Skip to content
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

Better cuBLAS handle management #1389

Open
wants to merge 5 commits into
base: main
Choose a base branch
from
Open

Better cuBLAS handle management #1389

Changes from 2 commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
999879e
Do not create multiple cublas handle
ptrendx Jan 2, 2025
7b868b0
[pre-commit.ci] auto fixes from pre-commit.com hooks
pre-commit-ci[bot] Jan 2, 2025
529aec2
Fix for multiple GPUs per thread
ptrendx Jan 6, 2025
44c7d70
[pre-commit.ci] auto fixes from pre-commit.com hooks
pre-commit-ci[bot] Jan 6, 2025
e62a128
Fix multithreaded execution
ptrendx Jan 6, 2025
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
22 changes: 20 additions & 2 deletions transformer_engine/common/gemm/cublaslt_gemm.cu
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -50,6 +50,25 @@ uint32_t _getAlignment(uintptr_t address) {

namespace transformer_engine {

class cublasHandleManager {
public:
static cublasHandleManager &Instance() {
static thread_local cublasHandleManager instance;
return instance;
}

cublasLtHandle_t GetHandle() {
static thread_local std::once_flag flag;
std::call_once(flag, [&] { NVTE_CHECK_CUBLAS(cublasLtCreate(&handle_)); });
return handle_;
}

~cublasHandleManager() {}

private:
cublasLtHandle_t handle_ = nullptr;
};

void cublas_gemm(const Tensor *inputA, const Tensor *inputB, Tensor *outputD,
const Tensor *inputBias, Tensor *outputPreGelu, int m, int n, int k, int lda,
int ldb, int ldd, cublasOperation_t transa, cublasOperation_t transb, bool grad,
Expand Down Expand Up @@ -98,8 +117,7 @@ void cublas_gemm(const Tensor *inputA, const Tensor *inputB, Tensor *outputD,
float zero = 0.0;
float beta = (accumulate) ? one : zero;

cublasLtHandle_t handle;
NVTE_CHECK_CUBLAS(cublasLtCreate(&handle));
cublasLtHandle_t handle = cublasHandleManager::Instance().GetHandle();

cublasLtMatmulDesc_t operationDesc = nullptr;
cublasLtMatrixLayout_t Adesc = nullptr, Bdesc = nullptr, Cdesc = nullptr, Ddesc = nullptr;
Expand Down