app.py

import gradio as gr
import spaces
from guidance import gen
from guidance.models import Transformers
from mymodule.utils import load_model_and_tok

model, tok = load_model_and_tok()


@spaces.GPU(duration=10)
def greet(prompt):
    temp = f"{prompt} Only use results from the genomic data commons in your response and provide frequencies as a percentage. Only report the final response."
    lm = Transformers(model=model, tokenizer=tok)
    lm += temp
    lm += gen(
        "response",
        n=1,
        temperature=0,
        max_tokens=1024,
        regex="The final answer is: \d*\.\d*%",
    )
    return lm["response"]


demo = gr.Interface(fn=greet, inputs=gr.Text(), outputs=gr.Text())
demo.launch()