v4.2.1 update. (#2666)

This commit is contained in:
Junkai-Wu
2025-09-24 01:25:43 +08:00
committed by GitHub
parent 2b8dff1f90
commit 7a6d4ee099
12 changed files with 163 additions and 65 deletions

View File

@@ -7467,8 +7467,9 @@ def GenerateSM100_TensorOp_fp8_UMMA_gemm_with_blockwise(manifest, cuda_version,
kernel_schedule = to_grouped_schedule(KernelScheduleType.BlockwiseTmaWarpSpecialized1SmSm100, grouped)
epi_schedule = to_grouped_schedule(EpilogueScheduleType.TmaWarpSpecialized1Sm, grouped)
epi_schedule_nosmem = to_grouped_schedule(EpilogueScheduleType.BlockwiseNoSmemWarpSpecialized1Sm, grouped)
CreateGemmUniversal3xOperator(manifest, layouts, tile_descriptions, data_type,
[[kernel_schedule, epi_schedule]],
[[kernel_schedule, epi_schedule], [kernel_schedule, epi_schedule_nosmem]],
tile_schedulers=tile_schedulers, gemm_kind=gemm_kind)
def GenerateSM100_TensorOp_mixed_8bits_UMMA_gemm(manifest, cuda_version, gemm_kind=GemmKind.Universal3x):