app.py

# app.py - ZeroGPU compatible version (NO Unsloth)

import gradio as gr
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
from peft import PeftModel
from huggingface_hub import spaces   # ← important!

# Your model paths
BASE_MODEL = "unsloth/Phi-3-mini-4k-instruct-bnb-4bit"
LORA_PATH = "saadkhi/SQL_Chat_finetuned_model"

print("Loading model on CPU first... (will use GPU only during @spaces.GPU)")
bnb_config = BitsAndBytesConfig(load_in_4bit=True)
model = AutoModelForCausalLM.from_pretrained(
    BASE_MODEL,
    quantization_config=bnb_config,
    device_map="auto",
    trust_remote_code=True
)
model = PeftModel.from_pretrained(model, LORA_PATH)
tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL)
model.eval()

@spaces.GPU   # ← this requests GPU slice only during this function
def generate_sql(prompt: str):
    messages = [{"role": "user", "content": prompt}]
    inputs = tokenizer.apply_chat_template(messages, tokenize=True, add_generation_prompt=True, return_tensors="pt").to("cuda")
    
    with torch.inference_mode():
        outputs = model.generate(
            inputs,
            max_new_tokens=180,
            temperature=0.0,
            do_sample=False,
            use_cache=True,
            pad_token_id=tokenizer.eos_token_id,
        )
    
    response = tokenizer.decode(outputs[0], skip_special_tokens=True)
    if "<|assistant|>" in response:
        response = response.split("<|assistant|>", 1)[-1].strip()
    return response.split("<|end|>")[0].strip()

demo = gr.Interface(
    fn=generate_sql,
    inputs=gr.Textbox(label="Your SQL question"),
    outputs="text",
    title="SQL Chatbot (ZeroGPU)",
    description="Free but limited daily GPU time"
)

if __name__ == "__main__":
    demo.launch()