test_shape.py

import torch
import warnings
warnings.filterwarnings("ignore")

# Let's check what generation returns for an empty input
from transformers import WhisperForConditionalGeneration, WhisperProcessor

model_name_or_path = "openai/whisper-large-v3"
processor = WhisperProcessor.from_pretrained(model_name_or_path)
model = WhisperForConditionalGeneration.from_pretrained(model_name_or_path, torch_dtype=torch.float16).to("cuda")

inputs = torch.zeros(1, 128, 3000, dtype=torch.float16, device="cuda")

out = model.generate(inputs, return_timestamps=True, num_beams=1)
print("Output shape:", out.shape)
print("Output type:", type(out))
out_list = out.cpu().numpy().tolist()
print("Output list type:", type(out_list))
if len(out_list) > 0 and isinstance(out_list[0], list):
    print("Depth 2:", type(out_list[0][0]))

try:
    processor.tokenizer.batch_decode(out_list, skip_special_tokens=False, decode_with_timestamps=True)
    print("Batch decode successful")
except Exception as e:
    import traceback
    traceback.print_exc()