[Bugfix] Capture final-layer aux hidden state in deepseek_v2 backbone (#46973)
Signed-off-by: mgoin <mgoin64@gmail.com> Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This commit is contained in:
co-authored by
Claude Opus 4.8
parent
6185d73882
commit
0ca39c4f1f
@@ -1470,6 +1470,9 @@ class DeepseekV2Model(nn.Module):
|
||||
[self.hidden_size, self.hidden_size], dim=-1
|
||||
)
|
||||
|
||||
if self.end_layer in self.aux_hidden_state_layers:
|
||||
aux_hidden_states.append(hidden_states + residual)
|
||||
|
||||
hidden_states, _ = self.norm(hidden_states, residual)
|
||||
if len(aux_hidden_states) > 0:
|
||||
return hidden_states, aux_hidden_states
|
||||
|
||||
Reference in New Issue
Block a user