back move ffn.

Author: wanghao7

lightllm/models/deepseek2/layer_infer/transformer_layer_infer.py

@@ -759,10 +759,6 @@ def overlap_tpsp_context_forward(
 
         _0_overlap_event = Buffer.capture()
 
-        # 0 shared expert
-        if self.n_shared_experts is not None:
-            _0_shared_output = LlamaTransformerLayerInfer._ffn(self, _0_input1, infer_state, layer_weight)
-
         # 1 attention
         _1_input1 = self._att_norm(input_embdings1, infer_state1, layer_weight)
         _1_cache_kv = self._pre_cache_kv(infer_state1, layer_weight)
@@ -801,6 +797,10 @@ def overlap_tpsp_context_forward(
 
         _1_overlap_event = Buffer.capture()
 
+        # 0 shared expert
+        if self.n_shared_experts is not None:
+            _0_shared_output = LlamaTransformerLayerInfer._ffn(self, _0_input1, infer_state, layer_weight)
+
         # 1 shared expert
         if self.n_shared_experts is not None:
             _1_shared_output = LlamaTransformerLayerInfer._ffn(self, _1_input1, infer_state1, layer_weight)