app.py

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

# Define your models
MODEL_PATHS = {
    "LeCarnet-3M": "MaxLSB/LeCarnet-3M",
    "LeCarnet-8M": "MaxLSB/LeCarnet-8M",
    "LeCarnet-21M": "MaxLSB/LeCarnet-21M",
}

# Add your Hugging Face token
hf_token = os.environ.get("HUGGINGFACEHUB_API_TOKEN")
if not hf_token:
    raise ValueError("HUGGINGFACEHUB_API_TOKEN environment variable not set.")

# Load tokenizers & models - only load one initially
tokenizer = None
model = None

def load_model(model_name: str):
    """Loads the specified model and tokenizer."""
    global tokenizer, model
    if model_name not in MODEL_PATHS:
        raise ValueError(f"Unknown model: {model_name}")

    print(f"Loading {model_name}...")
    repo = MODEL_PATHS[model_name]
    tokenizer = AutoTokenizer.from_pretrained(repo, use_auth_token=hf_token)
    model = AutoModelForCausalLM.from_pretrained(repo, use_auth_token=hf_token)
    model.eval()
    print(f"{model_name} loaded.")

# Initial model load
initial_model = list(MODEL_PATHS.keys())[0]
load_model(initial_model)


def respond(
    prompt: str,
    chat_history: list,
    model_choice: str,
    max_tokens: int,
    temperature: float,
    top_p: float,
):
    global tokenizer, model

    # Reload model if it's not the currently loaded one
    current_path = getattr(model.config, "_name_or_path", None)
    desired_path = MODEL_PATHS[model_choice]
    if current_path != desired_path:
        load_model(model_choice)

    # Tokenize
    inputs = tokenizer(prompt, return_tensors="pt")
    streamer = TextIteratorStreamer(
        tokenizer,
        skip_prompt=False,
        skip_special_tokens=True,
    )

    # Prepare generation kwargs
    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,
    )

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

    # Stream back to the UI
    accumulated = ""
    for new_text in streamer:
        accumulated += new_text
        yield accumulated


# If you have custom CSS, define it here; otherwise set to None or remove the css= line below
custom_css = None

with gr.Blocks(css=custom_css, fill_width=True) as demo:
    with gr.Row():
        with gr.Column(scale=1):
            model_dropdown = gr.Dropdown(
                choices=list(MODEL_PATHS.keys()),
                value=initial_model,
                label="Choose Model",
                interactive=True
            )
            max_tokens_slider = gr.Slider(
                minimum=1, maximum=512, value=512, step=1, label="Max new tokens"
            )
            temperature_slider = gr.Slider(
                minimum=0.1, maximum=2.0, value=0.7, step=0.1, label="Temperature"
            )
            top_p_slider = gr.Slider(
                minimum=0.1, maximum=1.0, value=0.9, step=0.05, label="Top‑p"
            )

        with gr.Column(scale=3):
            chatbot = gr.ChatInterface(
                fn=respond,
                additional_inputs=[
                    model_dropdown,
                    max_tokens_slider,
                    temperature_slider,
                    top_p_slider,
                ],
                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"],
                ],
                cache_examples=False,
                submit_btn="Generate",
                avatar_images=(None, "media/le-carnet.png")
            )

if __name__ == "__main__":
    demo.queue()
    demo.launch()