!pip install flask pyngrok transformers accelerate

from flask import Flask, request
from pyngrok import ngrok, conf
import threading
from transformers import pipeline

# Set your ngrok authtoken
conf.get_default().auth_token = "2zXMEMLtjmkTv94ApaPMVfYtYCa_WESER25tBVukfAB8ZDLU"

# Initialize text-generation pipeline
generator = pipeline(
    "text-generation",
    model="open-paws/8B-instruct-chat",
    torch_dtype="auto",
    device_map="auto"
)

# Flask app setup
app = Flask(__name__)

@app.route("/")
def home():
    return "Welcome to the Text Generation API! Use /generate?prompt=your_text"

@app.route("/generate")
def generate():
    prompt = request.args.get("prompt", default="Hello, how are you?", type=str)
    response = generator(
        prompt,
        max_new_tokens=128,
        do_sample=True,
        temperature=0.7
    )
    return response[0]['generated_text']

# Start ngrok tunnel
public_url = ngrok.connect(5000)
print(" * ngrok tunnel URL:", public_url)

# Run Flask app
def run():
    app.run()

thread = threading.Thread(target=run)
thread.start()