Skip to content

[Core] Fix L40 cublas heuristics #3440

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

Open
jberchtold-nvidia wants to merge 2 commits into
base: main
Choose a base branch
from
+12 −0
Open

[Core] Fix L40 cublas heuristics #3440

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
55f2514
Fix L40 cublas heuristics
jberchtold-nvidia Aug 28, 2026
a70ce1b
Merge branch 'main' into jberchtold/fix-l40-cublas-heuristic
jberchtold-nvidia Aug 28, 2026
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
12 changes: 12 additions & 0 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 @@ -747,6 +747,18 @@ void cublas_gemm(const Tensor *inputA, const Tensor *inputB, Tensor *outputD,
NVTE_CHECK_CUBLAS(cublasLtMatmulPreferenceCreate(&preference));
NVTE_CHECK_CUBLAS(cublasLtMatmulPreferenceSetAttribute(
preference, CUBLASLT_MATMUL_PREF_MAX_WORKSPACE_BYTES, &workspaceSize, sizeof(workspaceSize)));
// Avoid reduced-precision split-K reductions for BF16 GEMMs with FP32 compute. In particular,
// CUBLASLT_REDUCTION_SCHEME_OUTPUT_TYPE converts each split-K partial to BF16 before the final
// reduction, which can introduce significant error for cancellation-heavy dot products.
// Non-split algorithms remain eligible, as do split-K algorithms that reduce their partials in
// the FP32 compute type.
if (A_type == CUDA_R_16BF && B_type == CUDA_R_16BF && D_type == CUDA_R_16BF &&
gemm_compute_type == CUBLAS_COMPUTE_32F) {
const uint32_t reduction_scheme_mask = CUBLASLT_REDUCTION_SCHEME_COMPUTE_TYPE;
NVTE_CHECK_CUBLAS(cublasLtMatmulPreferenceSetAttribute(
preference, CUBLASLT_MATMUL_PREF_REDUCTION_SCHEME_MASK, &reduction_scheme_mask,
sizeof(reduction_scheme_mask)));
}
const auto A_alignment = _getAlignment(reinterpret_cast<uintptr_t>(param.A));
const auto B_alignment = _getAlignment(reinterpret_cast<uintptr_t>(param.B));
const auto C_alignment = _getAlignment(reinterpret_cast<uintptr_t>(C));
Expand Down
Loading