Skip to content

Check CUDA memory pool support #3931

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
wants to merge 8 commits into from
Closed

Check CUDA memory pool support #3931

Changes from 2 commits
Commits
Show all changes
8 commits
Select commit Hold shift + click to select a range
ce4df17
Check CUDA memory support in device properties.
young-developer Nov 3, 2023
bd56886
set nullptr to memory pool element if it failed during initialization.
young-developer Nov 3, 2023
c42ca8f
GGML_CUDA_FORCE_CUSTOM_MEMORY_POOL was added to force use only custom…
young-developer Nov 3, 2023
815bf1a
prop.memoryPoolsSupported cant be found in cuda 17. Revert back to ba…
young-developer Nov 3, 2023
56e5162
All memory pool operation are checked during init phase. For CUDA 12+…
young-developer Nov 4, 2023
81931b2
Multi GPU memory pool access + Check memory pool support of multiple …
young-developer Nov 4, 2023
863166b
Skip GPUs without mem pool support.
young-developer Nov 4, 2023
2b0303a
CUDA pool is optional now.
young-developer Nov 4, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
23 changes: 15 additions & 8 deletions ggml-cuda.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -5844,7 +5844,21 @@ void ggml_init_cublas() {
for (int id = 0; id < g_device_count; ++id) {
cudaDeviceProp prop;
CUDA_CHECK(cudaGetDeviceProperties(&prop, id));
fprintf(stderr, " Device %d: %s, compute capability %d.%d\n", id, prop.name, prop.major, prop.minor);
fprintf(stderr, " Device %d: %s, compute capability %d.%d", id, prop.name, prop.major, prop.minor);

// configure memory pool
if (prop.memoryPoolsSupported == 1) {
cudaError_t err = cudaDeviceGetMemPool(&g_cudaMemPools[id], id);
if (err == cudaSuccess) {
size_t treshold = UINT64_MAX;
CUDA_CHECK(cudaMemPoolSetAttribute(g_cudaMemPools[id], cudaMemPoolAttrReleaseThreshold, &treshold));
fprintf(stderr, ", CUDA memory pool is supported\n");
} else {
g_cudaMemPools[id] = nullptr;
}
} else {
fprintf(stderr, ", CUDA memory pool is not supported\n");
}

g_tensor_split[id] = total_vram;
total_vram += prop.totalGlobalMem;
Expand All @@ -5869,13 +5883,6 @@ void ggml_init_cublas() {
// create cublas handle
CUBLAS_CHECK(cublasCreate(&g_cublas_handles[id]));
CUBLAS_CHECK(cublasSetMathMode(g_cublas_handles[id], CUBLAS_TF32_TENSOR_OP_MATH));

// configure memory pool
cudaError_t err = cudaDeviceGetMemPool(&g_cudaMemPools[id], id);
if (err == cudaSuccess) {
size_t treshold = UINT64_MAX;
CUDA_CHECK(cudaMemPoolSetAttribute(g_cudaMemPools[id], cudaMemPoolAttrReleaseThreshold, &treshold));
}
}

// configure logging to stdout
Expand Down