[Whisper Tok] Move token ids to CPU when computing offsets (#28485)

* move token ids to cpu

* check for torch attr

src/transformers/models/whisper/tokenization_whisper.py

@@ -553,6 +553,9 @@ class WhisperTokenizer(PreTrainedTokenizer):
                 The time ratio to convert from token to time.
         """
         offsets = []
+        # ensure torch tensor of token ids is placed on cpu
+        if "torch" in str(type(token_ids)) and (hasattr(token_ids, "cpu") and callable(token_ids.cpu)):
+            token_ids = token_ids.cpu()
         token_ids = np.array(token_ids)
         if token_ids.shape[0] > 1 and len(token_ids.shape) > 1:
             raise ValueError("Can only process a single input at a time")

src/transformers/models/whisper/tokenization_whisper_fast.py

@@ -248,6 +248,9 @@ class WhisperTokenizerFast(PreTrainedTokenizerFast):
                 The time ratio to convert from token to time.
         """
         offsets = []
+        # ensure torch tensor of token ids is placed on cpu
+        if "torch" in str(type(token_ids)) and (hasattr(token_ids, "cpu") and callable(token_ids.cpu)):
+            token_ids = token_ids.cpu()
         token_ids = np.array(token_ids)
         if token_ids.shape[0] > 1 and len(token_ids.shape) > 1:
             raise ValueError("Can only process a single input at a time")