app.py

import gradio as gr
from unsloth import FastLanguageModel
import torch

# Load the model and tokenizer locally
max_seq_length = 2048
model_name_or_path = "unsloth/Llama-3.2-3B-Instruct"

# Load model and tokenizer using unsloth
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name=model_name_or_path,
    max_seq_length=max_seq_length,
    load_in_4bit=True,
)
FastLanguageModel.for_inference(model)  # Enable optimized inference

# Define the response function
def respond(message, history, system_message, max_tokens, temperature, top_p):
    # Start by preparing only the conversation history (user-assistant pairs)
    messages = []
    if history:
        for entry in history:
            messages.append({"role": "user", "content": entry["user"]})
            messages.append({"role": "assistant", "content": entry["assistant"]})
    
    # Add the user's new message to the list
    messages.append({"role": "user", "content": message})

    # Tokenize the input (prepare the data for the model)
    inputs = tokenizer.apply_chat_template(
        messages,
        tokenize=True,
        add_generation_prompt=True,
        return_tensors="pt",
    ).to("cuda" if torch.cuda.is_available() else "cpu")

    # Generate the response
    attention_mask = inputs.ne(tokenizer.pad_token_id).long()
    generated_tokens = model.generate(
        input_ids=inputs,
        attention_mask=attention_mask,
        max_new_tokens=max_tokens,
        use_cache=True,
        temperature=temperature,
        top_p=top_p,
    )

    response = tokenizer.decode(generated_tokens[0], skip_special_tokens=True)
    
    # Update the conversation history with the new user-assistant pair
    if history is None:
        history = []
    history.append({"user": message, "assistant": response})

    # Prepare the history for Gradio
    formatted_history = []
    for entry in history:
        formatted_history.append({"role": "user", "content": entry["user"]})
        formatted_history.append({"role": "assistant", "content": entry["assistant"]})

    # Return the formatted history for Gradio to display
    return formatted_history



# Define the Gradio interface
demo = gr.ChatInterface(
    fn=respond,
    additional_inputs=[
        gr.Textbox(value="You are a helpful assistant.", label="System message"),
        gr.Slider(minimum=1, maximum=2048, value=512, step=1, label="Max 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"),
    ],
    type="messages",
)


if __name__ == "__main__":
    demo.launch(share=False)  # Use share=False for local testing