[Perf] Remove redundant clone for GLM, Deepseek etc (#46651)

Signed-off-by: yewentao256 <zhyanwentao@126.com>

vllm/model_executor/models/AXK1.py

@@ -649,7 +649,7 @@ class AXK1DecoderLayer(nn.Module):
     ) -> tuple[torch.Tensor, torch.Tensor]:
         # Self Attention
         if residual is None:
-            residual = hidden_states.clone()
+            residual = hidden_states
             hidden_states = self.input_layernorm(hidden_states)
         else:
             hidden_states, residual = self.input_layernorm(hidden_states, residual)

vllm/model_executor/models/deepseek_v2.py

@@ -1186,7 +1186,7 @@ class DeepseekV2DecoderLayer(nn.Module):
     ) -> torch.Tensor:
         # Self Attention
         if residual is None:
-            residual = hidden_states.clone()
+            residual = hidden_states
             hidden_states = self.input_layernorm(hidden_states)
         else:
             hidden_states, residual = self.input_layernorm(hidden_states, residual)

vllm/model_executor/models/glm4_moe_lite.py

@@ -184,7 +184,7 @@ class Glm4MoeLiteDecoderLayer(nn.Module):
     ) -> torch.Tensor:
         # Self Attention
         if residual is None:
-            residual = hidden_states.clone()
+            residual = hidden_states
             hidden_states = self.input_layernorm(hidden_states)
         else:
             hidden_states, residual = self.input_layernorm(hidden_states, residual)

vllm/model_executor/models/openpangu.py

@@ -935,7 +935,7 @@ class OpenPanguDecoderLayer(nn.Module):
         residual: torch.Tensor | None,
     ) -> torch.Tensor:
         if residual is None:
-            residual = hidden_states.clone()
+            residual = hidden_states
             hidden_states = self.input_layernorm(hidden_states)
         else:
             hidden_states, residual = self.input_layernorm(hidden_states, residual)