app.py

import torch
import streamlit as st
from transformers import AutoTokenizer, OPTForCausalLM


@st.cache_resource
def load_model():
    tokenizer = AutoTokenizer.from_pretrained("facebook/galactica-30b")
    model = OPTForCausalLM.from_pretrained("facebook/galactica-30b", device_map='auto', low_cpu_mem_usage=True, torch_dtype=torch.float16)
    model.gradient_checkpointing_enable()
    return tokenizer, model


st.set_page_config(
    page_title='BioML-SVM',
    layout="wide"
)

with st.spinner("Loading Models and Tokens..."):
    tokenizer, model = load_model()

with st.form(key='my_form'):
    col1, col2 = st.columns([10, 1])
    text_input = col1.text_input(label='Enter the amino sequence')
    with col2:
        st.text('')
        st.text('')
        submit_button = st.form_submit_button(label='Submit')

    if submit_button:
        st.session_state['result_done'] = False
    # input_text = "[START_AMINO]GHMQSITAGQKVISKHKNGRFYQCEVVRLTTETFYEVNFDDGSFSDNLYPEDIVSQDCLQFGPPAEGEVVQVRWTDGQVYGAKFVASHPIQMYQVEFEDGSQLVVKRDDVYTLDEELP[END_AMINO]"
        with st.spinner('Generating...'):
            # formatted_text = f"[START_AMINO]{text_input}[END_AMINO]"
            # formatted_text = f"Here is the sequence: [START_AMINO]{text_input}[END_AMINO]"
            formatted_text = f"{text_input}"
            input_ids = tokenizer(formatted_text, return_tensors="pt").input_ids.to("cuda")
            outputs = model.generate(
                input_ids=input_ids,
                max_new_tokens=500
            )
            result = tokenizer.decode(outputs[0]).replace(formatted_text, "")
        st.markdown(result)

        if 'result_done' not in st.session_state or not st.session_state.result_done:
            st.session_state['result_done'] = True
            st.session_state['previous_state'] = result
    else:
        if 'result_done' in st.session_state and st.session_state.result_done:
            st.markdown(st.session_state.previous_state)

if 'result_done' in st.session_state and st.session_state.result_done:
    with st.form(key='ask_more'):
        col1, col2 = st.columns([10, 1])
        text_input = col1.text_input(label='Ask more question')
        with col2:
            st.text('')
            st.text('')
            submit_button = st.form_submit_button(label='Submit')

        if submit_button:
            with st.spinner('Generating...'):
                # formatted_text = f"[START_AMINO]{text_input}[END_AMINO]"
                formatted_text = f"Q:{text_input}\n\nA:\n\n"
                input_ids = tokenizer(formatted_text, return_tensors="pt").input_ids.to("cuda")

                outputs = model.generate(
                    input_ids=input_ids,
                    max_length=len(formatted_text) + 500,
                    do_sample=True,
                    top_k=40,
                    num_beams=1,
                    num_return_sequences=1
                )
                result = tokenizer.decode(outputs[0]).replace(formatted_text, "")
            st.markdown(result)