[BugFix]fix Qwen3 MoE call gate twice (#40664)

Signed-off-by: Kunshang Ji <kunshang.ji@intel.com>
This commit is contained in:
Kunshang Ji
2026-04-23 05:04:41 +00:00
committed by GitHub
parent fe9c3d6c5f
commit 342c58bc54
+13 -5
View File
@@ -231,11 +231,19 @@ class Qwen3MoeSparseMoeBlock(nn.Module):
if self.is_sequence_parallel:
hidden_states = sequence_parallel_chunk(hidden_states)
# router_logits: (num_tokens, n_experts)
router_logits, _ = self.gate(hidden_states)
final_hidden_states = self.experts(
hidden_states=hidden_states, router_logits=router_logits
)
if self.experts.is_internal_router:
# In this case, the gate/router runs inside the FusedMoE class
final_hidden_states = self.experts(
hidden_states=hidden_states, router_logits=hidden_states
)
else:
# Actually this will be dead code, since we always pass gate into
# FusedMoE in the current implementation. But we keep this code
# here for clarity and future flexibility.
router_logits, _ = self.gate(hidden_states)
final_hidden_states = self.experts(
hidden_states=hidden_states, router_logits=router_logits
)
if self.is_sequence_parallel:
final_hidden_states = tensor_model_parallel_all_gather(