[Bugfix] Fix double reduce in flashinfer_nvlink_two_sided and flashinfer_nvlink_one_sided backends (#41382)
Signed-off-by: amitz-nv <203509407+amitz-nv@users.noreply.github.com>
This commit is contained in:
+1
-1
@@ -67,7 +67,7 @@ class FlashInferNVLinkOneSidedPrepareAndFinalize(mk.FusedMoEPrepareAndFinalizeMo
|
||||
return self.num_dispatchers_
|
||||
|
||||
def output_is_reduced(self) -> bool:
|
||||
return False
|
||||
return True
|
||||
|
||||
def topk_indices_dtype(self) -> torch.dtype | None:
|
||||
return torch.int32
|
||||
|
||||
+1
-1
@@ -50,7 +50,7 @@ class FlashInferNVLinkTwoSidedPrepareAndFinalize(mk.FusedMoEPrepareAndFinalizeMo
|
||||
return self.num_dispatchers_
|
||||
|
||||
def output_is_reduced(self) -> bool:
|
||||
return False
|
||||
return True
|
||||
|
||||
def _apply_router_weight_on_input(
|
||||
self,
|
||||
|
||||
Reference in New Issue
Block a user