[MoE Refactor] Remove SharedFusedMoE class (#35782)

Signed-off-by: Bill Nell <bnell@redhat.com>
This commit is contained in:
bnellnm
2026-04-21 18:12:12 -04:00
committed by GitHub
parent 1842447c09
commit 5e584ce9ec
33 changed files with 112 additions and 141 deletions
+4 -4
View File
@@ -35,7 +35,7 @@ from vllm.distributed import (
get_tensor_model_parallel_world_size,
)
from vllm.model_executor.layers.activation import SiluAndMul
from vllm.model_executor.layers.fused_moe import SharedFusedMoE
from vllm.model_executor.layers.fused_moe import FusedMoE
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.linear import (
ColumnParallelLinear,
@@ -335,7 +335,7 @@ class SarvamMLAMoE(nn.Module):
else:
self.shared_experts = None
self.experts = SharedFusedMoE(
self.experts = FusedMoE(
shared_experts=self.shared_experts,
num_experts=self.num_experts,
top_k=self.top_k,
@@ -352,7 +352,7 @@ class SarvamMLAMoE(nn.Module):
routed_scaling_factor=self.routed_scaling_factor,
)
def maybe_get_fused_moe(self) -> SharedFusedMoE:
def maybe_get_fused_moe(self) -> FusedMoE:
return self.experts
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
@@ -529,7 +529,7 @@ class SarvamMLAModel(nn.Module):
return hidden_states
def get_expert_mapping(self) -> list[tuple[str, str, int, str]]:
return SharedFusedMoE.make_expert_params_mapping(
return FusedMoE.make_expert_params_mapping(
self,
ckpt_gate_proj_name="gate_proj",
ckpt_down_proj_name="down_proj",