Fix pad_to_max_length Whisper (#30787)

* fix pad_to_max_length Whisper

* add tests

* make style

src/transformers/models/whisper/generation_whisper.py

@@ -122,7 +122,9 @@ def _get_attr_from_logit_processors(logits_processor, logit_processor_class, att
     return None
 
 
-def _pad_to_max_length(current_segments, pad_token_id, padding="right", bos_token_tensor=None, cut_off_length=None):
+def _pad_to_max_length(
+    current_segments, pad_token_id, device, padding="right", bos_token_tensor=None, cut_off_length=None
+):
     max_total_length = 0
     sequences = []
     if padding not in ["right", "left"]:
@@ -143,7 +145,7 @@ def _pad_to_max_length(current_segments, pad_token_id, padding="right", bos_toke
         elif bos_token_tensor is not None:
             sequences.append(bos_token_tensor)
         else:
-            sequences.append(torch.tensor([]))
+            sequences.append(torch.tensor([], device=device))
 
     for i in range(len(current_segments)):
         pad_length = max_total_length - len(sequences[i])
@@ -733,7 +735,9 @@ class WhisperGenerationMixin:
             if (prompt_ids is not None and generation_config.prompt_condition_type == "first-segment")
             else current_segments
         )
-        sequences = _pad_to_max_length(final_segments, generation_config.pad_token_id, padding="right")
+        sequences = _pad_to_max_length(
+            final_segments, generation_config.pad_token_id, device=self.device, padding="right"
+        )
 
         # 8. If we return all segments, the predicted output sequences are put under `"sequences"`.
         if return_segments:
@@ -1506,6 +1510,7 @@ class WhisperGenerationMixin:
             prev_tokens = _pad_to_max_length(
                 active_segments,
                 generation_config.pad_token_id,
+                device=device,
                 padding="left",
                 bos_token_tensor=prev_ids,
                 cut_off_length=cut_off_length,

tests/models/whisper/test_modeling_whisper.py

@@ -35,6 +35,7 @@ from transformers.testing_utils import (
     require_torch,
     require_torch_fp16,
     require_torch_gpu,
+    require_torch_multi_gpu,
     require_torchaudio,
     slow,
     torch_device,
@@ -2866,6 +2867,81 @@ class WhisperModelIntegrationTests(unittest.TestCase):
         for i in range(num_samples):
             assert decoded_all[i] == EXPECTED_TEXT[i]
 
+    @require_torch_gpu
+    @slow
+    def test_whisper_empty_longform(self):
+        processor = WhisperProcessor.from_pretrained("openai/whisper-tiny")
+        model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-tiny")
+        model = model.to(torch_device)
+
+        ds = load_dataset("distil-whisper/meanwhile", "default")["test"]
+        ds = ds.cast_column("audio", Audio(sampling_rate=16000))
+
+        num_samples = 8
+
+        audio = ds[:num_samples]["audio"]
+        audios = [x["array"] for x in audio]
+        audios[0][:] = np.zeros(audios[0].shape)
+
+        inputs = processor(
+            audios,
+            return_tensors="pt",
+            truncation=False,
+            padding="longest",
+            return_attention_mask=True,
+            sampling_rate=16_000,
+        )
+        inputs = inputs.to(device=torch_device)
+
+        gen_kwargs = {
+            "no_speech_threshold": 0.2,
+            "temperature": (0.0,),
+            "logprob_threshold": 0.0,  # Ignore logprob, use only no-speech prob
+            "num_beams": 5,
+            "language": "fr",
+            "task": "transcribe",
+        }
+
+        torch.manual_seed(0)
+        model.generate(**inputs, **gen_kwargs)
+
+    @require_torch_multi_gpu
+    @slow
+    def test_whisper_empty_longform_multi_gpu(self):
+        processor = WhisperProcessor.from_pretrained("openai/whisper-tiny")
+        model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-tiny", device_map="auto")
+
+        ds = load_dataset("distil-whisper/meanwhile", "default")["test"]
+        ds = ds.cast_column("audio", Audio(sampling_rate=16000))
+
+        num_samples = 8
+
+        audio = ds[:num_samples]["audio"]
+        audios = [x["array"] for x in audio]
+        audios[0][:] = np.zeros(audios[0].shape)
+
+        inputs = processor(
+            audios,
+            return_tensors="pt",
+            truncation=False,
+            padding="longest",
+            return_attention_mask=True,
+            sampling_rate=16_000,
+        )
+        inputs = inputs.to(device=model.device)
+
+        gen_kwargs = {
+            "no_speech_threshold": 0.2,
+            "temperature": (0.0,),
+            "logprob_threshold": 0.0,  # Ignore logprob, use only no-speech prob
+            "num_beams": 5,
+            "language": "fr",
+            "task": "transcribe",
+        }
+
+        torch.manual_seed(0)
+        model.generate(**inputs, **gen_kwargs)
+
 
 def prepare_whisper_encoder_inputs_dict(config, input_features, head_mask=None):
     if head_mask is None: