app.py

import os
import torch
import gradio as gr
from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel
from huggingface_hub import login

# Login using HF token from secrets
hf_token = os.environ.get("HF_TOKEN")
if not hf_token:
    raise RuntimeError("Missing HF_TOKEN in secrets.")
login(token=hf_token)

# Base and LoRA model paths
base_model_id = "unsloth/gemma-2-9b-bnb-4bit"
lora_model_id = "Futuresony/future_12_10_2024"

# Load tokenizer and base model
tokenizer = AutoTokenizer.from_pretrained(base_model_id)
base_model = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    torch_dtype=torch.float16,
    device_map="auto"
)

# Load LoRA weights
model = PeftModel.from_pretrained(base_model, lora_model_id)
model.eval()

# Chat function
def generate_response(message, history, system_message, max_tokens, temperature, top_p):
    prompt = system_message + "\n\n"
    for user_input, bot_response in history:
        prompt += f"User: {user_input}\nAssistant: {bot_response}\n"
    prompt += f"User: {message}\nAssistant:"

    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)
    final_response = response.split("Assistant:")[-1].strip()
    return final_response

# Gradio interface
demo = gr.ChatInterface(
    fn=generate_response,
    additional_inputs=[
        gr.Textbox(value="You are a helpful assistant.", label="System Message"),
        gr.Slider(50, 1024, value=256, step=1, label="Max 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"),
    ],
    title="LoRA Chat Assistant (Gemma-2)",
    description="Chat with your fine-tuned Gemma-2 LoRA model"
)

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