[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
@@ -19,6 +19,7 @@ from vllm.logger import init_logger
from vllm.model_executor.layers.attention import Attention
from vllm.model_executor.layers.fused_moe import (
FusedMoEFactory,
GateLinear,
MoERunner,
)
from vllm.model_executor.layers.layernorm import RMSNorm
@@ -82,11 +83,12 @@ class AfmoeMoE(nn.Module):
)
# Router gate
self.gate = nn.Linear(
self.gate = GateLinear(
config.hidden_size,
config.num_experts,
bias=False,
dtype=torch.float32,
out_dtype=torch.float32,
params_dtype=torch.float32,
prefix=f"{prefix}.gate",
)
self.expert_bias = nn.Parameter(
torch.empty(config.num_experts, dtype=torch.float32)
@@ -140,7 +142,7 @@ class AfmoeMoE(nn.Module):
num_tokens, hidden_dim = hidden_states.shape
hidden_states = hidden_states.view(-1, hidden_dim)
router_logits = self.gate(hidden_states.to(dtype=torch.float32))
router_logits, _ = self.gate(hidden_states)
final_hidden_states = self.experts(
hidden_states=hidden_states, router_logits=router_logits