app.py

import random
import gradio as gr
from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline

### Model Setup ###
# Define the two model names.
MODEL1_NAME = "Qwen/Qwen2.5-14B-Instruct-1M"   # We'll refer to this as Qwen14B.
MODEL2_NAME = "Qwen/Qwen2.5-VL-7B-Instruct"       # We'll refer to this as QwenVL.

# Load Qwen14B model.
tokenizer1 = AutoTokenizer.from_pretrained(MODEL1_NAME)
model1 = AutoModelForCausalLM.from_pretrained(MODEL1_NAME, device_map="auto")
pipe1 = pipeline("text-generation", model=model1, tokenizer=tokenizer1)

# Load QwenVL model.
tokenizer2 = AutoTokenizer.from_pretrained(MODEL2_NAME)
model2 = AutoModelForCausalLM.from_pretrained(MODEL2_NAME, device_map="auto")
pipe2 = pipeline("text-generation", model=model2, tokenizer=tokenizer2)

def generate_response(prompt: str, model_choice: str) -> str:
    """
    Generate a response given the conversation prompt using the chosen model.
    """
    kwargs = {
        "max_length": 256,
        "do_sample": True,
        "top_p": 0.95,
        "temperature": 0.9,
    }
    if model_choice == "Qwen14B":
        result = pipe1(prompt, **kwargs)
    else:  # model_choice == "QwenVL"
        result = pipe2(prompt, **kwargs)
    return result[0]["generated_text"]

def chat_logic(user_input: str, chat_history: list):
    """
    Build the conversation prompt from the history, choose which model responds,
    generate the response, and update the conversation.
    """
    if chat_history is None:
        chat_history = []
    
    # If the user provides input, add it to the history and randomly choose a model.
    if user_input.strip():
        chat_history.append(("User", user_input.strip()))
        selected_model = random.choice(["Qwen14B", "QwenVL"])
    else:
        # When no user input is provided, let the models alternate.
        if not chat_history:
            selected_model = random.choice(["Qwen14B", "QwenVL"])
        else:
            last_speaker = chat_history[-1][0]
            if last_speaker == "Qwen14B":
                selected_model = "QwenVL"
            elif last_speaker == "QwenVL":
                selected_model = "Qwen14B"
            else:
                selected_model = random.choice(["Qwen14B", "QwenVL"])
    
    # Build the prompt from the conversation history.
    prompt = ""
    for speaker, message in chat_history:
        prompt += f"{speaker}: {message}\n"
    prompt += f"{selected_model}:"
    
    # Generate the model's response.
    model_response = generate_response(prompt, selected_model)
    chat_history.append((selected_model, model_response))
    
    # Return the updated conversation for display and to maintain state.
    return chat_history, chat_history

### Gradio Interface ###
with gr.Blocks() as demo:
    gr.Markdown("# Group Chat: Qwen Models")
    gr.Markdown(
        "This demo features two Qwen models conversing with each other. "
        "Leave the textbox blank to let the models converse automatically, "
        "or type a message to interject (a random model will then reply)."
    )
    
    # Chat display component.
    chatbot = gr.Chatbot()
    
    # Input row: textbox and send button.
    with gr.Row():
        user_message = gr.Textbox(
            placeholder="Type your message here or leave blank...",
            show_label=False
        )
        send_btn = gr.Button("Send")
    
    # Maintain the conversation history.
    state = gr.State([])

    # On clicking the button, update the conversation.
    send_btn.click(
        fn=chat_logic,
        inputs=[user_message, state],
        outputs=[chatbot, state]
    )

# Launch the Space.
demo.launch()