[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
-8
View File
@@ -153,7 +153,6 @@ class Qwen3NextSparseMoeBlock(nn.Module):
top_k=config.num_experts_per_tok,
hidden_size=config.hidden_size,
intermediate_size=config.moe_intermediate_size,
reduce_results=False,
renormalize=getattr(config, "norm_topk_prob", True),
quant_config=quant_config,
prefix=f"{prefix}.experts",
@@ -183,18 +182,11 @@ class Qwen3NextSparseMoeBlock(nn.Module):
hidden_states=hidden_states, router_logits=router_logits
)
if self.shared_expert is not None:
final_hidden_states = final_hidden_states[0] + final_hidden_states[1]
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 final_hidden_states.view(orig_shape)