[Kernel]Support W4A8 Grouped GEMM on Hopper (#29691)

Signed-off-by: czhu-cohere <conway.zhu@cohere.com>
This commit is contained in:
czhu-cohere
2025-12-08 19:29:06 -08:00
committed by GitHub
parent ea657f2078
commit f6227c22ab
22 changed files with 2045 additions and 101 deletions
+25 -1
View File
@@ -350,6 +350,29 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
ops.def("cutlass_encode_and_reorder_int4b(Tensor B) -> Tensor");
// conditionally compiled so impl registration is in source file
// CUTLASS w4a8 grouped GEMM
ops.def(
"cutlass_w4a8_moe_mm("
" Tensor! out_tensors,"
" Tensor a_tensors,"
" Tensor b_tensors,"
" Tensor a_scales,"
" Tensor b_scales,"
" Tensor b_group_scales,"
" int b_group_size,"
" Tensor expert_offsets,"
" Tensor problem_sizes,"
" Tensor a_strides,"
" Tensor b_strides,"
" Tensor c_strides,"
" Tensor group_scale_strides,"
" str? maybe_schedule"
") -> ()");
ops.def(
"cutlass_encode_and_reorder_int4b_grouped(Tensor b_tensors) -> (Tensor, "
"Tensor)");
// conditionally compiled so impl registration is in source file
#endif
// Dequantization for GGML.
@@ -466,7 +489,8 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
" Tensor! problem_sizes1, "
" Tensor! problem_sizes2, "
" int num_experts, int n, int k, "
" Tensor? blockscale_offsets) -> ()");
" Tensor? blockscale_offsets, "
" bool? force_swap_ab) -> ()");
ops.impl("get_cutlass_moe_mm_problem_sizes", torch::kCUDA,
&get_cutlass_moe_mm_problem_sizes);