import gradio as gr from openai import OpenAI import os import requests import json ACCESS_TOKEN = os.getenv("HF_TOKEN") print("Access token loaded.") # Initialize the OpenAI client for HF Inference hf_client = OpenAI( base_url="https://api-inference.huggingface.co/v1/", api_key=ACCESS_TOKEN, ) print("HF Inference OpenAI client initialized.") # Cerebras API endpoint CEREBRAS_API_URL = "https://router.huggingface.co/cerebras/v1/chat/completions" def respond( message, history: list[tuple[str, str]], system_message, max_tokens, temperature, top_p, frequency_penalty, seed, custom_model, provider # New parameter for provider selection ): 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}") print(f"Selected provider: {provider}") # Convert seed to None if -1 (meaning random) if seed == -1: seed = None # Prepare messages for API 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 = "" # Handle different providers if provider == "hf-inference": print("Using HF Inference API.") # Use the OpenAI client for HF Inference for message_chunk in hf_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 if token_text is not None: # Handle None values that might come in stream print(f"Received token: {token_text}") response += token_text yield response elif provider == "cerebras": print("Using Cerebras API via HF Router.") # Prepare headers and payload for the Cerebras API headers = { "Authorization": f"Bearer {ACCESS_TOKEN}", "Content-Type": "application/json" } payload = { "model": model_to_use, "messages": messages, "max_tokens": max_tokens, "temperature": temperature, "top_p": top_p, "frequency_penalty": frequency_penalty, "stream": True } if seed is not None: payload["seed"] = seed # Make the streaming request to Cerebras with requests.post( CEREBRAS_API_URL, headers=headers, json=payload, stream=True ) as req: # Handle Server-Sent Events (SSE) format for line in req.iter_lines(): if line: # Skip the "data: " prefix if line.startswith(b'data: '): line = line[6:] # Skip "[DONE]" message if line == b'[DONE]': continue try: # Parse the JSON chunk chunk = json.loads(line) token_text = chunk.get("choices", [{}])[0].get("delta", {}).get("content") if token_text: print(f"Received Cerebras token: {token_text}") response += token_text yield response except json.JSONDecodeError as e: print(f"Error decoding JSON: {e}, Line: {line}") continue print("Completed response generation.") # GRADIO UI chatbot = gr.Chatbot(height=600, show_copy_button=True, placeholder="Select a model and begin chatting", layout="panel") print("Chatbot interface created.") system_message_box = gr.Textbox(value="", placeholder="You are a helpful assistant.", label="System Prompt") 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.", placeholder="meta-llama/Llama-3.3-70B-Instruct" ) # New provider selection radio provider_radio = gr.Radio( choices=["hf-inference", "cerebras"], value="hf-inference", label="Inference Provider", info="Select which inference provider to use" ) 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. """ print(f"Featured model selected: {selected}") return selected 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, provider_radio, # Add provider selection to inputs ], fill_height=True, chatbot=chatbot, theme="Nymbo/Nymbo_Theme", ) print("ChatInterface object created.") with demo: with gr.Accordion("Model Selection", open=False): model_search_box = gr.Textbox( label="Filter Models", placeholder="Search for a featured model...", lines=1 ) print("Model search box created.") models_list = [ "meta-llama/Llama-3.3-70B-Instruct", "meta-llama/Llama-3.1-70B-Instruct", "meta-llama/Llama-3.0-70B-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", "NousResearch/Nous-Hermes-2-Mixtral-8x7B-DPO", "mistralai/Mistral-Nemo-Instruct-2407", "mistralai/Mixtral-8x7B-Instruct-v0.1", "mistralai/Mistral-7B-Instruct-v0.3", "mistralai/Mistral-7B-Instruct-v0.2", "Qwen/Qwen3-235B-A22B", "Qwen/Qwen3-32B", "Qwen/Qwen2.5-72B-Instruct", "Qwen/Qwen2.5-3B-Instruct", "Qwen/Qwen2.5-0.5B-Instruct", "Qwen/QwQ-32B", "Qwen/Qwen2.5-Coder-32B-Instruct", "microsoft/Phi-3.5-mini-instruct", "microsoft/Phi-3-mini-128k-instruct", "microsoft/Phi-3-mini-4k-instruct", "deepseek-ai/DeepSeek-R1-Distill-Qwen-32B", "deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B", "HuggingFaceH4/zephyr-7b-beta", "HuggingFaceTB/SmolLM2-360M-Instruct", "tiiuae/falcon-7b-instruct", "01-ai/Yi-1.5-34B-Chat", ] print("Models list initialized.") 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.") 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) model_search_box.change( fn=filter_models, inputs=model_search_box, outputs=featured_model_radio ) print("Model search box change event linked.") 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.") # Add new accordion for advanced settings including provider selection with gr.Accordion("Advanced Settings", open=False): # The provider_radio is already defined above, we're just adding it to the UI here gr.Markdown("### Inference Provider") gr.Markdown("Select which provider to use for inference. Default is Hugging Face Inference API.") # Provider radio is already included in the additional_inputs gr.Markdown("Note: Different providers may support different models and parameters.") print("Gradio interface initialized.") if __name__ == "__main__": print("Launching the demo application.") demo.launch(show_api=True)