app.py

# app.py - ZeroGPU safe version (no .to("cuda") outside decorated fn + no caching)

import torch
import gradio as gr
import spaces  # Correct import
from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
from peft import PeftModel

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

MAX_NEW_TOKENS = 180
TEMPERATURE    = 0.0
DO_SAMPLE      = False

print("Loading quantized base model on CPU (GPU only in @spaces.GPU)...")
bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16
)

model = AutoModelForCausalLM.from_pretrained(
    BASE_MODEL,
    quantization_config=bnb_config,
    device_map="cpu",  # ← Force CPU at load time (required for ZeroGPU)
    trust_remote_code=True
)

print("Loading LoRA...")
model = PeftModel.from_pretrained(model, LORA_PATH)
model = model.merge_and_unload()  # Merge for speed

tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL)
model.eval()

# ────────────────────────────────────────────────────────────────
@spaces.GPU(duration=60)  # 60s max is safe & gives good queue priority
def generate_sql(prompt: str):
    messages = [{"role": "user", "content": prompt}]
    
    # Tokenize on CPU first
    inputs = tokenizer.apply_chat_template(
        messages,
        tokenize=True,
        add_generation_prompt=True,
        return_tensors="pt"
    )
    
    # Move to CUDA ONLY inside here (GPU is now allocated)
    inputs = inputs.to("cuda")
    
    with torch.inference_mode():
        outputs = model.generate(
            input_ids=inputs,
            max_new_tokens=MAX_NEW_TOKENS,
            temperature=TEMPERATURE,
            do_sample=DO_SAMPLE,
            use_cache=True,
            pad_token_id=tokenizer.eos_token_id,
        )

    response = tokenizer.decode(outputs[0], skip_special_tokens=True)
    
    # Clean output
    if "<|assistant|>" in response:
        response = response.split("<|assistant|>", 1)[-1].strip()
    if "<|end|>" in response:
        response = response.split("<|end|>")[0].strip()

    return response

# ────────────────────────────────────────────────────────────────
demo = gr.Interface(
    fn=generate_sql,
    inputs=gr.Textbox(
        label="Ask SQL question",
        placeholder="Delete duplicate rows from users table based on email",
        lines=3
    ),
    outputs=gr.Textbox(label="Generated SQL"),
    title="SQL Chatbot (ZeroGPU Safe)",
    description="Phi-3-mini 4bit + LoRA - GPU allocated only during generation",
    examples=[
        ["Find duplicate emails in users table"],
        ["Top 5 highest paid employees"],
        ["Count orders per customer last month"]
    ],
    cache_examples=False  # ← CRITICAL: Disable caching to avoid startup .to("cuda") call
)

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