NVIDIA/cutlass
1
0
Fork 0
mirror of https://github.com/NVIDIA/cutlass.git synced 2026-04-20 14:59:01 +00:00
Code Issues Packages Projects Releases Wiki Activity

3.6.0 update (#2005)

Browse Source 
* 3.6.0 update

* doc and swap stuff

---------

Co-authored-by: yuzhai <yuzhai@nvidia.com>
Co-authored-by: Haicheng Wu <haichengw@nvidia.com>
This commit is contained in:
Yujia Zhai
2024-12-24 22:34:40 -08:00
committed by GitHub
parent e1cd8c7866
commit 3d261a5974
258 changed files with 10863 additions and 3883 deletions
Download Patch File Download Diff File Expand all files Collapse all files

26
python/cutlass_library/sm90_utils.py
View File

@@ -492,6 +492,21 @@ def get_valid_schedules(tile_description, cuda_version, is_aligned, data_types,
if not (is_fp8 and is_sparse):
schedules.append([KernelScheduleType.TmaWarpSpecialized, default_epilogue])
stream_k_schedules = []
if CudaToolkitVersionSatisfies(cuda_version, 12, 0):
if can_do_tma_epilogue:
assert not requires_transposed_epilogue
# Inconsistency: fp8 pingpong only gets stamped out with fast accum
if not is_fp8 or level >= 1:
schedules.append([
KernelScheduleType.TmaWarpSpecializedPingpong,
EpilogueScheduleType.TmaWarpSpecialized
])
if can_do_fp8_fast_accum:
schedules.append([
KernelScheduleType.TmaWarpSpecializedPingpongFP8FastAccum,
EpilogueScheduleType.TmaWarpSpecialized
])
if CudaToolkitVersionSatisfies(cuda_version, 12, 1):
# Pruning: don't stamp out fp8 ping-ponging kernel with non-tma epilogue
@@ -526,17 +541,6 @@ def get_valid_schedules(tile_description, cuda_version, is_aligned, data_types,
# persistent kernels with TMA epilogues
if can_do_tma_epilogue:
assert not requires_transposed_epilogue
# Inconsistency: fp8 pingpong only gets stamped out with fast accum
if not is_fp8 or level >= 1:
schedules.append([
KernelScheduleType.TmaWarpSpecializedPingpong,
EpilogueScheduleType.TmaWarpSpecialized
])
if can_do_fp8_fast_accum:
schedules.append([
KernelScheduleType.TmaWarpSpecializedPingpongFP8FastAccum,
EpilogueScheduleType.TmaWarpSpecialized
])
if can_do_cooperative:
# Sparse kernels only support FastAccum FP8 mainloop
if not (is_fp8 and is_sparse):