[Kernel] Support DCP for Triton backend (#25132)
Signed-off-by: Wei Wei <wwei6@meta.com>
This commit is contained in:
@@ -685,7 +685,7 @@ class DeepseekV2DecoderLayer(nn.Module):
|
||||
) -> torch.Tensor:
|
||||
# Self Attention
|
||||
if residual is None:
|
||||
residual = hidden_states
|
||||
residual = hidden_states.clone()
|
||||
hidden_states = self.input_layernorm(hidden_states)
|
||||
else:
|
||||
hidden_states, residual = self.input_layernorm(
|
||||
|
||||
Reference in New Issue
Block a user