@randombk This vLLM version is incredible! I'm getting like 16-18x speedup in batch mode! Thank you so much for making it! I have no idea how you did it, but it's amazing!

• Fix gradio passing overrides to tts.py
• Add diffusion steps to gradio with default 10

I found the audio quality of the vLLM to be appreciably worse than the original and went hunting for the reason. I saw the diffusion steps were set down and set them up and that fixed it! So I'd recommend defaulting to 10. vLLM is still loads faster than the original with 10 steps, so I think letting people manually set down the quality for even more speed is better.

Also fixes errors like

  File "/home/david/chatterbox-vllm/gradio_tts_app.py", line 29, in load_model
    global_model = ChatterboxTTS.from_pretrained(
                   ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/david/chatterbox-vllm/src/chatterbox_vllm/tts.py", line 165, in from_pretrained
    return cls.from_local(Path(local_path).parent, *args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/david/chatterbox-vllm/src/chatterbox_vllm/tts.py", line 123, in from_local
    t3 = LLM(
         ^^^^
TypeError: vllm.entrypoints.llm.LLM() got multiple values for keyword argument 'gpu_memory_utilization'