Skip to content

[https://nvbugs/6919819][perf] Autotune the CuTeDSL FC2 N tile again when the override is unset #20277

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
chenfeiz0326 merged 1 commit into from
Oct 16, 2026
Merged

[https://nvbugs/6919819][perf] Autotune the CuTeDSL FC2 N tile again when the override is unset #20277

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
11 changes: 6 additions & 5 deletions tensorrt_llm/_torch/custom_ops/cute_dsl_custom_ops.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -44,7 +44,6 @@
SWIGLU_LIMIT_SCALAR_DISABLED = -1.0
_CUTEDSL_FC2_N_TILE_SIZE_ENV = "TRTLLM_CUTEDSL_FC2_N_TILE_SIZE"
_CUTEDSL_FC2_N_TILE_SIZES = (128, 256)
_CUTEDSL_FC2_DEFAULT_N_TILE_SIZE = 128

# Legacy Rubin ops use a float sentinel for absent SiTU soft-caps.
# The Blackwell act-fusion op accepts Optional[float] directly.
Expand Down Expand Up @@ -192,10 +191,12 @@ def _get_sm107_nvfp4_default_mma_config(

@functools.lru_cache(maxsize=1)
def _get_cutedsl_fc2_n_tile_size_override() -> Optional[int]:
value = os.environ.get(
_CUTEDSL_FC2_N_TILE_SIZE_ENV,
str(_CUTEDSL_FC2_DEFAULT_N_TILE_SIZE),
)
"""Return the FC2 N tile pinned by the user, or None to autotune over all of them.

Unset and ``0`` both autotune. ``128`` or ``256`` pins that tile for shapes
where it is known to win end to end.
"""
value = os.environ.get(_CUTEDSL_FC2_N_TILE_SIZE_ENV, "0")
Comment thread
chenfeiz0326 marked this conversation as resolved.
if value == "0":
return None
valid_values = tuple(
Expand Down
3 changes: 2 additions & 1 deletion tests/unittest/_torch/moe/fused_moe/test_cutedsl_fc2_tuning.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -61,6 +61,7 @@ def _make_runner(
[
((128, 128), (1, 1), False),
((128, 128), (1, 2), False),
((128, 256), (1, 1), False),
],
),
(
Expand Down Expand Up @@ -97,7 +98,7 @@ def test_fc2_n_tile_size_override(
tactics = runner.get_valid_tactics(inputs, OptimizationProfile())

assert tactics == expected_tactics
expected_override = 128 if env_value is None else (None if env_value == "0" else int(env_value))
expected_override = None if env_value in (None, "0") else int(env_value)
assert runner._get_default_tactic()[0][1] == (expected_override or 128)
if expected_override is None:
assert "fc2_n_tile_size_override" not in runner.unique_id()
Expand Down
Loading