[feat]: CUTLASS block scaled group gemm for SM100 (#19757)

Signed-off-by: Duncan Moss <djm.moss@gmail.com>
Co-authored-by: Duncan Moss <dmoss@nvidia.com>
This commit is contained in:
Duncan Moss
2025-07-04 12:58:04 -06:00
committed by GitHub
co-authored by Duncan Moss
parent 2f35a022e6
commit 3d184b95b8
13 changed files with 726 additions and 30 deletions
+9
View File
@@ -393,6 +393,15 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
{stride_tag});
ops.impl("cutlass_scaled_fp4_mm", torch::kCUDA, &cutlass_scaled_fp4_mm);
// cutlass blockwise scaledgroup GEMM
ops.def(
"cutlass_blockwise_scaled_grouped_mm(Tensor! output, Tensor a, Tensor b, "
"Tensor scales_a, Tensor scales_b, "
"Tensor problem_sizes, Tensor expert_offsets) -> ()",
{stride_tag});
ops.impl("cutlass_blockwise_scaled_grouped_mm", torch::kCUDA,
&cutlass_blockwise_scaled_grouped_mm);
// cutlass nvfp4 block scaled group GEMM
ops.def(
"cutlass_fp4_group_mm(Tensor! out, Tensor a, Tensor b,"