v4.2 tag release. (#2638)

This commit is contained in:
Junkai-Wu
2025-09-16 00:21:53 +08:00
committed by GitHub
parent 56f0718a97
commit 6a35b4d22f
161 changed files with 14056 additions and 3793 deletions

View File

@@ -637,7 +637,10 @@ def get_valid_schedules(tile_description, cuda_version, is_aligned, data_types,
if CudaToolkitVersionSatisfies(cuda_version, 12, 1):
# Pruning: don't stamp out fp8 ping-pong kernel with non-tma epilogue
if not is_fp8 or level >= 1:
schedules.append([to_grouped_schedule(KernelScheduleType.TmaWarpSpecializedPingpong, grouped), to_grouped_schedule(default_epilogue, grouped)])
if not is_blockwise(gemm_kind):
schedules.append([to_grouped_schedule(KernelScheduleType.TmaWarpSpecializedPingpong, grouped), to_grouped_schedule(default_epilogue, grouped)])
else:
schedules.append([to_grouped_schedule(KernelScheduleType.BlockwiseTmaWarpSpecializedPingpong, grouped), to_grouped_schedule(default_epilogue, grouped)])
if can_do_fp8_fast_accum:
if not grouped: