app.py

import streamlit as st
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
import torch

# Load the tokenizer and models
tokenizer = AutoTokenizer.from_pretrained("cssupport/t5-small-awesome-text-to-sql")
original_model = AutoModelForSeq2SeqLM.from_pretrained("cssupport/t5-small-awesome-text-to-sql", torch_dtype=torch.bfloat16)
ft_model = AutoModelForSeq2SeqLM.from_pretrained("daljeetsingh/sql_ft_t5small_kag", torch_dtype=torch.bfloat16)

# Move models to GPU
device = 'cuda' if torch.cuda.is_available() else 'cpu'
original_model.to(device)
ft_model.to(device)

# Streamlit app layout
st.title("SQL Generation with T5 Models")

# Input text box
input_text = st.text_area("Enter your query:", height=150)

# Generate button
if st.button("Generate SQL"):
    if input_text:
        # Tokenize input
        inputs = tokenizer(input_text, return_tensors='pt').to(device)

        # Generate SQL queries
        with torch.no_grad():
            original_sql = tokenizer.decode(
                original_model.generate(inputs["input_ids"], max_new_tokens=200)[0],
                skip_special_tokens=True
            )
            ft_sql = tokenizer.decode(
                ft_model.generate(inputs["input_ids"], max_new_tokens=200)[0],
                skip_special_tokens=True
            )

        # Display results
        st.subheader("Original Model Output")
        st.write(original_sql)
        st.subheader("Fine-Tuned Model Output")
        st.write(ft_sql)
    else:
        st.warning("Please enter a query to generate SQL.")