app.py

import subprocess
import sys
import os

# Fonction pour installer un package si non présent
def install_package(package_name):
    subprocess.run([sys.executable, "-m", "pip", "install", package_name], check=True)

# Vérifiez si torch est installé, sinon installez-le
try:
    import torch
except ImportError:
    print("Torch n'est pas installé. Installation de torch...")
    install_package("torch")
    import torch

# Vérifiez si transformers est installé, sinon installez-le
try:
    from transformers import (
        AutoModelForCausalLM,
        AutoTokenizer,
        TextIteratorStreamer,
    )
except ImportError:
    print("Transformers n'est pas installé. Installation de transformers...")
    install_package("transformers")
    from transformers import (
        AutoModelForCausalLM,
        AutoTokenizer,
        TextIteratorStreamer,
    )

# Installer flash-attn
subprocess.run(
    "pip install flash-attn --no-build-isolation",
    env={"FLASH_ATTENTION_SKIP_CUDA_BUILD": "TRUE"},
    shell=True,
)

import gradio as gr
from threading import Thread

# Obtenir le token d'authentification Hugging Face
token = os.getenv("HF_TOKEN")
if not token:
    raise ValueError("Le token d'authentification HF_TOKEN n'est pas défini.")

# Charger le modèle et le tokenizer
model = AutoModelForCausalLM.from_pretrained(
    "CampAIgn/Phi-3-mini_16bit",
    token=token,
    trust_remote_code=True,
)
tok = AutoTokenizer.from_pretrained("CampAIgn/Phi-3-mini_16bit", token=token)

terminators = [tok.eos_token_id]

# Vérifier la disponibilité du GPU
if torch.cuda.is_available():
    device = torch.device("cuda")
    print(f"Using GPU: {torch.cuda.get_device_name(device)}")
else:
    device = torch.device("cpu")
    print("Using CPU")

model = model.to(device)

# Fonction de chat
def chat(message, history, temperature, do_sample, max_tokens):
    chat = [{"role": "user", "content": item[0]} for item in history]
    chat.extend({"role": "assistant", "content": item[1]} for item in history if item[1])
    chat.append({"role": "user", "content": message})
    messages = tok.apply_chat_template(chat, tokenize=False, add_generation_prompt=True)
    model_inputs = tok([messages], return_tensors="pt").to(device)
    streamer = TextIteratorStreamer(tok, timeout=20.0, skip_prompt=True, skip_special_tokens=True)
    
    generate_kwargs = {
        "input_ids": model_inputs.input_ids,
        "streamer": streamer,
        "max_new_tokens": max_tokens,
        "do_sample": do_sample,
        "temperature": temperature,
        "eos_token_id": terminators,
    }

    t = Thread(target=model.generate, kwargs=generate_kwargs)
    t.start()

    partial_text = ""
    for new_text in streamer:
        partial_text += new_text
        yield partial_text

    yield partial_text

# Configuration de Gradio
demo = gr.ChatInterface(
    fn=chat,
    examples=[["Write me a poem about Machine Learning."]],
    additional_inputs_accordion=gr.Accordion(
        label="⚙️ Parameters", open=False, render=False
    ),
    additional_inputs=[
        gr.Slider(minimum=0, maximum=1, step=0.1, value=0.9, label="Temperature"),
        gr.Checkbox(label="Sampling", value=True),
        gr.Slider(minimum=128, maximum=4096, step=1, value=512, label="Max new tokens"),
    ],
    stop_btn="Stop Generation",
    title="Chat With LLMs",
    description="Now Running [CampAIgn/Phi-3-mini_16bit](https://huggingface.co./CampAIgn/Phi-3-mini_16bit)",
)

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