app.py

import gradio as gr
from openai import OpenAI
import os

ACCESS_TOKEN = os.getenv("HF_TOKEN")
print("Access token loaded.")

client = OpenAI(
    base_url="https://api-inference.huggingface.co/v1/",
    api_key=ACCESS_TOKEN,
)
print("OpenAI client initialized.")

def respond(
    message,
    history: list[tuple[str, str]],
    system_message="You are a helpful assistant.",
    max_tokens=512,
    temperature=0.7,
    top_p=0.95,
    frequency_penalty=0.0,
    seed=-1
):
    print(f"Received message: {message}")
    print(f"History: {history}")
    print(f"System message: {system_message}")
    print(f"Max tokens: {max_tokens}, Temperature: {temperature}, Top-P: {top_p}")
    print(f"Frequency Penalty: {frequency_penalty}, Seed: {seed}")

    # Convert seed to None if -1 (meaning random)
    if seed == -1:
        seed = None

    messages = [{"role": "system", "content": system_message}]
    print("Initial messages array constructed.")

    # Add conversation history to the context
    for val in history:
        user_part = val[0]
        assistant_part = val[1]
        if user_part:
            messages.append({"role": "user", "content": user_part})
            print(f"Added user message to context: {user_part}")
        if assistant_part:
            messages.append({"role": "assistant", "content": assistant_part})
            print(f"Added assistant message to context: {assistant_part}")

    # Append the latest user message
    messages.append({"role": "user", "content": message})
    print("Latest user message appended.")

    # Set the model to "meta" by default
    model_to_use = "meta-llama/Llama-3.3-70B-Instruct"
    print(f"Model selected for inference: {model_to_use}")

    # Start with an empty string to build the response as tokens stream in
    response = ""
    print("Sending request to OpenAI API.")

    for message_chunk in client.chat.completions.create(
        model=model_to_use,
        max_tokens=max_tokens,
        stream=True,
        temperature=temperature,
        top_p=top_p,
        frequency_penalty=frequency_penalty,
        seed=seed,
        messages=messages,
    ):
        token_text = message_chunk.choices[0].delta.content
        print(f"Received token: {token_text}")
        response += token_text
        yield response

    print("Completed response generation.")

# GRADIO UI

chatbot = gr.Chatbot(height=600, show_copy_button=True, placeholder="Start chatting!", likeable=True, layout="panel")
print("Chatbot interface created.")

system_message_box = gr.Textbox(value="You are a helpful assistant.", label="System Prompt", visible=False)

max_tokens_slider = gr.Slider(
    minimum=1,
    maximum=4096,
    value=512,
    step=1,
    label="Max new tokens"
)
temperature_slider = gr.Slider(
    minimum=0.1,
    maximum=4.0,
    value=0.7,
    step=0.1,
    label="Temperature"
)
top_p_slider = gr.Slider(
    minimum=0.1,
    maximum=1.0,
    value=0.95,
    step=0.05,
    label="Top-P"
)
frequency_penalty_slider = gr.Slider(
    minimum=-2.0,
    maximum=2.0,
    value=0.0,
    step=0.1,
    label="Frequency Penalty"
)
seed_slider = gr.Slider(
    minimum=-1,
    maximum=65535,
    value=-1,
    step=1,
    label="Seed (-1 for random)"
)

# Removed the custom_model_box as the model is pre-set

demo = gr.ChatInterface(
    fn=respond,
    additional_inputs=[
        system_message_box,
        max_tokens_slider,
        temperature_slider,
        top_p_slider,
        frequency_penalty_slider,
        seed_slider,
    ],
    fill_height=True,
    chatbot=chatbot,
    theme="Nymbo/Nymbo_Theme",
)
print("ChatInterface object created.")

with demo:
    # No need for a model selection accordion since the model is fixed to "meta-llama"
    pass

print("Gradio interface initialized.")

if __name__ == "__main__":
    print("Launching the demo application.")
    demo.launch()