[Perf] Remove redundant clone for GLM, Deepseek etc (#46651)
Signed-off-by: yewentao256 <zhyanwentao@126.com>
This commit is contained in:
@@ -649,7 +649,7 @@ class AXK1DecoderLayer(nn.Module):
|
|||||||
) -> tuple[torch.Tensor, torch.Tensor]:
|
) -> tuple[torch.Tensor, torch.Tensor]:
|
||||||
# Self Attention
|
# Self Attention
|
||||||
if residual is None:
|
if residual is None:
|
||||||
residual = hidden_states.clone()
|
residual = hidden_states
|
||||||
hidden_states = self.input_layernorm(hidden_states)
|
hidden_states = self.input_layernorm(hidden_states)
|
||||||
else:
|
else:
|
||||||
hidden_states, residual = self.input_layernorm(hidden_states, residual)
|
hidden_states, residual = self.input_layernorm(hidden_states, residual)
|
||||||
|
|||||||
@@ -1186,7 +1186,7 @@ class DeepseekV2DecoderLayer(nn.Module):
|
|||||||
) -> torch.Tensor:
|
) -> torch.Tensor:
|
||||||
# Self Attention
|
# Self Attention
|
||||||
if residual is None:
|
if residual is None:
|
||||||
residual = hidden_states.clone()
|
residual = hidden_states
|
||||||
hidden_states = self.input_layernorm(hidden_states)
|
hidden_states = self.input_layernorm(hidden_states)
|
||||||
else:
|
else:
|
||||||
hidden_states, residual = self.input_layernorm(hidden_states, residual)
|
hidden_states, residual = self.input_layernorm(hidden_states, residual)
|
||||||
|
|||||||
@@ -184,7 +184,7 @@ class Glm4MoeLiteDecoderLayer(nn.Module):
|
|||||||
) -> torch.Tensor:
|
) -> torch.Tensor:
|
||||||
# Self Attention
|
# Self Attention
|
||||||
if residual is None:
|
if residual is None:
|
||||||
residual = hidden_states.clone()
|
residual = hidden_states
|
||||||
hidden_states = self.input_layernorm(hidden_states)
|
hidden_states = self.input_layernorm(hidden_states)
|
||||||
else:
|
else:
|
||||||
hidden_states, residual = self.input_layernorm(hidden_states, residual)
|
hidden_states, residual = self.input_layernorm(hidden_states, residual)
|
||||||
|
|||||||
@@ -935,7 +935,7 @@ class OpenPanguDecoderLayer(nn.Module):
|
|||||||
residual: torch.Tensor | None,
|
residual: torch.Tensor | None,
|
||||||
) -> torch.Tensor:
|
) -> torch.Tensor:
|
||||||
if residual is None:
|
if residual is None:
|
||||||
residual = hidden_states.clone()
|
residual = hidden_states
|
||||||
hidden_states = self.input_layernorm(hidden_states)
|
hidden_states = self.input_layernorm(hidden_states)
|
||||||
else:
|
else:
|
||||||
hidden_states, residual = self.input_layernorm(hidden_states, residual)
|
hidden_states, residual = self.input_layernorm(hidden_states, residual)
|
||||||
|
|||||||
Reference in New Issue
Block a user