[Bugfix][MoE] Write Humming results to the supplied output buffer (#50338)
Signed-off-by: Netanel Haber <58652339+netanel-haber@users.noreply.github.com> Co-authored-by: OpenAI Codex <noreply@openai.com>
Netanel Haber committed
Jul 30, 2026 at 21:34 UTC
61cacd272f5a95aa2a091d711ed91a4aa85522d5
1 file changed
+3
-12
vllm/model_executor/layers/fused_moe/experts/fused_humming_moe.py
+3
-12
@@ -662,12 +662,9 @@ class HummingIndexedExperts(HummingExpertsBase):
662
topk_weights=topk_weights,
663
topk_ids=topk_ids,
664
expert_map=expert_map,
665
- outputs=buffers["output"],
665
+ outputs=output,
666
)
667
668
- # Note: output is already written to buffers["output"]
669
- # which aliases workspace13/output
670
-
668
669
class HummingGroupedExperts(HummingExpertsBase):
670
def finalize_weight_and_reduce_impl(self) -> mk.TopKWeightAndReduce:
@@ -777,16 +774,13 @@ class HummingGroupedExperts(HummingExpertsBase):
774
)
775
776
moe_unpermute(
780
- out=buffers["output"],
777
+ out=output,
778
permuted_hidden_states=buffers["down_output"].view(*topk_ids.shape, -1),
779
topk_weights=topk_weights,
780
inv_permuted_idx=inv_perm,
781
expert_first_token_offset=expert_first_token_offset,
782
)
783
787
- # Note: output is already written to buffers["output"]
788
- # which aliases workspace13/output
789
-
784
785
class BatchedHummingGroupedExperts(HummingExpertsBase):
786
def finalize_weight_and_reduce_impl(self) -> mk.TopKWeightAndReduce:
@@ -880,13 +874,10 @@ class BatchedHummingGroupedExperts(HummingExpertsBase):
874
layer=self.layer,
875
inputs=inputs,
876
input_scale=input_scale,
883
- outputs=buffers["down_output"].view(-1, hidden_states.size(-1)),
877
+ outputs=output.view(-1, hidden_states.size(-1)),
878
valid_shape_m=valid_shape_m,
879
expert_layout=expert_num_tokens,
880
compute_config=self.compute_config_str,
881
tuning_config=self.w2_tuning_config_str,
882
sublayer_name="w2",
883
)
890
-
891
- # Note: output is already written to buffers["down_output"]
892
- # which aliases workspace13/output