[bugfix] fix reranker_padding_free (#6989)

Author: Jintao-Huang

swift/plugin/loss.py

@@ -575,7 +575,7 @@ def generative_reranker_loss(outputs,
 
     # Extract logits at the last valid (non-padding) token position for each sample
     batch_size = logits.shape[0]
-    last_valid_indices = get_last_valid_indices(attention_mask)
+    last_valid_indices = -1 if attention_mask is None else get_last_valid_indices(attention_mask)
     batch_indices = torch.arange(batch_size, device=logits.device)
     last_valid_logits = logits[batch_indices, last_valid_indices, :]
 
@@ -743,7 +743,7 @@ def listwise_generative_reranker_loss(outputs,
 
     # Extract logits at the last valid (non-padding) token position for each sample
     batch_size = logits.shape[0]
-    last_valid_indices = get_last_valid_indices(attention_mask)
+    last_valid_indices = -1 if attention_mask is None else get_last_valid_indices(attention_mask)
     batch_indices = torch.arange(batch_size, device=logits.device)
     last_valid_logits = logits[batch_indices, last_valid_indices, :]
 

swift/trainers/trainers.py

@@ -190,7 +190,7 @@ def compute_loss(self, model, inputs, return_outputs=False, num_items_in_batch=N
                     labels,
                     num_items_in_batch=num_items_in_batch,
                     trainer=self,
-                    attention_mask=inputs['attention_mask'])
+                    attention_mask=inputs.get('attention_mask'))
             else:
                 # Fallback to model's loss
                 loss = outputs.loss