[Bugfix] Fix weight loading for Chameleon when TP>1 (#7410)
This commit is contained in:
@@ -399,8 +399,11 @@ class Phi3SmallForCausalLM(nn.Module):
|
||||
def get_decoder(self):
|
||||
return self.model
|
||||
|
||||
def compute_logits(self, hidden_states: torch.Tensor,
|
||||
sampling_metadata: SamplingMetadata) -> torch.Tensor:
|
||||
def compute_logits(
|
||||
self,
|
||||
hidden_states: torch.Tensor,
|
||||
sampling_metadata: SamplingMetadata,
|
||||
) -> Optional[torch.Tensor]:
|
||||
logits = self.logits_processor(self.lm_head, hidden_states,
|
||||
sampling_metadata)
|
||||
if self.dummy_token_indices is not None and logits is not None:
|
||||
|
||||
Reference in New Issue
Block a user