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

fix: 修复MOE量化tensor对于其他模型的影响

See merge request dcutoolkit/deeplearing/vllm!500

vllm/model_executor/layers/fused_moe/layer.py

@@ -1721,11 +1721,16 @@ class FusedMoE(CustomOp):
                     hidden_states, router_logits
                 )
             else:
-                shared_output, fused_output = torch.ops.vllm.moe_forward_shared(
-                    hidden_states, router_logits, encode_layer_name(),
-                    i_q=i_q,
-                    i_s=i_s
-                )
+                if envs.USE_FUSED_RMS_QUANT:
+                    shared_output, fused_output = torch.ops.vllm.moe_forward_shared(
+                        hidden_states, router_logits, encode_layer_name(),
+                        i_q=i_q,
+                        i_s=i_s
+                    )
+                else:
+                    shared_output, fused_output = torch.ops.vllm.moe_forward_shared(
+                        hidden_states, router_logits, encode_layer_name()
+                    )
             return (
                 reduce_output(shared_output)[..., :og_hidden_states],
                 reduce_output(fused_output)[..., :og_hidden_states],
@@ -1976,7 +1981,10 @@ class FusedMoE(CustomOp):
             # because matrix multiply maybe modify the hidden_states.
             if has_separate_shared_experts and not use_shared_experts_stream:
                 assert self.shared_experts is not None
-                shared_output = self.shared_experts(hidden_states, iqis=(i_q, i_s))
+                if envs.USE_FUSED_RMS_QUANT:
+                    shared_output = self.shared_experts(hidden_states, iqis=(i_q, i_s))
+                else:
+                    shared_output = self.shared_experts(hidden_states)
 
             # NOTE: Similar with DP, PCP also needs dispatch and combine. For
             # simplicity, AgRsAll2All was added separately for PCP here. Maybe