[Kernel] Batch invariant NVFP4 MoE using cutlass (#40372)

Signed-off-by: Jakub Zakrzewski <jzakrzewski@nvidia.com>
Co-authored-by: OpenAI Codex <codex@openai.com>
Co-authored-by: Wentao Ye <44945378+yewentao256@users.noreply.github.com>
This commit is contained in:
Jakub Zakrzewski
2026-08-05 04:31:32 +00:00
committed by GitHub
co-authored by OpenAI Codex Wentao Ye
parent 41a7e7da0c
commit c25093a305
5 changed files with 316 additions and 102 deletions
@@ -734,6 +734,10 @@ class CutlassExpertsFp4(mk.FusedMoEExpertsModular):
# needed for STANDARD activation format kernels in EP mode.
return moe_parallel_config.ep_size == 1
@staticmethod
def _supports_batch_invariance() -> bool:
return True
@staticmethod
def activation_format() -> mk.FusedMoEActivationFormat:
return mk.FusedMoEActivationFormat.Standard