[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
+5 -5
View File
@@ -35,10 +35,12 @@ from vllm.compilation.decorators import support_torch_compile
from vllm.config import CacheConfig, VllmConfig
from vllm.distributed import get_pp_group, get_tensor_model_parallel_world_size
from vllm.model_executor.layers.attention import Attention
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 (
QKVParallelLinear,
ReplicatedLinear,
RowParallelLinear,
)
from vllm.model_executor.layers.logits_processor import LogitsProcessor
@@ -262,12 +264,10 @@ class PhiMoE(nn.Module):
self.hidden_size = hidden_size
# Gate always runs at half / full precision for now.
self.gate = ReplicatedLinear(
self.gate = GateLinear(
hidden_size,
num_experts,
bias=False,
params_dtype=params_dtype,
quant_config=None,
prefix=f"{prefix}.gate",
)