""" NL2SQL — live demo Space (ZeroGPU). Model is loaded lazily inside the GPU function because ZeroGPU only attaches a GPU during an @spaces.GPU call. """ import spaces import gradio as gr import torch from transformers import AutoModelForCausalLM, AutoTokenizer from peft import PeftModel BASE_MODEL = "Qwen/Qwen2.5-Coder-1.5B-Instruct" ADAPTER_REPO = "JayaHareeswarReddy/nl2sql-qlora" SYSTEM_PROMPT = ( "You are a text-to-SQL assistant. Given a database schema and a question, " "respond with a single correct SQL query and nothing else." ) tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL) _model = None def _get_model(): global _model if _model is None: base = AutoModelForCausalLM.from_pretrained( BASE_MODEL, torch_dtype=torch.float16, device_map="cuda" ) m = PeftModel.from_pretrained(base, ADAPTER_REPO) m.eval() _model = m return _model @spaces.GPU(duration=120) def generate_sql(schema: str, question: str): schema = (schema or "").strip() question = (question or "").strip() if not schema or not question: return "-- Please provide both a schema and a question." model = _get_model() messages = [ {"role": "system", "content": SYSTEM_PROMPT}, {"role": "user", "content": f"Schema:\n{schema}\n\nQuestion: {question}"}, ] prompt = tokenizer.apply_chat_template( messages, tokenize=False, add_generation_prompt=True ) inputs = tokenizer(prompt, return_tensors="pt").to("cuda") with torch.no_grad(): out = model.generate( **inputs, max_new_tokens=256, do_sample=False, pad_token_id=tokenizer.eos_token_id, ) text = tokenizer.decode( out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True ) return text.strip() EXAMPLE_SCHEMA = ( "CREATE TABLE employees (id INT, name TEXT, salary INT, department TEXT);\n" "CREATE TABLE departments (name TEXT, budget INT);" ) EXAMPLES = [ [EXAMPLE_SCHEMA, "List the names of employees in Sales earning more than 50000."], [EXAMPLE_SCHEMA, "What is the average salary per department?"], [EXAMPLE_SCHEMA, "How many employees are in each department, highest first?"], ] with gr.Blocks(title="NL2SQL", theme=gr.themes.Soft()) as demo: gr.Markdown( """ # 🧮 NL2SQL — Natural Language to SQL A **Qwen2.5-Coder-1.5B** model fine-tuned with **QLoRA** to turn an English question and a table schema into a SQL query. *(First query loads the model and may take ~30s; later ones are fast.)* """ ) with gr.Row(): with gr.Column(): schema_in = gr.Textbox( label="Database schema (CREATE TABLE ...)", lines=6, value=EXAMPLE_SCHEMA, ) question_in = gr.Textbox(label="Question", lines=2) btn = gr.Button("Generate SQL", variant="primary") with gr.Column(): sql_out = gr.Code(label="Generated SQL", language="sql") btn.click(generate_sql, [schema_in, question_in], sql_out) gr.Examples(EXAMPLES, inputs=[schema_in, question_in]) gr.Markdown( "Base: Qwen2.5-Coder-1.5B-Instruct · Fine-tuned with QLoRA on " "gretelai/synthetic_text_to_sql · Built with Gradio, PEFT & 🤗 Transformers" ) if __name__ == "__main__": demo.launch()