[MoE Refactor] FusedMoE/MoERunner inversion refactor (#41184)

Signed-off-by: Bill Nell <bnell@redhat.com>
This commit is contained in:
bnellnm
2026-06-08 10:42:58 -04:00
committed by GitHub
parent 753e9d55e6
commit dc68bd8c41
90 changed files with 2738 additions and 2031 deletions
+2 -1
View File
@@ -25,6 +25,7 @@ from vllm.model_executor.layers.activation import SiluAndMul, SwigluStepAndMul
from vllm.model_executor.layers.attention import Attention
from vllm.model_executor.layers.fused_moe import (
FusedMoE,
MoERunner,
fused_moe_make_expert_params_mapping,
)
from vllm.model_executor.layers.layernorm import GemmaRMSNorm
@@ -954,7 +955,7 @@ class Step3p5ForCausalLM(nn.Module, SupportsPP, MixtureOfExperts):
) -> None:
for layer_idx, layer in enumerate(self.moe_layers):
experts = layer.experts
assert isinstance(experts, FusedMoE)
assert isinstance(experts, MoERunner)
# Register the expert weights.
self.expert_weights.append(experts.get_expert_weights())
experts.set_eplb_state(