@tduncan / Nvidia-Optimization-Stack-VLLM / commits / 8b4d93ba2b

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

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

Wentao Ye committed Jun 25, 2026 at 16:09 UTC 8b4d93ba2b4e4e79062fd59a35a623e381dcf6b0

4 files changed +4 -4

vllm/model_executor/models/AXK1.py

+1 -1

@@ -649,7 +649,7 @@ class AXK1DecoderLayer(nn.Module):

649 ) -> tuple[torch.Tensor, torch.Tensor]:

650 # Self Attention

651 if residual is None:

652 - residual = hidden_states.clone()

652 + residual = hidden_states

653 hidden_states = self.input_layernorm(hidden_states)

654 else:

655 hidden_states, residual = self.input_layernorm(hidden_states, residual)

vllm/model_executor/models/deepseek_v2.py

+1 -1

@@ -1186,7 +1186,7 @@ class DeepseekV2DecoderLayer(nn.Module):

1186 ) -> torch.Tensor:

1187 # Self Attention

1188 if residual is None:

1189 - residual = hidden_states.clone()

1189 + residual = hidden_states

1190 hidden_states = self.input_layernorm(hidden_states)

1191 else:

1192 hidden_states, residual = self.input_layernorm(hidden_states, residual)

vllm/model_executor/models/glm4_moe_lite.py

+1 -1

@@ -184,7 +184,7 @@ class Glm4MoeLiteDecoderLayer(nn.Module):

184 ) -> torch.Tensor:

185 # Self Attention

186 if residual is None:

187 - residual = hidden_states.clone()

187 + residual = hidden_states

188 hidden_states = self.input_layernorm(hidden_states)

189 else:

190 hidden_states, residual = self.input_layernorm(hidden_states, residual)

vllm/model_executor/models/openpangu.py

+1 -1

@@ -935,7 +935,7 @@ class OpenPanguDecoderLayer(nn.Module):

935 residual: torch.Tensor | None,

936 ) -> torch.Tensor:

937 if residual is None:

938 - residual = hidden_states.clone()

938 + residual = hidden_states

939 hidden_states = self.input_layernorm(hidden_states)

940 else:

941 hidden_states, residual = self.input_layernorm(hidden_states, residual)