[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 -12
View File
@@ -131,7 +131,6 @@ class AfmoeMoE(nn.Module):
top_k=config.num_experts_per_tok,
hidden_size=config.hidden_size,
intermediate_size=config.moe_intermediate_size,
reduce_results=False,
renormalize=self.route_norm if self.score_func == "sigmoid" else False,
quant_config=quant_config,
use_grouped_topk=True,
@@ -152,20 +151,10 @@ class AfmoeMoE(nn.Module):
router_logits = self.gate(hidden_states.to(dtype=torch.float32))
fused_moe_out = self.experts(
final_hidden_states = self.experts(
hidden_states=hidden_states, router_logits=router_logits
)
if self.shared_experts is not None:
shared_output, final_hidden_states = fused_moe_out
final_hidden_states = final_hidden_states + shared_output
else:
final_hidden_states = fused_moe_out
if self.tp_size > 1:
final_hidden_states = self.experts.maybe_all_reduce_tensor_model_parallel(
final_hidden_states
)
return final_hidden_states.view(num_tokens, hidden_dim)