app.py

import gradio as gr
from huggingface_hub import InferenceClient
import os
import json
import base64
from PIL import Image
import io

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

# Function to encode image to base64
def encode_image(image_path):
    if not image_path:
        print("No image path provided")
        return None
    
    try:
        print(f"Encoding image from path: {image_path}")
        
        # If it's already a PIL Image
        if isinstance(image_path, Image.Image):
            image = image_path
        else:
            # Try to open the image file
            image = Image.open(image_path)
        
        # Convert to RGB if image has an alpha channel (RGBA)
        if image.mode == 'RGBA':
            image = image.convert('RGB')
        
        # Encode to base64
        buffered = io.BytesIO()
        image.save(buffered, format="JPEG")
        img_str = base64.b64encode(buffered.getvalue()).decode("utf-8")
        print("Image encoded successfully")
        return img_str
    except Exception as e:
        print(f"Error encoding image: {e}")
        return None

def respond(
    message,
    image_files,
    history: list[tuple[str, str]],
    system_message,
    max_tokens,
    temperature,
    top_p,
    frequency_penalty,
    seed,
    provider,
    custom_api_key,
    custom_model,    
    model_search_term,
    selected_model
):
    print(f"Received message: {message}")
    print(f"Received {len(image_files) if image_files else 0} images")
    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 provider: {provider}")         
    print(f"Custom API Key provided: {bool(custom_api_key.strip())}")
    print(f"Selected model (custom_model): {custom_model}")  
    print(f"Model search term: {model_search_term}")
    print(f"Selected model from radio: {selected_model}")

    # Determine which token to use
    token_to_use = custom_api_key if custom_api_key.strip() != "" else ACCESS_TOKEN
    
    if custom_api_key.strip() != "":
        print("USING CUSTOM API KEY: BYOK token provided by user is being used for authentication")
    else:
        print("USING DEFAULT API KEY: Environment variable HF_TOKEN is being used for authentication")
    
    # Initialize the Inference Client with the provider and appropriate token
    client = InferenceClient(token=token_to_use, provider=provider)
    print(f"Hugging Face Inference Client initialized with {provider} provider.")

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

    # Prepare messages for the API
    user_content = []
    
    # Add text if there is any
    if message and message.strip():
        user_content.append({
            "type": "text",
            "text": message
        })
    
    # Add images if any
    if image_files and len(image_files) > 0:
        for file_path in image_files:
            if not file_path:
                continue
                
            try:
                print(f"Processing image file: {file_path}")
                # For direct file paths, no need to encode as base64
                user_content.append({
                    "type": "image_url",
                    "image_url": {
                        "url": f"file://{file_path}"
                    }
                })
            except Exception as e:
                print(f"Error processing image file: {e}")
    
    # If empty content, set to text only
    if not user_content:
        user_content = ""
    
    # Prepare messages in the format expected by the API
    messages = [{"role": "system", "content": system_message}]
    print("Initial messages array constructed.")

    # Add conversation history to the context
    for val in history:
        user_msg = val[0]
        assistant_msg = val[1]
        
        # Process user message
        if user_msg:
            if isinstance(user_msg, dict) and "text" in user_msg:
                # This is a MultimodalTextbox message
                hist_text = user_msg.get("text", "")
                hist_files = user_msg.get("files", [])
                
                hist_content = []
                if hist_text:
                    hist_content.append({
                        "type": "text",
                        "text": hist_text
                    })
                
                for hist_file in hist_files:
                    if hist_file:
                        hist_content.append({
                            "type": "image_url",
                            "image_url": {
                                "url": f"file://{hist_file}"
                            }
                        })
                
                if hist_content:
                    messages.append({"role": "user", "content": hist_content})
            else:
                # Regular text message
                messages.append({"role": "user", "content": user_msg})
        
        # Process assistant message
        if assistant_msg:
            messages.append({"role": "assistant", "content": assistant_msg})

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

    # Determine which model to use, prioritizing custom_model if provided
    model_to_use = custom_model.strip() if custom_model.strip() != "" else selected_model
    print(f"Model selected for inference: {model_to_use}")

    # Start with an empty string to build the response as tokens stream in
    response = ""
    print(f"Sending request to {provider} provider.")

