[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
-11
View File
@@ -194,7 +194,6 @@ class Ernie4_5_MoeMoE(nn.Module):
top_k=config.moe_k,
hidden_size=config.hidden_size,
intermediate_size=config.moe_intermediate_size,
reduce_results=False,
renormalize=True,
quant_config=quant_config,
prefix=f"{prefix}.experts",
@@ -215,16 +214,6 @@ class Ernie4_5_MoeMoE(nn.Module):
hidden_states=hidden_states, router_logits=router_logits
)
if self.has_shared_experts:
final_hidden_states = final_hidden_states[0] + final_hidden_states[1]
else:
final_hidden_states = final_hidden_states[1]
if self.tp_size > 1:
final_hidden_states = self.experts.maybe_all_reduce_tensor_model_parallel(
final_hidden_states
)
return final_hidden_states.view(orig_shape)