Fix Humming MoE activation_output aliasing (#54056)
Elvir Crnčević committed
Aug 27, 2026 at 20:28 UTC
674c284dbd2012f4b3e0da362aa11b75044c5d0c
1 file changed
+9
vllm/model_executor/layers/fused_moe/experts/fused_humming_moe.py
+9
@@ -541,6 +541,15 @@ class HummingExpertsBase(mk.FusedMoEExpertsModular):
541
"down_output",
542
]
543
544
+ # Note: The fused SITU+FP8 quant goes gate_up_output ->
545
+ # quanted_down_input directly, never materializing
546
+ # activation_output. Dropping activation_output from the chain
547
+ # flips the even/odd workspace 2-coloring below so
548
+ # gate_up_output and quanted_down_input land
549
+ # on DIFFERENT workspaces.
550
+ if self.fused_situ_quant_enabled(activation):
551
+ required_buffers.remove("activation_output")
552
+
553
# batched moe use down_output as output
554
if not self.is_batched():
555
required_buffers.append("output")