Merge branch 'v0.15.1-dev-wm' into 'v0.15.1-dev'

支持非deepseek模型使用moe_fused_gate

See merge request dcutoolkit/deeplearing/vllm!439

vllm/model_executor/layers/fused_moe/layer.py

@@ -602,9 +602,9 @@ class FusedMoE(CustomOp):
         # moe_fused_gate kernel ensure that num_experts/num_expert_group does not exceed MAX_VPT=32 now. And when kernel can handle MAX_VPT > 32, we can remove this assertion.
         self.use_fused_gate = envs.VLLM_ENABLE_MOE_FUSED_GATE \
             and self.e_score_correction_bias is not None \
-                 and num_expert_group is not None \
-                    and self.global_num_experts // num_expert_group <= 32 \
-                        and is_power_of_two(e_score_correction_bias.shape[0])
+                 and num_expert_group is not None #\
+                    #and self.global_num_experts // num_expert_group <= 32 \
+                    #    and is_power_of_two(e_score_correction_bias.shape[0])
                         
         self.router = create_fused_moe_router(
             top_k=top_k,