CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-10-31 08:51:55 +00:00
Code Issues Packages Projects Releases Wiki Activity

HIP: Only call rocblas_initialize on rocblas versions with the multiple instantation bug (#11080)

Browse Source 
This disables the workaround on rocblas fixed versions (>=4.0.0) to eliminate the runtime cost and unnecessary VRAM allocation of loading all tensile objects.
This commit is contained in:
Nikita Sarychev
2025-01-28 07:42:20 -08:00
committed by GitHub
parent 7fee2889e6
commit cae9fb4361
1 changed files with 20 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

22
ggml/src/ggml-cuda/ggml-cuda.cu
View File

@@ -42,6 +42,7 @@
#include <algorithm>
#include <array>
#include <atomic>
#include <charconv>
#include <cinttypes>
#include <cstddef>
#include <cstdint>
@@ -172,8 +173,25 @@ static ggml_cuda_device_info ggml_cuda_init() {
#ifdef __HIP_PLATFORM_AMD__
// Workaround for a rocBLAS bug when using multiple graphics cards:
// https://github.com/ROCmSoftwarePlatform/rocBLAS/issues/1346
rocblas_initialize();
CUDA_CHECK(cudaDeviceSynchronize());
{
int major_version = 0;
size_t version_length = 0;
if (rocblas_get_version_string_size(&version_length) == rocblas_status_success) {
std::string version(version_length, '\0');
if (rocblas_get_version_string(version.data(), version.size()) == rocblas_status_success) {
version.resize(::strlen(version.c_str()));
int parsed_value = 0;
if (std::from_chars(version.c_str(), version.c_str() + version.length(), parsed_value).ec == std::errc()) {
major_version = parsed_value;
}
}
}
if (major_version < 4) {
GGML_LOG_DEBUG(GGML_CUDA_NAME " calling rocblas_initialize as a workaround for a rocBLAS bug\n");
rocblas_initialize();
CUDA_CHECK(cudaDeviceSynchronize());
}
}
#endif
ggml_cuda_device_info info = {};