Whisper: move to tensor cpu before converting to np array at decode time (#31954)

src/transformers/models/whisper/tokenization_whisper.py

@@ -872,7 +872,10 @@ class WhisperTokenizer(PreTrainedTokenizer):
     @staticmethod
     def _convert_to_list(token_ids):
         # convert type to ndarray if necessary
-        if "torch" in str(type(token_ids)) or "tensorflow" in str(type(token_ids)) and hasattr(token_ids, "numpy"):
+        if hasattr(token_ids, "numpy"):
+            if "torch" in str(type(token_ids)):
+                token_ids = token_ids.cpu().numpy()
+            elif "tensorflow" in str(type(token_ids)):
                 token_ids = token_ids.numpy()
         # now the token ids are either a numpy array, or a list of lists
         if isinstance(token_ids, np.ndarray):

src/transformers/models/whisper/tokenization_whisper_fast.py

@@ -605,7 +605,10 @@ class WhisperTokenizerFast(PreTrainedTokenizerFast):
     # Copied from transformers.models.whisper.tokenization_whisper.WhisperTokenizer._convert_to_list
     def _convert_to_list(token_ids):
         # convert type to ndarray if necessary
-        if "torch" in str(type(token_ids)) or "tensorflow" in str(type(token_ids)) and hasattr(token_ids, "numpy"):
+        if hasattr(token_ids, "numpy"):
+            if "torch" in str(type(token_ids)):
+                token_ids = token_ids.cpu().numpy()
+            elif "tensorflow" in str(type(token_ids)):
                 token_ids = token_ids.numpy()
         # now the token ids are either a numpy array, or a list of lists
         if isinstance(token_ids, np.ndarray):