[MoE] Use GateLinear for all MoE models (#58234)

Signed-off-by: Jee Jee Li <jeejeelee@inferact.ai>
This commit is contained in:
Jee Jee Li
2026-09-23 05:40:13 +00:00
committed by GitHub
parent e975732152
commit de6674b002
34 changed files with 221 additions and 225 deletions
@@ -17,7 +17,10 @@ from vllm.distributed import (
tensor_model_parallel_all_gather,
tensor_model_parallel_reduce_scatter,
)
from vllm.model_executor.layers.fused_moe import FusedMoEFactory
from vllm.model_executor.layers.fused_moe import (
FusedMoEFactory,
GateLinear,
)
from vllm.model_executor.layers.linear import ReplicatedLinear
from vllm.model_executor.layers.logits_processor import LogitsProcessor
from vllm.model_executor.layers.mamba.gdn.qwen_gdn_linear_attn import (
@@ -93,11 +96,9 @@ class InternS2MobiusMetaMoeBlock(nn.Module):
self.n_physical_experts = self.n_logical_experts + self.n_redundant_experts
self.n_local_physical_experts = self.n_physical_experts // self.ep_size
self.gate = ReplicatedLinear(
self.gate = GateLinear(
config.hidden_size,
config.num_experts,
bias=False,
quant_config=None,
prefix=f"{prefix}.gate",
)
self.experts = FusedMoEFactory(