app.py

import gradio as gr
import torch
import re
from transformers import AutoTokenizer, AutoModelForCausalLM

MODEL = "jinesh90/qwen2.5-coder-sql-generator"

print("Loading model...")
tokenizer = AutoTokenizer.from_pretrained(MODEL)
model = AutoModelForCausalLM.from_pretrained(
    MODEL,
    torch_dtype = torch.float16,
    device_map = "auto",
    low_cpu_mem_usage = True,
)
model.eval()
print("Ready!")

def clean_sql(text):
    text = text.strip()
    clean = re.sub(r'[^\x00-\x7F].*', '', text).strip()
    for stop in ["###", "assistant", "\n\n"]:
        if stop in clean:
            clean = clean.split(stop)[0].strip()
    return clean

def build_prompt(question, schema):
    return f"""You are a SQL expert. Generate the simplest and most direct SQL query.
Use JOINs only when multiple tables are needed.

### Schema:
{schema}

### Question:
{question}

### SQL:"""

def generate(question, schema):
    if not question or not schema:
        return "Please provide both a question and schema!"
    
    messages = [{"role": "user", "content": build_prompt(question, schema)}]
    text = tokenizer.apply_chat_template(
        messages,
        tokenize = False,
        add_generation_prompt = True
    )
    inputs = tokenizer(
        text,
        return_tensors = "pt",
        truncation = True,
        max_length = 1024
    ).to(model.device)
    
    stop_tokens = [
        tokenizer.eos_token_id,
        tokenizer.convert_tokens_to_ids("<|im_end|>"),
    ]
    
    with torch.no_grad():
        outputs = model.generate(
            **inputs,
            max_new_tokens = 200,
            do_sample = False,
            temperature = 0,
            repetition_penalty = 1.3,
            eos_token_id = stop_tokens,
            pad_token_id = tokenizer.eos_token_id,
        )
    
    input_len = inputs["input_ids"].shape[1]
    raw = tokenizer.decode(outputs[0, input_len:], skip_special_tokens=True)
    return clean_sql(raw)

# Example schemas for demo
example_schema = """CREATE TABLE employees (
    id INTEGER,
    name VARCHAR,
    salary REAL,
    department VARCHAR,
    age INTEGER
);"""

with gr.Blocks(title="SQL Query Generator") as demo:
    gr.Markdown("# 🗄️ SQL Query Generator")
    gr.Markdown("Fine-tuned Qwen2.5-Coder 7B on Spider dataset | 42% execution accuracy")
    
    with gr.Row():
        with gr.Column():
            schema = gr.Textbox(
                label = "Database Schema (CREATE TABLE statements)",
                value = example_schema,
                lines = 10
            )
            question = gr.Textbox(
                label = "Question",
                placeholder = "How many employees have salary > 50000?",
                lines = 2
            )
            btn = gr.Button("🚀 Generate SQL", variant="primary")
        
        with gr.Column():
            output = gr.Code(
                label = "Generated SQL",
                language = "sql"
            )
            gr.Markdown("""
            ### 📊 Model Stats
            - **Base model**: Qwen2.5-Coder-7B
            - **Training data**: Spider dataset (7.9k samples)
            - **Simple queries**: 64.2% accuracy
            - **Complex queries**: 17.0% accuracy
            - **Overall**: 42% execution accuracy
            """)
    
    btn.click(fn=generate, inputs=[question, schema], outputs=output)
    
    gr.Examples(
        examples=[
            ["How many employees are there?", example_schema],
            ["Find all employees with salary greater than 50000", example_schema],
            ["What is the average salary by department?", example_schema],
        ],
        inputs=[question, schema]
    )

demo.launch()