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, 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' - model: the selected model 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}, Model: {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}) # 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, # Use the selected 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 response print("Completed response generation.") # Create a Chatbot component with a specified height chatbot = gr.Chatbot(height=600) print("Chatbot interface created.") # List of featured models (placeholder models for now) featured_models = [ "meta-llama/Llama-3.3-70B-Instruct", "gpt-3.5-turbo", "gpt-4", "mistralai/Mistral-7B-Instruct-v0.1", "tiiuae/falcon-40b-instruct" ] # Function to filter models based on search input def filter_models(search_term): filtered_models = [m for m in featured_models if search_term.lower() in m.lower()] return gr.update(choices=filtered_models) # Create the Gradio ChatInterface demo = gr.ChatInterface( respond, additional_inputs=[ gr.Textbox(value="", label="System message"), gr.Slider(minimum=1, maximum=4096, 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"), gr.Slider(minimum=-2.0, maximum=2.0, value=0.0, step=0.1, label="Frequency Penalty"), gr.Slider(minimum=-1, maximum=65535, value=-1, step=1, label="Seed (-1 for random)"), gr.Radio(label="Select a model below", value="meta-llama/Llama-3.3-70B-Instruct", choices=featured_models, interactive=True, elem_id="model-radio") ], fill_height=True, chatbot=chatbot, theme="Nymbo/Nymbo_Theme", ) # Add a "Custom Model" text box and "Featured Models" accordion with demo: with gr.Tab("Model Settings"): with gr.Row(): with gr.Column(): # Textbox for custom model input custom_model = gr.Textbox(label="Custom Model", info="Hugging Face model path (optional)", placeholder="username/model-name") # Accordion for selecting featured models with gr.Accordion("Featured Models", open=True): # Textbox for searching models model_search = gr.Textbox(label="Filter Models", placeholder="Search for a featured model...", lines=1, elem_id="model-search-input") # Radio buttons to select the desired model model_radio = gr.Radio(label="Select a model below", value="meta-llama/Llama-3.3-70B-Instruct", choices=featured_models, interactive=True, elem_id="model-radio") # Update model list when search box is used model_search.change(filter_models, inputs=model_search, outputs=model_radio) # Add an "Information" tab with accordions with gr.Tab("Information"): with gr.Row(): # Accordion for "Featured Models" with a table with gr.Accordion("Featured Models (WiP)", open=False): gr.HTML( """

See all available models

Model Name Typical Use Case Notes
meta-llama/Llama-3.3-70B-Instruct General-purpose instruction following High-quality, large-scale model
gpt-3.5-turbo Chat and general text generation Fast and efficient
gpt-4 Advanced text generation State-of-the-art performance
mistralai/Mistral-7B-Instruct-v0.1 Instruction following Lightweight and efficient
tiiuae/falcon-40b-instruct Instruction following High-quality, large-scale model
""" ) # Accordion for "Parameters Overview" with markdown with gr.Accordion("Parameters Overview", open=False): gr.Markdown( """ ## System Message ###### This is the initial prompt that sets the behavior of the model. It can be used to define the tone, style, or role of the assistant. ## Max Tokens ###### This controls the maximum length of the generated response. Higher values allow for longer responses but may take more time to generate. ## Temperature ###### This controls the randomness of the output. Lower values make the model more deterministic, while higher values make it more creative. ## Top-P ###### This controls the diversity of the output by limiting the model to the most likely tokens. Lower values make the output more focused, while higher values allow for more diversity. ## Frequency Penalty ###### This penalizes repeated tokens in the output. Higher values discourage repetition, while lower values allow for more repetitive outputs. ## Seed ###### This sets a fixed seed for reproducibility. A value of -1 means the seed is random. ## Model ###### This selects the model used for text generation. You can choose from featured models or specify a custom model. """ ) print("Gradio interface initialized.") if __name__ == "__main__": print("Launching the demo application.") demo.launch()