happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

Merge 5d9f64c54e into 0356e33aaf

This commit is contained in:
Russell 2026-04-01 15:38:36 +02:00 committed by GitHub
parent 0356e33aaf 5d9f64c54e
commit 5f4610fd99
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 3 additions and 9 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

12
ggml/src/ggml-cuda/ggml-cuda.cu
View File

@ -98,16 +98,9 @@ void ggml_cuda_error(const char * stmt, const char * func, const char * file, in
GGML_ABORT(GGML_CUDA_NAME " error");
}
// this is faster on Windows
// probably because the Windows CUDA libraries forget to make this check before invoking the drivers
// always set device explicitly — early-return optimization is unsafe on ROCm multi-GPU
// with uninitialized thread contexts (see https://github.com/ggml-org/llama.cpp/issues/21140)
void ggml_cuda_set_device(int device) {
int current_device;
CUDA_CHECK(cudaGetDevice(&current_device));
if (device == current_device) {
return;
}
CUDA_CHECK(cudaSetDevice(device));
}
@ -2851,6 +2844,7 @@ static void ggml_backend_cuda_set_tensor_async(ggml_backend_t backend, ggml_tens
GGML_ASSERT(buf->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device) && "unsupported buffer type");
ggml_cuda_set_device(cuda_ctx->device);
CUDA_CHECK(cudaMemcpyAsync((char *)tensor->data + offset, data, size, cudaMemcpyHostToDevice, cuda_ctx->stream()));
}