[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
-5
View File
@@ -36,7 +36,6 @@ from vllm.config import CacheConfig, ModelConfig, VllmConfig
from vllm.distributed import (
get_pp_group,
get_tensor_model_parallel_world_size,
tensor_model_parallel_all_reduce,
)
from vllm.model_executor.layers.attention import Attention
from vllm.model_executor.layers.fused_moe import FusedMoE
@@ -104,7 +103,6 @@ class MiniMaxM2MoE(nn.Module):
e_score_correction_bias=self.e_score_correction_bias,
hidden_size=config.hidden_size,
intermediate_size=config.intermediate_size,
reduce_results=False,
renormalize=True,
quant_config=quant_config,
prefix=f"{prefix}.experts",
@@ -134,9 +132,6 @@ class MiniMaxM2MoE(nn.Module):
final_hidden_states = self.experts(
hidden_states=hidden_states, router_logits=router_logits
)
final_hidden_states = final_hidden_states
if self.tp_size > 1:
final_hidden_states = tensor_model_parallel_all_reduce(final_hidden_states)
return final_hidden_states.view(num_tokens, hidden_dim)