[MoE Refactor] Make SharedExperts class for use with DefaultMoERunner (#35153)

Signed-off-by: Bill Nell <bnell@redhat.com>
This commit is contained in:
bnellnm
2026-04-01 09:44:08 -04:00
committed by GitHub
parent 5e30e9b9a9
commit 7cf56a59a2
34 changed files with 556 additions and 397 deletions
@@ -935,7 +935,7 @@ class ModelOptFp8MoEMethod(FusedMoEMethodBase):
layer: FusedMoE,
x: torch.Tensor,
router_logits: torch.Tensor,
) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
) -> torch.Tensor:
assert self.is_monolithic
assert self.moe_kernel is not None
return self.moe_kernel.apply_monolithic(
@@ -960,7 +960,7 @@ class ModelOptFp8MoEMethod(FusedMoEMethodBase):
topk_weights: torch.Tensor,
topk_ids: torch.Tensor,
shared_experts_input: torch.Tensor | None,
) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
) -> torch.Tensor:
assert not self.is_monolithic
assert self.moe_kernel is not None
return self.moe_kernel.apply(
@@ -1419,7 +1419,7 @@ class ModelOptNvFp4FusedMoE(FusedMoEMethodBase):
layer: FusedMoE,
x: torch.Tensor,
router_logits: torch.Tensor,
) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
) -> torch.Tensor:
assert self.is_monolithic
assert self.moe_kernel is not None
return self.moe_kernel.apply_monolithic(
@@ -1444,7 +1444,7 @@ class ModelOptNvFp4FusedMoE(FusedMoEMethodBase):
topk_weights: torch.Tensor,
topk_ids: torch.Tensor,
shared_experts_input: torch.Tensor | None,
) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
) -> torch.Tensor:
assert not self.is_monolithic
assert self.moe_kernel is not None
return self.moe_kernel.apply(