app.py

import os
import gradio as gr
from transformers import AutoTokenizer, AutoModelForCausalLM
from transformers import BitsAndBytesConfig
import torch

# ✅ Load the model and tokenizer
MODEL_ID = "pareshmishra/mt564-gemma-lora"
API_TOKEN = os.getenv("HUGGINGFACEHUB_API_TOKEN")
if not API_TOKEN:
    raise ValueError("HUGGINGFACEHUB_API_TOKEN environment variable not set")

# Configure 4-bit quantization
quantization_config = BitsAndBytesConfig(
    load_in_4bit=True,  # Enable 4-bit quantization
    bnb_4bit_compute_dtype=torch.float16,  # Use fp16 for computation
    bnb_4bit_quant_type="nf4",  # Normal Float 4-bit quantization
    bnb_4bit_use_double_quant=True  # Nested quantization for efficiency
)

tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, token=API_TOKEN)
model = AutoModelForCausalLM.from_pretrained(
    MODEL_ID,
    token=API_TOKEN,
    torch_dtype=torch.float16,  # fp16 as per model card
    device_map="auto",  # Auto-map to GPU/CPU
    quantization_config=quantization_config  # Use BitsAndBytesConfig
)

def respond(messages, chatbot_history, system_message, max_tokens, temperature, top_p):
    try:
        # Build prompt from history
        prompt = f"{system_message.strip()}\n\n"
        for msg in messages:
            if isinstance(msg, dict):
                role = msg.get("role")
                content = msg.get("content", "")
                if role == "user":
                    prompt += f"User: {content.strip()}\n"
                elif role == "assistant":
                    prompt += f"Assistant: {content.strip()}\n"
        prompt += "Assistant:"

        # Tokenize and generate
        inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
        outputs = model.generate(
            **inputs,
            max_new_tokens=max_tokens,
            temperature=temperature,
            top_p=top_p,
            do_sample=True,
            pad_token_id=tokenizer.eos_token_id
        )

        response = tokenizer.decode(outputs[0], skip_special_tokens=True)
        response = response[len(prompt):].strip()

        yield response if response else "⚠️ No response returned from the model."

    except Exception as e:
        yield f"❌ Error: {str(e)}\nDetails: {e.__class__.__name__}"

# Gradio Interface
demo = gr.ChatInterface(
    fn=respond,
    type="messages",
    additional_inputs=[
        gr.Textbox(
            lines=3,
            label="System message",
            value="You are an expert in SWIFT MT564 financial messaging. Analyze, validate, and answer related user questions.",
        ),
        gr.Slider(50, 2048, value=512, step=1, label="Max new tokens"),
        gr.Slider(0.1, 1.5, value=0.7, step=0.1, label="Temperature"),
        gr.Slider(0.1, 1.0, value=0.95, step=0.05, label="Top-p sampling"),
    ],
    title="💬 MT564 Chat Assistant",
    description="Analyze SWIFT MT564 messages or ask financial-related questions.",
    theme="default"
)

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