Update after #36962

src/transformers/cache_utils.py

@@ -631,8 +631,10 @@ def __init__(self) -> None:
     def prefetch_layer(self, layer_idx: int):
         "Starts prefetching the next layer cache"
         if layer_idx < len(self):
-            with self.prefetch_stream if is_torch_greater_or_equal("2.7", accept_dev=True) else torch.cuda.stream(
+            with (
                 self.prefetch_stream
+                if is_torch_greater_or_equal("2.7", accept_dev=True)
+                else torch.cuda.stream(self.prefetch_stream)
             ):
                 # Prefetch next layer tensors to GPU
                 device = self.original_device[layer_idx]

src/transformers/models/qwen2_audio/processing_qwen2_audio.py

@@ -129,7 +129,7 @@ def __call__(
         if audio is not None:
             # ensure we have as much audios as audio tokens
             num_audio_tokens = sum(sample.count(self.audio_token) for sample in text)
-            num_audios = 1 if type(audio) == np.ndarray else len(audio)
+            num_audios = 1 if type(audio) is np.ndarray else len(audio)
             if num_audio_tokens != num_audios:
                 raise ValueError(
                     f"Found {num_audio_tokens} {self.audio_token} token{'s' if num_audio_tokens > 1 else ''} in provided text but received {num_audios} audio{'s' if num_audios > 1 else ''}"