[MoE] Defer the TRTLLM-Gen top-k finalize on the modular path (#58635)

Signed-off-by: Yongye Zhu <zyy1102000@gmail.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
Yongye Zhu
2026-09-25 15:38:29 +08:00
committed by GitHub
co-authored by Claude Opus 5.5
parent 24928a4d62
commit 378504a544
26 changed files with 652 additions and 117 deletions
@@ -690,7 +690,7 @@ class ModelOptFp8MoEMethod(FusedMoEMethodBase):
topk_ids: torch.Tensor,
shared_experts: SharedExperts | None,
shared_experts_input: torch.Tensor | None,
) -> torch.Tensor:
) -> torch.Tensor | UnfinalizedMoEOutput:
assert not self.is_monolithic
assert self.moe_kernel is not None
return self.moe_kernel.apply(
@@ -1102,7 +1102,7 @@ class ModelOptNvFp4FusedMoE(FusedMoEMethodBase):
topk_ids: torch.Tensor,
shared_experts: SharedExperts | None,
shared_experts_input: torch.Tensor | None,
) -> torch.Tensor:
) -> torch.Tensor | UnfinalizedMoEOutput:
assert not self.is_monolithic
assert self.moe_kernel is not None
return self.moe_kernel.apply(
@@ -1471,7 +1471,7 @@ class ModelOptMxFp8FusedMoE(FusedMoEMethodBase):
topk_ids: torch.Tensor,
shared_experts: SharedExperts | None,
shared_experts_input: torch.Tensor | None,
) -> torch.Tensor:
) -> torch.Tensor | UnfinalizedMoEOutput:
assert not self.is_monolithic
assert self.moe_kernel is not None
return self.moe_kernel.apply(