app.py

import gradio as gr
import torch
from unsloth import FastLanguageModel
from transformers import TextStreamer

# Configuration Variables
model_name = "unsloth/Llama-3.2-3B-Instruct-bnb-4bit"  # Replace with your actual model name
lora_adapter = "Braszczynski/Llama-3.2-3B-Instruct-bnb-4bit-460steps"

max_seq_length = 512  # Adjust as needed
dtype = None   # Example dtype, adjust based on your setup
load_in_4bit = True     # Set to True if you want to use 4-bit quantization

# Load the model and tokenizer using FastLanguageModel
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name=model_name,
    max_seq_length=max_seq_length,
    dtype=dtype,
    load_in_4bit=load_in_4bit,
)

# Load the adapter
model.load_adapter(lora_adapter)

# Enable native 2x faster inference
FastLanguageModel.for_inference(model)

# Optional: Initialize TextStreamer if you plan to use streaming
# text_streamer = TextStreamer(tokenizer, skip_prompt=True)

def respond(message, history, system_message, max_tokens, temperature, top_p):
    # Combine system message and chat history
    chat_history = f"{system_message}\n"
    for user_msg, bot_reply in history:
        chat_history += f"User: {user_msg}\nAssistant: {bot_reply}\n"
    chat_history += f"User: {message}\nAssistant:"

    # Apply chat template and tokenize the input
    inputs = tokenizer.apply_chat_template(
        [{"role": "user", "content": message}] if not history else [
            {"role": "system", "content": system_message}] + [
            {"role": "user", "content": msg} for msg, _ in history
        ] + [{"role": "assistant", "content": reply} for _, reply in history] + [
            {"role": "user", "content": message}
        ],
        tokenize=True,
        add_generation_prompt=True,  # Must add for generation
        return_tensors="pt",
    ).to("cuda")

    # Generate response
    outputs = model.generate(
        input_ids=inputs["input_ids"],
        max_new_tokens=max_tokens,
        temperature=temperature,
        top_p=top_p,
        pad_token_id=tokenizer.eos_token_id,
        use_cache=True
        # streamer=text_streamer  # Uncomment if using streaming
    )

    # Decode and format the output
    response = tokenizer.decode(outputs[0], skip_special_tokens=True)
    response = response[len(chat_history):].strip()  # Remove input context from output
    return response

# Define the Gradio interface
demo = gr.ChatInterface(
    respond,
    additional_inputs=[
        gr.Textbox(value="You are a friendly assistant.", label="System message"),
        gr.Slider(minimum=1, maximum=2048, value=512, step=1, label="Max new tokens"),
        gr.Slider(minimum=0.1, maximum=4.0, value=0.7, step=0.1, label="Temperature"),
        gr.Slider(minimum=0.1, maximum=1.0, value=0.95, step=0.05, label="Top-p (nucleus sampling)"),
    ],
)

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