app.py

import gradio as gr
import time
from transformers import Qwen2AudioForConditionalGeneration, AutoProcessor
from io import BytesIO
from urllib.request import urlopen
import librosa
import os, json
from sys import argv
from vllm import LLM, SamplingParams

def load_model_processor(model_path):
    processor = AutoProcessor.from_pretrained(model_path)
    llm = LLM(
        model=model_path, trust_remote_code=True, gpu_memory_utilization=0.4,  
        enforce_eager=True,   
        limit_mm_per_prompt={"audio": 5},
    )
    return llm, processor

model_path1 = "Qwen/Qwen2-Audio-7B-Instruct" #argv[1]
model1, processor1 = load_model_processor(model_path1)

def response_to_audio_conv(conversation, model=None, processor=None, temperature = 0.1,repetition_penalty=1.1, top_p = 0.9,
                    max_new_tokens = 2048):
    text = processor.apply_chat_template(conversation, add_generation_prompt=True, tokenize=False)
    audios = []
    for message in conversation:
        if isinstance(message["content"], list):
            for ele in message["content"]:
                if ele["type"] == "audio":
                    if ele['audio_url'] != None:
                        audios.append(librosa.load(
                            ele['audio_url'], 
                            sr=processor.feature_extractor.sampling_rate)[0]
                        )

    sampling_params = SamplingParams(
        temperature=temperature, max_tokens=max_new_tokens, repetition_penalty=repetition_penalty, top_p=top_p, top_k=20,
        stop_token_ids=[],
    )

    input = {
            'prompt': text,
            'multi_modal_data': {
                'audio': [(audio, 16000) for audio in audios]
            }
            }

    output = model.generate([input], sampling_params=sampling_params)[0]
    response = output.outputs[0].text
    return response

def print_like_dislike(x: gr.LikeData):
    print(x.index, x.value, x.liked)

def add_message(history, message):
    paths = []
    for turn in history: 
        if turn['role'] == "user" and type(turn['content']) != str:
            paths.append(turn['content'][0])
    for x in message["files"]:
        if x not in paths:
            history.append({"role": "user", "content": {"path": x}})
    if message["text"] is not None:
        history.append({"role": "user", "content": message["text"]})
    return history, gr.MultimodalTextbox(value=None, interactive=False)

def format_user_messgae(message):
    if type(message['content']) == str:
        return {"role": "user", "content": [{"type": "text", "text": message['content']}]}
    else:
        return {"role": "user", "content": [{"type": "audio", "audio_url": message['content'][0]}]}

def history_to_conversation(history):
    conversation = []
    audio_paths = []
    for turn in history:
        if turn['role'] == "user":
            if not turn['content']: 
                continue
            turn = format_user_messgae(turn)
            if turn['content'][0]['type'] == 'audio':
                if turn['content'][0]['audio_url'] in audio_paths:
                    continue
                else: 
                    audio_paths.append(turn['content'][0]['audio_url'])

            if len(conversation) > 0 and conversation[-1]["role"] == "user":
                conversation[-1]['content'].append(turn['content'][0])
            else:
                conversation.append(turn)
        else:
            conversation.append(turn)
    
    print(json.dumps(conversation, indent=4, ensure_ascii=False))
    return conversation

def bot(history: list, temperature = 0.1,repetition_penalty=1.1, top_p = 0.9,
                    max_new_tokens = 2048):
    conversation = history_to_conversation(history)
    response = response_to_audio_conv(conversation, model=model1, processor=processor1, temperature = temperature,repetition_penalty=repetition_penalty, top_p = top_p, max_new_tokens = max_new_tokens)
    # response = "Nice to meet you!"
    print("Bot:",response)

    history.append({"role": "assistant", "content": ""})
    for character in response:
        history[-1]["content"] += character
        time.sleep(0.01)
        yield history

insturctions = """**Instruction**: there are three input format: 
    1. text: input text message only 
    2. audio: upload audio file or record a voice message
    3. audio + text: record a voice message and input text message"""

with gr.Blocks() as demo:
    # gr.Markdown("""<p align="center"><img src="images/seal_logo.png" style="height: 80px"/><p>""")
    # gr.Image("images/seal_logo.png", elem_id="seal_logo", show_label=False,height=80,show_fullscreen_button=False)
    gr.Markdown(
        """<div style="text-align: center; font-size: 32px; font-weight: bold;">SeaLLMs-Audio ChatBot</div>""",
    )

    # Description text
    gr.Markdown(
        """<div style="text-align: center; font-size: 16px;">
    This WebUI is based on SeaLLMs-Audio-7B-Chat, developed by Alibaba DAMO Academy.<br>
    You can interact with the chatbot in <b>English, Chinese, Indonesian, Thai, or Vietnamese</b>.<br>
    For each round, you can input <b>audio and/or text</b>.
    </div>""",
    )

    # Links with proper formatting
    gr.Markdown(
        """<div style="text-align: center; font-size: 16px;">
        <a href="https://huggingface.co/SeaLLMs/SeaLLMs-v3-7B-Chat">[Website]</a> &nbsp; 
        <a href="https://huggingface.co/SeaLLMs/SeaLLMs-v3-7B-Chat">[Model🤗]</a> &nbsp; 
        <a href="https://github.com/liuchaoqun/SeaLLMs-Audio">[Github]</a>
        </div>""",
    )

    # gr.Markdown(insturctions)
    # with gr.Row():
    #     with gr.Column():
    #         temperature = gr.Slider(minimum=0, maximum=1, value=0.3, step=0.1, label="Temperature")
    #     with gr.Column():
    #         top_p = gr.Slider(minimum=0.1, maximum=1, value=0.5, step=0.1, label="Top P")
    #     with gr.Column():
    #         repetition_penalty = gr.Slider(minimum=0, maximum=2, value=1.1, step=0.1, label="Repetition Penalty")
    chatbot = gr.Chatbot(elem_id="chatbot", bubble_full_width=False, type="messages")

    chat_input = gr.MultimodalTextbox(
        interactive=True,
        file_count="single",
        file_types=['.wav'],
        placeholder="Enter message (optional) ...",
        show_label=False,
        sources=["microphone", "upload"],
    )

    chat_msg = chat_input.submit(
        add_message, [chatbot, chat_input], [chatbot, chat_input]
    )
    bot_msg = chat_msg.then(bot, chatbot, chatbot, api_name="bot_response")
    # bot_msg = chat_msg.then(bot, [chatbot, temperature, repetition_penalty, top_p], chatbot, api_name="bot_response")
    bot_msg.then(lambda: gr.MultimodalTextbox(interactive=True), None, [chat_input])

    # chatbot.like(print_like_dislike, None, None, like_user_message=True)

    clear_button = gr.ClearButton([chatbot, chat_input])
        
# PORT = 7950
# demo.launch(server_port=PORT, show_api = True, allowed_paths = [], 
#     root_path = f"https://dsw-gateway.alibaba-inc.com/dsw81322/proxy/{PORT}/")

demo.launch(
    share=False,
    inbrowser=True,
    server_port=7950,
    server_name="0.0.0.0",
    max_threads=40
)