app.py

import os
import threading
from collections import defaultdict

import gradio as gr
from transformers import (
    AutoModelForCausalLM,
    AutoTokenizer,
    TextIteratorStreamer,
)

# Define model paths
model_name_to_path = {
    "LeCarnet-3M": "MaxLSB/LeCarnet-3M",
    "LeCarnet-8M": "MaxLSB/LeCarnet-8M",
    "LeCarnet-21M": "MaxLSB/LeCarnet-21M",
}

# Load Hugging Face token
hf_token = os.environ["HUGGINGFACEHUB_API_TOKEN"]

# Preload models and tokenizers
loaded_models = defaultdict(dict)

for name, path in model_name_to_path.items():
    loaded_models[name]["tokenizer"] = AutoTokenizer.from_pretrained(path, token=hf_token)
    loaded_models[name]["model"] = AutoModelForCausalLM.from_pretrained(path, token=hf_token)
    loaded_models[name]["model"].eval()

def respond(message, history, model_name, max_tokens, temperature, top_p):
    """
    Generate a response from the selected model, streaming the output and updating chat history.
    
    Args:
        message (str): User's input message.
        history (list): Current chat history as list of (user_msg, bot_msg) tuples.
        model_name (str): Selected model name.
        max_tokens (int): Maximum number of tokens to generate.
        temperature (float): Sampling temperature.
        top_p (float): Top-p sampling parameter.
    
    Yields:
        list: Updated chat history with the user's message and streaming bot response.
    """
    # Append user's message to history with an empty bot response
    history = history + [(message, "")]
    yield history  # Display user's message immediately

    # Select tokenizer and model
    tokenizer = loaded_models[model_name]["tokenizer"]
    model = loaded_models[model_name]["model"]

    # Tokenize input
    inputs = tokenizer(message, return_tensors="pt")

    # Set up streaming
    streamer = TextIteratorStreamer(
        tokenizer,
        skip_prompt=False,
        skip_special_tokens=True,
    )

    # Configure generation parameters
    generate_kwargs = dict(
        **inputs,
        streamer=streamer,
        max_new_tokens=max_tokens,
        do_sample=True,
        temperature=temperature,
        top_p=top_p,
        eos_token_id=tokenizer.eos_token_id,
    )

    # Start generation in a background thread
    thread = threading.Thread(target=model.generate, kwargs=generate_kwargs)
    thread.start()

    # Stream the response with model name prefix
    accumulated = f"**{model_name}:** "
    for new_text in streamer:
        accumulated += new_text
        history[-1] = (message, accumulated)
        yield history

def submit(message, history, model_name, max_tokens, temperature, top_p):
    """
    Handle form submission by calling respond and clearing the input box.
    
    Args:
        message (str): User's input message.
        history (list): Current chat history.
        model_name (str): Selected model name.
        max_tokens (int): Max tokens parameter.
        temperature (float): Temperature parameter.
        top_p (float): Top-p parameter.
    
    Yields:
        tuple: (updated chat history, cleared user input)
    """
    for updated_history in respond(message, history, model_name, max_tokens, temperature, top_p):
        yield updated_history, ""

# Create the Gradio interface with Blocks
with gr.Blocks() as demo:
    # Title and description
    gr.Markdown("# LeCarnet")
    gr.Markdown("Select a model on the right and type a message to chat.")

    # Two-column layout
    with gr.Row():
        # Left column: Chat interface
        with gr.Column():
            chatbot = gr.Chatbot(
                avatar_images=(None, "media/le-carnet.png"),  # User avatar: None, Bot avatar: Logo
                label="Chat"
            )
            user_input = gr.Textbox(placeholder="Type your message here...", label="Message")
            submit_btn = gr.Button("Send")

        # Right column: Model selection and parameters
        with gr.Column():
            model_dropdown = gr.Dropdown(
                choices=["LeCarnet-3M", "LeCarnet-8M", "LeCarnet-21M"],
                value="LeCarnet-8M",
                label="Model"
            )
            max_tokens = gr.Slider(1, 512, value=512, step=1, label="Max New Tokens")
            temperature = gr.Slider(0.1, 2.0, value=0.7, step=0.1, label="Temperature")
            top_p = gr.Slider(0.1, 1.0, value=0.9, step=0.05, label="Top-p")

    # Example prompts
    examples = gr.Examples(
        examples=[
            ["Il était une fois un petit garçon qui vivait dans un village paisible."],
            ["Il était une fois une grenouille qui rêvait de toucher les étoiles chaque nuit depuis son étang."],
            ["Il était une fois un petit lapin perdu"],
        ],
        inputs=user_input,
    )

    # Event handling for submit button
    submit_btn.click(
        fn=submit,
        inputs=[user_input, chatbot, model_dropdown, max_tokens, temperature, top_p],
        outputs=[chatbot, user_input],
    )

if __name__ == "__main__":
    demo.queue(default_concurrency_limit=10, max_size=10).launch(ssr_mode=False, max_threads=10)