    # Prepare parameters for the chat completion request
    parameters = {
        "max_tokens": max_tokens,
        "temperature": temperature,
        "top_p": top_p,
        "frequency_penalty": frequency_penalty,
    }
    
    if seed is not None:
        parameters["seed"] = seed

    # Use the InferenceClient for making the request
    try:
        # Create a generator for the streaming response
        stream = client.chat_completion(
            model=model_to_use,
            messages=messages,
            stream=True,
            **parameters
        )
        
        print("Received tokens: ", end="", flush=True)
        
        # Process the streaming response
        for chunk in stream:
            if hasattr(chunk, 'choices') and len(chunk.choices) > 0:
                # Extract the content from the response
                if hasattr(chunk.choices[0], 'delta') and hasattr(chunk.choices[0].delta, 'content'):
                    token_text = chunk.choices[0].delta.content
                    if token_text:
                        print(token_text, end="", flush=True)
                        response += token_text
                        yield response
        
        print()
    except Exception as e:
        print(f"Error during inference: {e}")
        response += f"\nError: {str(e)}"
        yield response

    print("Completed response generation.")

# Function to validate provider selection based on BYOK
def validate_provider(api_key, provider):
    if not api_key.strip() and provider != "hf-inference":
        return gr.update(value="hf-inference")
    return gr.update(value=provider)

# GRADIO UI
with gr.Blocks(theme="Nymbo/Nymbo_Theme") as demo:
    # Create the chatbot component
    chatbot = gr.Chatbot(
        height=600, 
        show_copy_button=True, 
        placeholder="Select a model and begin chatting",
        layout="panel"
    )
    print("Chatbot interface created.")
    
    # Multimodal textbox for messages (combines text and file uploads)
    msg = gr.MultimodalTextbox(
        placeholder="Type a message or upload images...",
        show_label=False,
        container=False,
        scale=12,
        file_types=["image"],
        file_count="multiple",
        sources=["upload"]
    )
    
    # Note: We're removing the separate submit button since MultimodalTextbox has its own

    # Create accordion for settings
    with gr.Accordion("Settings", open=False):
        # System message
        system_message_box = gr.Textbox(
            value="You are a helpful AI assistant that can understand images and text.", 
            placeholder="You are a helpful assistant.",
            label="System Prompt"
        )
        
        # Generation parameters
        with gr.Row():
            with gr.Column():
                max_tokens_slider = gr.Slider(
                    minimum=1,
                    maximum=4096,
                    value=512,
                    step=1,
                    label="Max 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"
                )
                
            with gr.Column():
                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)"
                )
        
        # Provider selection
        providers_list = [
            "hf-inference",  # Default Hugging Face Inference
            "cerebras",      # Cerebras provider
            "together",      # Together AI
            "sambanova",     # SambaNova
            "novita",        # Novita AI
            "cohere",        # Cohere
            "fireworks-ai",  # Fireworks AI
            "hyperbolic",    # Hyperbolic
            "nebius",        # Nebius
        ]
        
        provider_radio = gr.Radio(
            choices=providers_list,
            value="hf-inference",
            label="Inference Provider",
            info="[View all models here](https://huggingface.co/models?inference_provider=all&sort=trending)"
        )
        
        # New BYOK textbox
        byok_textbox = gr.Textbox(
            value="",
            label="BYOK (Bring Your Own Key)",
            info="Enter a custom Hugging Face API key here. When empty, only 'hf-inference' provider can be used.",
            placeholder="Enter your Hugging Face API token",
            type="password"  # Hide the API key for security
        )
        
        # Custom model box
        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"
        )
        
        # Model search
        model_search_box = gr.Textbox(
            label="Filter Models",
            placeholder="Search for a featured model...",
            lines=1
        )
        
        # Featured models list
        # Updated to include multimodal models
        models_list = [
            # Multimodal models
            "meta-llama/Llama-3.3-70B-Vision",
            "Alibaba-NLP/NephilaV-16B-Chat",
            "mistralai/Mistral-Large-Vision-2407",
            "OpenGVLab/InternVL-Chat-V1-5",
            "microsoft/Phi-3.5-vision-instruct",
            "Qwen/Qwen2.5-VL-7B-Instruct",
            "liuhaotian/llava-v1.6-mistral-7b",
            
