app.py

import torch
import gradio as gr
from PIL import Image
from transformers import AutoModel, AutoTokenizer

# Initialize model and tokenizer
model = AutoModel.from_pretrained('openbmb/MiniCPM-V-4_5', trust_remote_code=True,
    attn_implementation='sdpa', torch_dtype=torch.bfloat16)
model = model.eval().cuda()
tokenizer = AutoTokenizer.from_pretrained('openbmb/MiniCPM-V-4_5', trust_remote_code=True)

# Default chat parameters
enable_thinking = False

def chat_with_model(image, question, history):
    """
    Chat with the MiniCPM model.
    
    Args:
        image: PIL Image or None
        question: str, user's text question
        history: list of previous conversation turns
    
    Returns:
        str: model's response
        list: updated conversation history
    """
    # Build messages with history context
    msgs = []
    for h in history:
        msgs.append({"role": "user", "content": h[0]})
        msgs.append({"role": "assistant", "content": h[1]})
    
    # Add current user message
    if image is not None:
        msgs.append({"role": "user", "content": [image, question]})
    else:
        msgs.append({"role": "user", "content": question})
    
    # Generate model response
    answer = model.chat(
        msgs=msgs,
        tokenizer=tokenizer,
        enable_thinking=enable_thinking
    )
    
    # Update history
    history.append((question if image is None else [image, question], answer))
    return answer, history

# Create Gradio interface
with gr.Blocks() as demo:
    gr.Markdown("# MiniCPM-V-4.5 Chat Interface")
    gr.Markdown("Upload an image and ask questions, or chat without an image")
    
    # Store conversation history
    chat_history = gr.State([])
    
    with gr.Row():
        with gr.Column():
            image_input = gr.Image(type="pil", label="Upload Image (optional)")
            question_input = gr.Textbox(label="Your Question", placeholder="Enter your question here...")
            submit_btn = gr.Button("Submit")
            clear_btn = gr.Button("Clear History")
        
        with gr.Column():
            response_output = gr.Textbox(label="Model Response", interactive=False)
            chat_display = gr.Chatbot(label="Conversation History")
    
    # Handle submit action
    submit_btn.click(
        fn=chat_with_model,
        inputs=[image_input, question_input, chat_history],
        outputs=[response_output, chat_history]
    ).then(
        fn=lambda history: history,
        inputs=[chat_history],
        outputs=[chat_display]
    )
    
    # Clear history
    def clear_history():
        return [], [], ""
    
    clear_btn.click(
        fn=clear_history,
        outputs=[chat_history, chat_display, response_output]
    )

# Launch demo
if __name__ == "__main__":
    demo.launch()