[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
@@ -358,7 +358,6 @@ class BailingMoeV25(nn.Module):
top_k=self.top_k,
hidden_size=self.hidden_size,
intermediate_size=config.moe_intermediate_size,
reduce_results=False,
renormalize=self.norm_expert_prob,
quant_config=quant_config,
prefix=f"{prefix}.experts",
@@ -368,6 +367,8 @@ class BailingMoeV25(nn.Module):
topk_group=self.topk_group,
use_grouped_topk=self.use_grouped_topk,
router_logits_dtype=self.router_dtype,
routed_scaling_factor=self.routed_scaling_factor,
apply_routed_scale_to_output=True,
)
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
@@ -383,22 +384,6 @@ class BailingMoeV25(nn.Module):
hidden_states=hidden_states, router_logits=router_logits
)
# Handle tuple return from SharedFusedMoE
if self.shared_experts is not None:
shared_output, final_hidden_states = final_hidden_states
else:
shared_output = None
final_hidden_states *= self.routed_scaling_factor
if shared_output is not None:
final_hidden_states = final_hidden_states + shared_output
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_size)