A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 13, 04:30 UTC
[Bugfix] Fix incorrect use of hidden_states for shared_experts due to do_naive_dispatch_combine (#28740)
Signed-off-by: Alexander Matveev <amatveev@redhat.com>
Alexander Matveev committed
Nov 14, 2025 at 17:13 UTC
e5c78956c0c576d8f7230c29550ff09ffff0c064
1 file changed
+4
-2
vllm/model_executor/layers/fused_moe/layer.py
+4
-2
@@ -1749,14 +1749,16 @@ class FusedMoE(CustomOp):
1749
1750
with sp_ctx:
1751
if do_naive_dispatch_combine:
1752
- hidden_states, router_logits = get_ep_group().dispatch(
1752
+ hidden_states_combined, router_logits = get_ep_group().dispatch(
1753
hidden_states, router_logits, self.is_sequence_parallel
1754
)
1755
1756
# Matrix multiply.
1757
final_hidden_states = self.quant_method.apply(
1758
layer=self,
1759
- x=hidden_states,
1759
+ x=hidden_states_combined
1760
+ if do_naive_dispatch_combine
1761
+ else hidden_states,
1762
router_logits=router_logits,
1763
top_k=self.top_k,
1764
renormalize=self.renormalize,