app.py

import gradio as gr
from openai import OpenAI
import os

# Retrieve the access token from the environment variable
ACCESS_TOKEN = os.getenv("HF_TOKEN")
print("Access token loaded.")

# Initialize the OpenAI client with the Hugging Face Inference API endpoint
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,
    max_tokens,
    temperature,
    top_p,
    frequency_penalty,
    seed,
    custom_model
):
    """
    This function handles the chatbot response. It takes in:
    - message: the user's new message
    - history: the list of previous messages, each as a tuple (user_msg, assistant_msg)
    - system_message: the system prompt
    - max_tokens: the maximum number of tokens to generate in the response
    - temperature: sampling temperature
    - top_p: top-p (nucleus) sampling
    - frequency_penalty: penalize repeated tokens in the output
    - seed: a fixed seed for reproducibility; -1 will mean 'random'
    - custom_model: the final model name in use, which may be set by selecting from the Featured Models radio or by typing a custom model
    """

    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}")
    print(f"Selected model (custom_model): {custom_model}")

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

    # Construct the messages array required by the API
    messages = [{"role": "system", "content": system_message}]

    # 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})

    # If user provided a model, use that; otherwise, fall back to a default
    model_to_use = custom_model.strip() if custom_model.strip() != "" else "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.")

    # Make the streaming request to the HF Inference API via openai-like client
    for message_chunk in client.chat.completions.create(
        model=model_to_use,              # Use either the user-provided or default model
        max_tokens=max_tokens,
        stream=True,                     # Stream the response
        temperature=temperature,
        top_p=top_p,
        frequency_penalty=frequency_penalty,
        seed=seed,
        messages=messages,
    ):
        # Extract the token text from the response chunk
        token_text = message_chunk.choices[0].delta.content
        print(f"Received token: {token_text}")
        response += token_text
        # Yield the partial response to Gradio so it can display in real-time
        yield response

    print("Completed response generation.")

# -------------------------
# GRADIO UI CONFIGURATION
# -------------------------

# Create a Chatbot component with a specified height
chatbot = gr.Chatbot(height=600)
print("Chatbot interface created.")

# We'll create text boxes & sliders for system prompt, tokens, etc.
system_message_box = gr.Textbox(value="", label="System message")

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)"
)

# The custom_model_box is what the respond function sees as "custom_model"
custom_model_box = gr.Textbox(
    value="",
    label="Custom Model",
    info="(Optional) Provide a custom Hugging Face model path. Overrides any selected featured model."
)

# Define a function that, when a user selects a model from the radio, populates `custom_model_box`
def set_custom_model_from_radio(selected):
    """
    This function will get triggered whenever someone picks a model from the 'Featured Models' radio.
    We will update the Custom Model text box with that selection automatically.
    """
    return selected

# The main ChatInterface object
demo = gr.ChatInterface(
    fn=respond,
    # For ChatInterface, we can pass additional inputs in order to feed them into the "respond" function
    additional_inputs=[
        system_message_box,
        max_tokens_slider,
        temperature_slider,
        top_p_slider,
        frequency_penalty_slider,
        seed_slider,
        custom_model_box
    ],
    fill_height=True,
    chatbot=chatbot,
    theme="Nymbo/Nymbo_Theme",
)

# -----------
# ADDING THE "FEATURED MODELS" ACCORDION
# -----------
with demo:
    with gr.Accordion("Featured Models", open=False):
        model_search_box = gr.Textbox(
            label="Filter Models",
            placeholder="Search for a featured model...",
            lines=1
        )

        # Sample list of popular text models
        models_list = [
            "meta-llama/Llama-3.3-70B-Instruct",
            "bigscience/bloomz-7b1",
            "OpenAssistant/oasst-sft-1-pythia-12b",
            "OpenAssistant/oasst-sft-4-pythia-12b-epoch-3.5",
            "tiiuae/falcon-7b-instruct",
            "OpenAI/gpt-3.5-turbo",
            "OpenAI/gpt-4-32k",
            "meta-llama/Llama-2-13B-chat-hf",
            "meta-llama/Llama-2-70B-chat-hf",
        ]

        featured_model_radio = gr.Radio(
            label="Select a model below",
            choices=models_list,
            value="meta-llama/Llama-3.3-70B-Instruct",
            interactive=True
        )

        # Filter function for the radio
        def filter_models(search_term):
            filtered = [m for m in models_list if search_term.lower() in m.lower()]
            return gr.update(choices=filtered)

        # Whenever we type in the search box, update the radio with the filtered list
        model_search_box.change(
            fn=filter_models,
            inputs=model_search_box,
            outputs=featured_model_radio
        )

        # Whenever we select a featured model, populate the 'Custom Model' textbox
        featured_model_radio.change(
            fn=set_custom_model_from_radio,
            inputs=featured_model_radio,
            outputs=custom_model_box
        )

print("Gradio interface initialized.")

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