[MoE Refactor] Move the shared/fused expert output sum into MoERunnerBase (#35949)

Signed-off-by: Bill Nell <bnell@redhat.com>
Co-authored-by: Robert Shaw <114415538+robertgshaw2-redhat@users.noreply.github.com>
This commit is contained in:
bnellnm
2026-04-20 12:28:46 -04:00
committed by GitHub
co-authored by Robert Shaw
parent 595562651a
commit 726efe177b
52 changed files with 325 additions and 700 deletions
+1 -9
View File
@@ -212,7 +212,6 @@ class Qwen3MoeSparseMoeBlock(nn.Module):
top_k=config.num_experts_per_tok,
hidden_size=config.hidden_size,
intermediate_size=config.moe_intermediate_size,
reduce_results=False,
renormalize=config.norm_topk_prob,
quant_config=quant_config,
prefix=f"{prefix}.experts",
@@ -234,22 +233,15 @@ class Qwen3MoeSparseMoeBlock(nn.Module):
# router_logits: (num_tokens, n_experts)
router_logits, _ = self.gate(hidden_states)
shared_out, fused_out = self.experts(
final_hidden_states = self.experts(
hidden_states=hidden_states, router_logits=router_logits
)
final_hidden_states = (
shared_out + fused_out if shared_out is not None else fused_out
)
if self.is_sequence_parallel:
final_hidden_states = tensor_model_parallel_all_gather(
final_hidden_states, 0
)
final_hidden_states = final_hidden_states[:num_tokens]
elif self.tp_size > 1:
final_hidden_states = self.experts.maybe_all_reduce_tensor_model_parallel( # noqa E501
final_hidden_states
)
# return to 1d if input is 1d
return final_hidden_states.squeeze(0) if is_input_1d else final_hidden_states