[WhisperModel] fix bug in reshaping labels (#21653)

fix bug in reshaping labels

src/transformers/models/whisper/modeling_whisper.py

@@ -1211,7 +1211,7 @@ class WhisperForConditionalGeneration(WhisperPreTrainedModel):
         loss = None
         if labels is not None:
             loss_fct = CrossEntropyLoss()
-            loss = loss_fct(lm_logits.view(-1, self.config.vocab_size), labels.view(-1))
+            loss = loss_fct(lm_logits.view(-1, self.config.vocab_size), labels.reshape(-1))
 
         if not return_dict:
             output = (lm_logits,) + outputs[1:]