[11b/n] Migrate Machete kernels to torch stable ABI (#45304)

Signed-off-by: Chris Leonard <chleonar@redhat.com>
Signed-off-by: Shengqi Chen <harry-chen@outlook.com>
Co-authored-by: Shengqi Chen <harry-chen@outlook.com>
This commit is contained in:
Chris Leonard
2026-06-12 10:36:49 +00:00
committed by GitHub
co-authored by Shengqi Chen
parent a37b4a940e
commit a014dddbaa
17 changed files with 341 additions and 323 deletions
+33
View File
@@ -34,6 +34,39 @@ STABLE_TORCH_LIBRARY_FRAGMENT(_C, ops) {
// TODO: Remove this once ROCm upgrade to torch 2.11.
ops.def("get_cuda_view_from_cpu_tensor(Tensor cpu_tensor) -> Tensor");
// Machete (Dense) Optimized Mixed Precision GEMM for Hopper.
ops.def(
"machete_supported_schedules("
" ScalarType a_type,"
" int b_type,"
" ScalarType? maybe_group_scales_type,"
" ScalarType? maybe_group_zeros_type,"
" ScalarType? maybe_channel_scales_type,"
" ScalarType? maybe_token_scales_type,"
" ScalarType? maybe_out_type"
") -> str[]");
ops.def(
"machete_mm("
" Tensor A,"
" Tensor B,"
" int b_type,"
" ScalarType? out_type,"
" Tensor? group_scales,"
" Tensor? group_zeros,"
" int? group_size,"
" Tensor? channel_scales,"
" Tensor? token_scales,"
" str? schedule"
") -> Tensor");
ops.def(
"machete_prepack_B("
" Tensor B,"
" ScalarType a_type,"
" int b_type,"
" ScalarType? group_scales_type"
") -> Tensor");
// conditionally compiled so impl registration is in source file
// Marlin GEMM
ops.def(
"marlin_gemm(Tensor a, Tensor? c_or_none, Tensor b_q_weight, "