app.py

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

MODEL_NAME = "MaxLSB/LeCarnet-8M"

# Load tokenizer & model locally
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
model = AutoModelForCausalLM.from_pretrained(MODEL_NAME)
model.eval()

def respond(
    prompt: str,
    chat_history,
    max_tokens: int,
    temperature: float,
    top_p: float,
):
    inputs = tokenizer(prompt, return_tensors="pt")

    # Text streamer to get one token at a time
    streamer = TextIteratorStreamer(
        tokenizer,
        skip_prompt=True,
        skip_special_tokens=True,
    )

    generate_kwargs = dict(
        **inputs,
        streamer=streamer,
        max_new_tokens=max_tokens,
        do_sample=True,
        temperature=temperature,
        top_p=top_p,
    )

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

    # Stream out partial completions
    accumulated = ""
    for new_text in streamer:
        accumulated += new_text
        yield accumulated

# Wire it up in Gradio
demo = gr.ChatInterface(
    fn=respond,
    additional_inputs=[
        gr.Slider(1, 512, value=128, step=1, label="Max new tokens"),
        gr.Slider(0.1, 2.0, value=0.7, step=0.1, label="Temperature"),
        gr.Slider(0.1, 1.0, value=0.9, step=0.05, label="Top‑p"),
    ],
    title="Prefix Completion Demo",
    description="Type the beginning of a sentence and watch the model finish it.",
)

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