xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-14 06:45:00 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Bugfix] Fix incorrect use of hidden_states for shared_experts due to do_naive_dispatch_combine (#28740)

Browse Source 
Signed-off-by: Alexander Matveev <amatveev@redhat.com>
This commit is contained in:
Alexander Matveev 2025-11-14 17:13:46 -05:00 committed by GitHub
parent 2e0ad629b0
commit e5c78956c0
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 4 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
vllm/model_executor/layers/fused_moe/layer.py
View File

@ -1749,14 +1749,16 @@ class FusedMoE(CustomOp):
with sp_ctx: with sp_ctx:
if do_naive_dispatch_combine: if do_naive_dispatch_combine:
hidden_states, router_logits = get_ep_group().dispatch( hidden_states_combined, router_logits = get_ep_group().dispatch(
hidden_states, router_logits, self.is_sequence_parallel hidden_states, router_logits, self.is_sequence_parallel
) )
# Matrix multiply. # Matrix multiply.
final_hidden_states = self.quant_method.apply( final_hidden_states = self.quant_method.apply(
layer=self, layer=self,
x=hidden_states, x=hidden_states_combined
if do_naive_dispatch_combine
else hidden_states,
router_logits=router_logits, router_logits=router_logits,
top_k=self.top_k, top_k=self.top_k,
renormalize=self.renormalize, renormalize=self.renormalize,