app.py

import gradio as gr
from transformers import Qwen2AudioForConditionalGeneration, AutoProcessor
import librosa

import spaces


MODEL_ID = "Qwen/Qwen2-Audio-7B-Instruct"

processor = AutoProcessor.from_pretrained(MODEL_ID)
model = Qwen2AudioForConditionalGeneration.from_pretrained(MODEL_ID, device_map="auto")


@spaces.GPU
def run_qwen2audio(audio_path: str, instruction: str) -> str:
    if not audio_path:
        return "Please upload an audio file."

    conversation = [
        {
            "role": "user",
            "content": [
                {"type": "audio", "audio_url": audio_path},
                {"type": "text", "text": instruction},
            ],
        }
    ]

    text = processor.apply_chat_template(conversation, add_generation_prompt=True, tokenize=False)

    audios = []
    target_sr = processor.feature_extractor.sampling_rate
    audio, _ = librosa.load(audio_path, sr=target_sr)
    audios.append(audio)

    inputs = processor(text=text, audio=audios, return_tensors="pt", padding=True)
    inputs = inputs.to(model.device)

    output_ids = model.generate(**inputs, max_new_tokens=4096)
    output_ids = output_ids[:, inputs.input_ids.size(1):]
    response = processor.batch_decode(output_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
    return response


with gr.Blocks(title="Qwen2-Audio Demo") as demo:
    gr.Markdown("# Qwen2-Audio Demo")
    gr.Markdown("Upload audio and run an instruction with Qwen2-Audio.")

    with gr.Row():
        with gr.Column():
            audio_input = gr.Audio(type="filepath", label="Upload Audio")
            instruction = gr.Textbox(label="Instruction", value="Transcribe the audio.")
            submit_btn = gr.Button("Run", variant="primary")
        with gr.Column():
            output_text = gr.Textbox(label="Response", lines=12)

    submit_btn.click(run_qwen2audio, [audio_input, instruction], output_text)


if __name__ == "__main__":
    demo.queue().launch(share=False, ssr_mode=False)