app.py

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

# device = "cuda" if torch.cuda.is_available() else "cpu"

# base_model = "unsloth/Phi-3-mini-4k-instruct-bnb-4bit"
# finetuned_model = "saadkhi/SQL_Chat_finetuned_model"

# tokenizer = AutoTokenizer.from_pretrained(base_model)

# bnb = BitsAndBytesConfig(load_in_4bit=True)

# model = AutoModelForCausalLM.from_pretrained(
#     base_model,
#     quantization_config=bnb,
#     torch_dtype=torch.bfloat16 if device == "cuda" else torch.float32,
#     device_map="auto"
# )

# model = PeftModel.from_pretrained(model, finetuned_model).to(device)
# model.eval()

# def chat(prompt):
#     inputs = tokenizer(prompt, return_tensors="pt").to(device)

#     with torch.inference_mode():
#         output = model.generate(
#             **inputs,
#             max_new_tokens=60,
#             temperature=0.1,
#             do_sample=False
#         )

#     return tokenizer.decode(output[0], skip_special_tokens=True)

# iface = gr.Interface(fn=chat, inputs="text", outputs="text", title="SQL Chatbot")
# iface.launch()









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

# Best 4-bit config for speed + low memory
quant_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16,
    bnb_4bit_use_double_quant=True,
)

# Load base + your LoRA once
base_model_name = "unsloth/Phi-3-mini-4k-instruct-bnb-4bit"
lora_model_name = "saadkhi/SQL_Chat_finetuned_model"

print("Loading model (20–40s first time)...")
base_model = AutoModelForCausalLM.from_pretrained(
    base_model_name,
    quantization_config=quant_config,
    device_map="auto",
    trust_remote_code=True,
    # Removed flash_attention_2 — avoids install issues
)

model = PeftModel.from_pretrained(base_model, lora_model_name)
tokenizer = AutoTokenizer.from_pretrained(base_model_name, trust_remote_code=True)

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

def chat(message, history):
    # Full conversation history
    messages = []
    for user, assistant in history:
        messages.append({"role": "user", "content": user})
        if assistant:
            messages.append({"role": "assistant", "content": assistant})
    messages.append({"role": "user", "content": message})

    inputs = tokenizer.apply_chat_template(
        messages,
        tokenize=True,
        add_generation_prompt=True,
        return_tensors="pt"
    ).to(model.device)

    # Optimized generation
    outputs = model.generate(
        inputs,
        max_new_tokens=256,
        temperature=0.7,
        do_sample=True,
        top_p=0.9,
        repetition_penalty=1.1,
        use_cache=True,  # KV cache = faster sequential tokens
        eos_token_id=tokenizer.eos_token_id,
    )

    response = tokenizer.decode(outputs[0][inputs.shape[-1]:], skip_special_tokens=True)
    
    history.append((message, response))
    return history, ""

# UI
with gr.Blocks(title="SQL Chatbot", theme=gr.themes.Soft()) as demo:
    gr.Markdown("# SQL Chat Assistant")
    gr.Markdown("Fine-tuned Phi-3 Mini for SQL. Fast responses (3–8s on GPU).")
    
    chatbot = gr.Chatbot(height=500)
    msg = gr.Textbox(label="Your Question", placeholder="e.g., delete duplicate rows from users table based on email", lines=2)
    clear = gr.Button("Clear")

    msg.submit(chat, [msg, chatbot], [chatbot, msg])
    clear.click(lambda: ([], ""), None, chatbot)

demo.queue(max_size=30)
demo.launch()