.txt
View File

@ -103,6 +103,8 @@ set(LLAMA_CUDA_KQUANTS_ITER "2" CACHE STRING "llama: iters./thread per block for
set(LLAMA_CUDA_PEER_MAX_BATCH_SIZE "128" CACHE STRING set(LLAMA_CUDA_PEER_MAX_BATCH_SIZE "128" CACHE STRING
"llama: max. batch size for using peer access") "llama: max. batch size for using peer access")
option(LLAMA_CUDA_NO_PEER_COPY "llama: do not use peer to peer copies" OFF) option(LLAMA_CUDA_NO_PEER_COPY "llama: do not use peer to peer copies" OFF)
option(LLAMA_CUDA_NO_VMM "llama: do not try to use CUDA VMM" OFF)
option(LLAMA_CURL "llama: use libcurl to download model from an URL" OFF) option(LLAMA_CURL "llama: use libcurl to download model from an URL" OFF)
option(LLAMA_HIPBLAS "llama: use hipBLAS" OFF) option(LLAMA_HIPBLAS "llama: use hipBLAS" OFF)
option(LLAMA_HIP_UMA "llama: use HIP unified memory architecture" OFF) option(LLAMA_HIP_UMA "llama: use HIP unified memory architecture" OFF)
@ -409,6 +411,9 @@ if (LLAMA_CUDA)
if (LLAMA_CUDA_FORCE_MMQ) if (LLAMA_CUDA_FORCE_MMQ)
add_compile_definitions(GGML_CUDA_FORCE_MMQ) add_compile_definitions(GGML_CUDA_FORCE_MMQ)
endif() endif()
if (LLAMA_CUDA_NO_VMM)
add_compile_definitions(GGML_CUDA_NO_VMM)
endif()
add_compile_definitions(GGML_CUDA_DMMV_X=${LLAMA_CUDA_DMMV_X}) add_compile_definitions(GGML_CUDA_DMMV_X=${LLAMA_CUDA_DMMV_X})
add_compile_definitions(GGML_CUDA_MMV_Y=${LLAMA_CUDA_MMV_Y}) add_compile_definitions(GGML_CUDA_MMV_Y=${LLAMA_CUDA_MMV_Y})
if (DEFINED LLAMA_CUDA_DMMV_Y) if (DEFINED LLAMA_CUDA_DMMV_Y)
@ -434,7 +439,11 @@ if (LLAMA_CUDA)
set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} CUDA::cudart CUDA::cublas CUDA::cublasLt) set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} CUDA::cudart CUDA::cublas CUDA::cublasLt)
endif() endif()
set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} CUDA::cuda_driver) if (LLAMA_CUDA_NO_VMM)
# No VMM requested, no need to link directly with the cuda driver lib (libcuda.so)
else()
set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} CUDA::cuda_driver) # required by cuDeviceGetAttribute(), cuMemGetAllocationGranularity(...), ...
endif()
if (NOT DEFINED CMAKE_CUDA_ARCHITECTURES) if (NOT DEFINED CMAKE_CUDA_ARCHITECTURES)
# 52 == lowest CUDA 12 standard # 52 == lowest CUDA 12 standard

6
ggml-cuda.cu
View File

@ -113,7 +113,7 @@ static ggml_cuda_device_info ggml_cuda_init() {
for (int id = 0; id < info.device_count; ++id) { for (int id = 0; id < info.device_count; ++id) {
int device_vmm = 0; int device_vmm = 0;
#if !defined(GGML_USE_HIPBLAS) #if !defined(GGML_USE_HIPBLAS) && !defined(GGML_CUDA_NO_VMM)
CUdevice device; CUdevice device;
CU_CHECK(cuDeviceGet(&device, id)); CU_CHECK(cuDeviceGet(&device, id));
CU_CHECK(cuDeviceGetAttribute(&device_vmm, CU_DEVICE_ATTRIBUTE_VIRTUAL_MEMORY_MANAGEMENT_SUPPORTED, device)); CU_CHECK(cuDeviceGetAttribute(&device_vmm, CU_DEVICE_ATTRIBUTE_VIRTUAL_MEMORY_MANAGEMENT_SUPPORTED, device));
@ -259,7 +259,7 @@ struct ggml_cuda_pool_leg : public ggml_cuda_pool {
}; };
// pool with virtual memory // pool with virtual memory
#if !defined(GGML_USE_HIPBLAS) #if !defined(GGML_USE_HIPBLAS) && !defined(GGML_CUDA_NO_VMM)
struct ggml_cuda_pool_vmm : public ggml_cuda_pool { struct ggml_cuda_pool_vmm : public ggml_cuda_pool {
static const size_t CUDA_POOL_VMM_MAX_SIZE = 1ull << 35; // 32 GB static const size_t CUDA_POOL_VMM_MAX_SIZE = 1ull << 35; // 32 GB
@ -356,7 +356,7 @@ struct ggml_cuda_pool_vmm : public ggml_cuda_pool {
#endif // !defined(GGML_USE_HIPBLAS) #endif // !defined(GGML_USE_HIPBLAS)
std::unique_ptr<ggml_cuda_pool> ggml_backend_cuda_context::new_pool_for_device(int device) { std::unique_ptr<ggml_cuda_pool> ggml_backend_cuda_context::new_pool_for_device(int device) {
#if !defined(GGML_USE_HIPBLAS) #if !defined(GGML_USE_HIPBLAS) && !defined(GGML_CUDA_NO_VMM)
if (ggml_cuda_info().devices[device].vmm) { if (ggml_cuda_info().devices[device].vmm) {
return std::unique_ptr<ggml_cuda_pool>(new ggml_cuda_pool_vmm(device)); return std::unique_ptr<ggml_cuda_pool>(new ggml_cuda_pool_vmm(device));
} }