[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
+3 -13
View File
@@ -37,7 +37,6 @@ 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.activation import SiluAndMul
from vllm.model_executor.layers.attention import Attention
@@ -120,7 +119,6 @@ class Dots1MoE(nn.Module):
prefix: str = "",
):
super().__init__()
self.tp_size = get_tensor_model_parallel_world_size()
self.routed_scaling_factor = config.routed_scaling_factor
self.n_shared_experts = config.n_shared_experts
@@ -163,7 +161,6 @@ class Dots1MoE(nn.Module):
top_k=config.num_experts_per_tok,
hidden_size=config.hidden_size,
intermediate_size=config.moe_intermediate_size,
reduce_results=False,
renormalize=config.norm_topk_prob,
quant_config=quant_config,
use_grouped_topk=True,
@@ -171,9 +168,9 @@ class Dots1MoE(nn.Module):
topk_group=config.topk_group,
prefix=f"{prefix}.experts",
scoring_func=config.scoring_func,
# we do scaling outside, set factor to 1.0 to avoid double mul
routed_scaling_factor=1.0,
e_score_correction_bias=self.gate.e_score_correction_bias,
routed_scaling_factor=self.routed_scaling_factor,
apply_routed_scale_to_output=True,
)
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
@@ -182,16 +179,9 @@ class Dots1MoE(nn.Module):
router_logits, _ = self.gate(hidden_states)
shared_out, routed_out = self.experts(
final_hidden_states = self.experts(
hidden_states=hidden_states, router_logits=router_logits
)
if self.shared_experts is not None:
final_hidden_states = (routed_out + shared_out) * self.routed_scaling_factor
else:
final_hidden_states = routed_out * self.routed_scaling_factor
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)