app.py

"""
The Streamlit app for the project demo.
In the demo, the user can write a prompt
 and the model will generate a response using the grouped sampling algorithm.
"""
import os

import streamlit as st
from torch.cuda import CudaError
from huggingface_hub import logging as hf_hub_logging

from available_models import AVAILABLE_MODELS
from hanlde_form_submit import on_form_submit


hf_hub_logging.set_verbosity_error()

st.set_page_config(
    page_title="דגימה בקבוצות - שימוש יעיל במודלי שפה סיבתיים",
    layout="wide",
)

with st.form("request_form"):
    selected_model_name: str = st.selectbox(
        label="בחרו מודל",
        options=AVAILABLE_MODELS,
        help="llama-30b-hf generates better texts but is slower",
    )

    output_length: int = st.number_input(
        label="כמות המילים המקסימלית בפלט - בין 1 ל-1024",
        min_value=1,
        max_value=1024,
        value=5,
    )

    submitted_prompt: str = st.text_area(
        label="הקלט לאלוגריתם (באנגלית בלבד)",
        value="Instruction: Answer in yes or no.\n"
              "Question: Is the sky blue?\n"
              "Answer:",
        max_chars=2048,
    )

    submitted: bool = st.form_submit_button(
        label="צור טקסט",
        disabled=False,
    )

    if submitted:
        try:
            output = on_form_submit(
                selected_model_name,
                output_length,
                submitted_prompt,
            )
        except CudaError as e:
            st.error("Out of memory. Please try a smaller model, shorter prompt, or a smaller output length.")
        except (ValueError, TypeError, RuntimeError) as e:
            st.error(e)
        else:
            st.write(f"Generated text: {output}")


user_instructions_file = os.path.join(
    os.path.dirname(__file__),
    "user_instructions_hebrew.md",
)
with open(user_instructions_file, "r") as fh:
    long_description = fh.read()
st.markdown(long_description)