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,
    max_tokens,
    temperature,
    top_p,
    frequency_penalty,
    seed,
    custom_model
):
    """
    This function handles the conversation logic and streams the response.

    Arguments:
    - message: The new user message
    - history: Chat history in the form of a list of (user_message, assistant_message) pairs
    - system_message: The system prompt specifying how the assistant should behave
    - max_tokens, temperature, top_p, frequency_penalty, seed, custom_model: Various parameters for text generation
    """
    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

    # Create the base system-level message
    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.")

    # If user provided a model, use that; otherwise, fall back to a default model
    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.")

    # Stream tokens from the HF inference endpoint
    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 definitions
# -------------------------

# Chatbot interface
chatbot = gr.Chatbot(
    height=600, 
    show_copy_button=True, 
    placeholder="Select a model and begin chatting", 
    likeable=True, 
    layout="panel"
)
print("Chatbot interface created.")

# System prompt textbox
system_message_box = gr.Textbox(
    value="", 
    placeholder="You are a helpful assistant.", 
    label="System Prompt"
)

# Sliders
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)"
)

# This textbox is what the respond() function sees as "custom_model"
# We will visually place it inside the Model Selection accordion (below),
# but we define it here so it can be passed to the ChatInterface.
custom_model_box = gr.Textbox(
    value="",
    label="Custom Model",
    info="(Optional) Provide a custom Hugging Face model path. Overrides any selected featured model.",
    placeholder="meta-llama/Llama-3.3-70B-Instruct"
)

# Create the ChatInterface, referencing the respond function and including all inputs
demo = gr.ChatInterface(
    fn=respond,
    additional_inputs=[
        system_message_box,
        max_tokens_slider,
        temperature_slider,
        top_p_slider,
        frequency_penalty_slider,
        seed_slider,
        custom_model_box,  # We pass it here to the ChatInterface function
    ],
    fill_height=True,
    chatbot=chatbot,
    theme="Nymbo/Nymbo_Theme",
)
print("ChatInterface object created.")


# --------------------------
# Additional Model Selection
# --------------------------

# This is the function that updates the Custom Model textbox whenever the user picks a model from the Radio
def set_custom_model_from_radio(selected):
    """
    Triggered when the user picks a model from the 'Featured Models' radio.
    We will update the Custom Model text box with that selection automatically.
    """
    print(f"Featured model selected: {selected}")
    return selected

# The set of models displayed in the radio
models_list = [
    "meta-llama/Llama-3.3-70B-Instruct",
    "meta-llama/Llama-3.2-3B-Instruct",
    "meta-llama/Llama-3.2-1B-Instruct",
    "meta-llama/Llama-3.1-8B-Instruct",
    "NousResearch/Hermes-3-Llama-3.1-8B",
    "google/gemma-2-27b-it",
    "google/gemma-2-9b-it",
    "google/gemma-2-2b-it",
    "mistralai/Mistral-Nemo-Instruct-2407",
    "mistralai/Mixtral-8x7B-Instruct-v0.1",
    "mistralai/Mistral-7B-Instruct-v0.3",
    "Qwen/Qwen2.5-72B-Instruct",
    "Qwen/QwQ-32B-Preview",
    "PowerInfer/SmallThinker-3B-Preview",
    "HuggingFaceTB/SmolLM2-1.7B-Instruct",
    "TinyLlama/TinyLlama-1.1B-Chat-v1.0",
    "microsoft/Phi-3.5-mini-instruct",
]
print("Models list initialized.")

# This function handles searching for models by a user-provided filter
def filter_models(search_term):
    print(f"Filtering models with search term: {search_term}")
    filtered = [m for m in models_list if search_term.lower() in m.lower()]
    print(f"Filtered models: {filtered}")
    return gr.update(choices=filtered)


# --------------------------------
# Advanced UI arrangement with demo
# --------------------------------
with demo:
    # Create an Accordion for model selection
    with gr.Accordion("Model Selection", open=False):
        # Place the Filter Models textbox and the Custom Model textbox side by side
        with gr.Row():
            model_search_box = gr.Textbox(
                label="Filter Models",
                placeholder="Search for a featured model...",
                lines=1
            )
            # Render the already-defined 'custom_model_box' so it appears in this row
            custom_model_box.render()

        # Create the Radio for featured models
        featured_model_radio = gr.Radio(
            label="Select a model below",
            choices=models_list,
            value="meta-llama/Llama-3.3-70B-Instruct",
            interactive=True
        )
        print("Featured models radio button created.")

        # Link the search box to the filtering function
        model_search_box.change(
            fn=filter_models,
            inputs=model_search_box,
            outputs=featured_model_radio
        )
        print("Model search box change event linked.")

        # Link the radio to the function that sets the custom model textbox
        featured_model_radio.change(
            fn=set_custom_model_from_radio,
            inputs=featured_model_radio,
            outputs=custom_model_box
        )
        print("Featured model radio button change event linked.")

print("Gradio interface initialized.")


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