Skip to content

Tune SM100 GEMM scheduling for dW shapes #1380

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
kolehma8 merged 2 commits into from
Aug 14, 2026
Merged

Tune SM100 GEMM scheduling for dW shapes #1380

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
fae0e22
perf: tune SM100 GEMM scheduling for dW shapes
Aug 14, 2026
a54c2cb
Remove SM100 GEMM dispatch test
Aug 14, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
14 changes: 12 additions & 2 deletions src/liger_kernel/ops/cutedsl/ops/_sm100_gemm.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -878,12 +878,22 @@ def _validate_epilogue_inputs(a, b, out):
)


def _select_epilogue_config(a):
"""Select measured SM100 scheduling knobs while retaining the two-CTA kernel."""
m_tiles = (a.shape[0] + _CTA_M - 1) // _CTA_M
if a.shape[0] >= 4096 and m_tiles % 2 == 0:
if a.shape[1] in (256, 2048):
return 4, 2
if a.shape[1] in (512, 1024):
return 6, 2
return _NUM_AB_STAGES, 1


def _run_epilogue_gemm(a, b, out, epilogue):
_validate_epilogue_inputs(a, b, out)
epilogue_key = _validate_epilogue_callback(epilogue)
current_stream = _current_stream(a.device)
num_ab_stages = _NUM_AB_STAGES
swizzle_size = 1
num_ab_stages, swizzle_size = _select_epilogue_config(a)
use_tma_output = out.stride(0) * out.element_size() % 16 == 0
max_active_clusters = _max_active_clusters(a.device.index)

Expand Down