            # Standard text models
            "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",
        ]

        featured_model_radio = gr.Radio(
            label="Select a model below",
            choices=models_list,
            value="meta-llama/Llama-3.3-70B-Vision",  # Default to a multimodal model
            interactive=True
        )
        
        gr.Markdown("[View all multimodal models](https://huggingface.co/models?pipeline_tag=image-to-text&sort=trending)")

    # Chat history state
    chat_history = gr.State([])
    
    # Function to filter models
    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)

    # Function to set custom model from radio
    def set_custom_model_from_radio(selected):
        print(f"Featured model selected: {selected}")
        return selected

    # Function for the chat interface
    def user(user_message, history):
        # Debug logging for troubleshooting
        print(f"User message received: {user_message}")
        
        # Skip if message is empty (no text and no files)
        if not user_message or (not user_message.get("text") and not user_message.get("files")):
            print("Empty message, skipping")
            return history
        
        # Extract data from the MultimodalTextbox
        text_content = user_message.get("text", "").strip()
        file_paths = user_message.get("files", [])
        
        print(f"Text content: {text_content}")
        print(f"Files: {file_paths}")
        
        # Process the message
        if file_paths and len(file_paths) > 0:
            # We have files - create a multimodal message
            file_path = file_paths[0]  # For simplicity, use the first file
            print(f"Using file: {file_path}")
            
            # Add the message with both text and file as separate components
            history.append([user_message, None])  # Keep the original format for processing
        else:
            # Text-only message
            history.append([{"text": text_content, "files": []}, None])
            
        return history
    
    # Define bot response function
    def bot(history, system_msg, max_tokens, temperature, top_p, freq_penalty, seed, provider, api_key, custom_model, search_term, selected_model):
        # Check if history is valid
        if not history or len(history) == 0:
            print("No history to process")
            return history
        
        # Extract the last user message
        user_message = history[-1][0]
        print(f"Processing user message: {user_message}")
        
        # Get text and files from the message
        if isinstance(user_message, dict) and "text" in user_message:
            text_content = user_message.get("text", "")
            image_files = user_message.get("files", [])
        else:
            text_content = ""
            image_files = []
        
        # Process message through respond function
        history[-1][1] = ""
        for response in respond(
            text_content,
            image_files,
            history[:-1],
            system_msg,
            max_tokens,
            temperature,
            top_p,
            freq_penalty,
            seed,
            provider,
            api_key,
            custom_model,
            search_term,
            selected_model
        ):
            history[-1][1] = response
            yield history

    # Event handlers - only using the MultimodalTextbox's built-in submit functionality
    msg.submit(
        user,
        [msg, chatbot],
        [chatbot],
        queue=False
    ).then(
        bot,
        [chatbot, system_message_box, max_tokens_slider, temperature_slider, top_p_slider, 
         frequency_penalty_slider, seed_slider, provider_radio, byok_textbox, custom_model_box, 
         model_search_box, featured_model_radio],
        [chatbot]
    ).then(
        lambda: {"text": "", "files": []},  # Clear inputs after submission
        None,
        [msg]
    )
    
    # Connect the model filter to update the radio choices
    model_search_box.change(
        fn=filter_models,
        inputs=model_search_box,
        outputs=featured_model_radio
    )
    print("Model search box change event linked.")

    # Connect the featured model radio to update the custom model box
    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.")
    
    # Connect the BYOK textbox to validate provider selection
    byok_textbox.change(
        fn=validate_provider,
        inputs=[byok_textbox, provider_radio],
        outputs=provider_radio
    )
    print("BYOK textbox change event linked.")

    # Also validate provider when the radio changes to ensure consistency
    provider_radio.change(
        fn=validate_provider,
        inputs=[byok_textbox, provider_radio],
        outputs=provider_radio
    )
    print("Provider radio button change event linked.")

print("Gradio interface initialized.")

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