app.py

#import gradio as gr

#gr.load("models/KvrParaskevi/Llama-2-7b-Hotel-Booking-Model-8Bit").launch()

import gradio as gr
import spaces

from transformers import pipeline

pipe = pipeline("text-generation", model="KvrParaskevi/Llama-2-7b-Hotel-Booking-Model")

@spaces.GPU
def predict(text):
  answer = pipe(text)[0]
  print(answer)
  return pipe(text)[0]["text"]

demo = gr.Interface(
  fn=predict,
  inputs='text',
  outputs='text',
)

demo.launch()