[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
+2 -3
View File
@@ -20,6 +20,7 @@ from vllm.model_executor.layers.activation import SiluAndMul
from vllm.model_executor.layers.attention import Attention
from vllm.model_executor.layers.fused_moe import (
FusedMoEFactory,
GateLinear,
)
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.linear import (
@@ -77,11 +78,9 @@ class FusedMoEBlock(nn.Module):
quant_config=quant_config,
prefix=f"{prefix}.experts",
)
self.gate = ReplicatedLinear(
self.gate = GateLinear(
config.hidden_size,
config.moe_num_experts,
bias=False,
quant_config=None,
prefix=f"{prefix}.gate",
)