app.py

import gradio as gr
import spaces
import torch
from huggingface_hub import hf_hub_download

zero = torch.Tensor([0]).cuda()
print(zero.device) # <-- 'cpu' 🤔

@spaces.GPU
def greet(n):
    print(zero.device) # <-- 'cuda:0' 🤗
    return f"Hello {zero + n} Tensor"

def download_model():

REPO_ID = "TheBloke/Llama-2-7B-GGUF"
FILENAME = "llama-2-7b.Q5_K_S.gguf"

hf_hub_download(repo_id=REPO_ID, filename=FILENAME)

def load_model():
    from llama_cpp import Llama, LlamaGrammar
    model_url="https://huggingface.co/TheBloke/Llama-2-7B-GGUF/resolve/main/llama-2-7b.Q5_K_S.gguf"
    llm = Llama(
        model_path=model_url,
        n_gpu_layers=-1, verbose=False
    )

    grammar = LlamaGrammar.from_string('''
    root ::= sentence
    answer ::= (weather | complaint | yesno | gen)
    weather ::= ("Sunny." | "Cloudy." | "Rainy.")
    complaint ::= "I don't like talking about the weather."
    yesno ::= ("Yes." | "No.")
    gen ::= "1. " [A-Z] [a-z] [a-z]*
    sentence ::= [A-Z] [A-Za-z0-9 ,-]* ("." | "!" | "?")
    ''')

    prompts = [
        "How's the weather in London?",
        "How's the weather in Munich?",
        "How's the weather in Barcelona?",
    ]

    for prompt in prompts:
        output = llm(
                prompt,
                max_tokens=512,
                temperature=0.4,
                grammar=grammar
        )

        s = output['choices'][0]['text']
        print(f'{s} , len(s) = {len(s)}')
        print(output['choices'])
        print(output['choices'][0]['text'])
        print()

download_model()
load_model()
demo = gr.Interface(fn=greet, inputs=gr.Number(), outputs=gr.Text())
demo.launch(share=False)