[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
+6 -4
View File
@@ -28,12 +28,15 @@ from vllm.model_executor.layers.attention import Attention
from vllm.model_executor.layers.attention.mm_encoder_attention import (
MMEncoderAttention,
)
from vllm.model_executor.layers.fused_moe import MoEActivation, fused_experts
from vllm.model_executor.layers.fused_moe import (
GateLinear,
MoEActivation,
fused_experts,
)
from vllm.model_executor.layers.fused_moe.config import biased_moe_quant_config
from vllm.model_executor.layers.linear import (
ColumnParallelLinear,
QKVParallelLinear,
ReplicatedLinear,
RowParallelLinear,
)
from vllm.model_executor.layers.logits_processor import LogitsProcessor
@@ -453,11 +456,10 @@ class Moondream3TextMoE(nn.Module):
self.num_local_experts = self.experts_per_rank
# Router (gate) - use ReplicatedLinear for compatibility
self.gate = ReplicatedLinear(
self.gate = GateLinear(
hidden_size,
num_experts,
bias=True,
quant_config=None,
prefix=f"{prefix}.gate",
)