fix dtype/device

app.py

@@ -103,10 +103,11 @@ def transcribe(inputs):
         yield distil_text, distil_runtime_pipeline, text, runtime_pipeline
 
     else:
-        input_features = processor(inputs, sampling_rate=processor.feature_extractor.sampling_rate, return_tensors="pt")
+        input_features = processor(inputs, sampling_rate=processor.feature_extractor.sampling_rate, return_tensors="pt").input_features
+        input_features = input_features.to(device, dtype=torch_dtype)
 
         # Run the generation in a separate thread, so that we can fetch the generated text in a non-blocking way.
-        generation_kwargs = dict(input_features, streamer=streamer, max_new_tokens=128, language="en", task="transcribe")
+        generation_kwargs = dict(input_features=input_features, streamer=streamer, max_new_tokens=128, language="en", task="transcribe")
         thread = Thread(target=distilled_model.generate, kwargs=generation_kwargs)
 
         thread.start()