app.py

# app.py - Fixed for recent Gradio versions (no allow_flagging)

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

# ────────────────────────────────────────────────────────────────
#   Fastest practical configuration
# ────────────────────────────────────────────────────────────────

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

MAX_NEW_TOKENS = 180
TEMPERATURE    = 0.0          # greedy = fastest
DO_SAMPLE      = False

# ────────────────────────────────────────────────────────────────
#   4-bit quantization (very important for speed)
# ────────────────────────────────────────────────────────────────

bnb_config = BitsAndBytesConfig(
    load_in_4bit              = True,
    bnb_4bit_quant_type       = "nf4",
    bnb_4bit_use_double_quant = True,
    bnb_4bit_compute_dtype    = torch.bfloat16
)

print("Loading quantized base model...")
model = AutoModelForCausalLM.from_pretrained(
    BASE_MODEL,
    quantization_config = bnb_config,
    device_map          = "auto",
    trust_remote_code   = True,
    torch_dtype         = torch.bfloat16
)

print("Loading LoRA adapters...")
model = PeftModel.from_pretrained(model, LORA_PATH)

# Merge LoRA into base model → much faster inference
model = model.merge_and_unload()

tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL)

model.eval()
print("Model ready!")

# ────────────────────────────────────────────────────────────────
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(model.device)

    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()
    response = response.split("<|end|>")[0].strip() if "<|end|>" in response else response

    return response

# ────────────────────────────────────────────────────────────────
#   Gradio interface - modern style (no allow_flagging)
# ────────────────────────────────────────────────────────────────

demo = gr.Interface(
    fn=generate_sql,
    inputs=gr.Textbox(
        label="Ask SQL related question",
        placeholder="Show me all employees with salary > 50000...",
        lines=3
    ),
    outputs=gr.Textbox(label="Generated SQL / Answer"),
    title="SQL Chatbot - Optimized",
    description="Phi-3-mini 4bit + LoRA merged",
    examples=[
        ["Find duplicate emails in users table"],
        ["Top 5 highest paid employees"],
        ["Count orders per customer last month"]
    ],
    # flag button is disabled by default in newer versions → no need for allow_flagging
)

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