app.py

import spaces
import gradio as gr
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, TextIteratorStreamer
from threading import Thread

# Remove GPU decorator since we are CPU-only
def predict(message, history):
    # Load model and tokenizer on CPU
    model_id = "kurakurai/Luth-0.6B-Instruct"
    tokenizer = AutoTokenizer.from_pretrained(model_id)
    model = AutoModelForCausalLM.from_pretrained(
        model_id,
        device_map="cpu",       # CPU only
        torch_dtype=torch.float16,
        trust_remote_code=True,
        load_in_4bit=False      # 4-bit quantization not supported on CPU
    )

    # Format conversation history for chat template
    messages = [{"role": "user" if i % 2 == 0 else "assistant", "content": msg} 
                for conv in history for i, msg in enumerate(conv) if msg]
    messages.append({"role": "user", "content": message})
    
    # Apply chat template
    input_ids = tokenizer.apply_chat_template(
        messages,
        add_generation_prompt=True,
        return_tensors="pt",
        tokenize=True
    ).to('cpu')  # CPU device
    
    # Setup streamer for real-time output
    streamer = TextIteratorStreamer(tokenizer, timeout=10., skip_prompt=True, skip_special_tokens=True)
    
    # Generation parameters
    generate_kwargs = dict(
        input_ids=input_ids,
        streamer=streamer,
        max_new_tokens=256,
        do_sample=True,
        temperature=0.3,
        min_p=0.15,
        repetition_penalty=1.05,
        pad_token_id=tokenizer.eos_token_id
    )
    
    # Start generation in separate thread
    t = Thread(target=model.generate, kwargs=generate_kwargs)
    t.start()
    
    # Stream tokens
    partial_message = ""
    for new_token in streamer:
        partial_message += new_token
        yield partial_message

# Setup Gradio interface
gr.ChatInterface(
    predict,
    description="""
    <center><h2>Kurakura AI Luth-0.6B-Instruct Chat</h2></center>
    
    Chat with [Luth-0.6B-Instruct](https://huggingface.co/kurakurai/Luth-0.6B-Instruct), a French-tuned version of Qwen3-0.6B.
    """,
    examples=[
        "Peux-tu résoudre l'équation 3x - 7 = 11 pour x ?",
        "Explique la photosynthèse en termes simples.",
        "Écris un petit poème sur l'intelligence artificielle."
    ],
    theme=gr.themes.Soft(primary_hue="purple"),
).launch()