app.py

import gradio as gr
from huggingface_hub import InferenceClient
from huggingface_hub.errors import HfHubHTTPError
import os
import base64
from PIL import Image
import io
from typing import Optional

HF_READ_TOKEN = os.getenv("HF_READ_TOKEN")
if HF_READ_TOKEN:
    print("Default Hugging Face token available from environment.")
else:
    print("No default Hugging Face token configured; relying on user sign-in.")

# 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 render_profile(profile: Optional[gr.OAuthProfile]) -> str:
    """Display the current authentication status in the UI."""
    if profile is None:
        return "Not signed in. Use the button above to sign in with Hugging Face and run inference with your own account."

    display_name = getattr(profile, "name", None) or getattr(profile, "username", "Hugging Face user")
    return f"Signed in as **{display_name}**."

def refresh_auth_info(
    profile: Optional[gr.OAuthProfile],
    oauth_token: Optional[gr.OAuthToken],
):
    """Capture OAuth credentials for downstream callbacks."""
    return oauth_token, render_profile(profile)


def respond(
    message,
    image_files,  # Changed parameter name and structure
    history: list[tuple[str, str]],
    system_message,
    max_tokens,
    temperature,
    top_p,
    frequency_penalty,
    seed,
    custom_model,    
    model_search_term,
    selected_model,
    oauth_token: Optional[gr.OAuthToken] = None
):
    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 model (custom_model): {custom_model}")  
    print(f"Model search term: {model_search_term}")

    api_token: Optional[str] = None
    token_source = "anonymous"

    if oauth_token is not None and getattr(oauth_token, "token", None):
        api_token = oauth_token.token
        token_source = "user"
        print("Using OAuth token from signed-in user for inference.")
    elif HF_READ_TOKEN:
        api_token = HF_READ_TOKEN
        token_source = "space"
        print("Using server-configured Hugging Face token for inference.")
    else:
        print("No default token configured; relying on user sign-in or anonymous access.")

    # Initialize the Inference Client with appropriate credentials
    client_kwargs = {}
    if api_token is not None:
        client_kwargs["token"] = api_token
    else:
        print("No Hugging Face token available; attempting anonymous inference (may fail for private models).")

    client = InferenceClient(**client_kwargs)
    print("Hugging Face Inference Client initialized.")

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

    # Create multimodal content if images are present
    if image_files and len(image_files) > 0:
        # Process the user message to include images
        user_content = []
        
        # Add text part if there is any
        if message and message.strip():
            user_content.append({
                "type": "text",
                "text": message
            })
        
        # Add image parts
        for img in image_files:
            if img is not None:
                # Get raw image data from path
                try:
                    encoded_image = encode_image(img)
                    if encoded_image:
                        user_content.append({
                            "type": "image_url",
                            "image_url": {
                                "url": f"data:image/jpeg;base64,{encoded_image}"
                            }
                        })
                except Exception as e:
                    print(f"Error encoding image: {e}")
    else:
        # Text-only message
        user_content = message

    # 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_part = val[0]
        assistant_part = val[1]
        if user_part:
            # Handle both text-only and multimodal messages in history
            if isinstance(user_part, tuple) and len(user_part) == 2:
                # This is a multimodal message with text and images
                history_content = []
                if user_part[0]:  # Text
                    history_content.append({
                        "type": "text",
                        "text": user_part[0]
                    })
                
                for img in user_part[1]:  # Images
                    if img:
                        try:
                            encoded_img = encode_image(img)
                            if encoded_img:
                                history_content.append({
                                    "type": "image_url",
                                    "image_url": {
                                        "url": f"data:image/jpeg;base64,{encoded_img}"
                                    }
                                })
                        except Exception as e:
                            print(f"Error encoding history image: {e}")
                
                messages.append({"role": "user", "content": history_content})
            else:
                # Regular text message
                messages.append({"role": "user", "content": user_part})
            print(f"Added user message to context (type: {type(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": 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 Hugging Face inference.")

    # 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 HfHubHTTPError as e:
        status = getattr(e.response, "status_code", None)
        if status in (401, 403):
            if token_source == "user":
                raise gr.Error(
                    (
                        "Failed to generate response: {}\n\n"
                        "Your Hugging Face session must grant the **inference-api** (Make calls to Inference Providers) "
                        "permission. Sign out, then sign back in and approve the requested scopes, or update the Space "
                        "metadata to request that scope."
                    ).format(e)
                ) from e
            if token_source == "space":
                raise gr.Error(
                    (
                        "Failed to generate response: {}\n\n"
                        "The Space-level token lacks Inference Provider access. Update the `HF_READ_TOKEN` secret with "
                        "a token that has the `Make calls to Inference Providers` permission, or have users sign in with "
                        "their own accounts."
                    ).format(e)
                ) from e
            raise gr.Error(
                "Failed to generate response: {}. Sign in with your Hugging Face account and retry."
                .format(e)
            ) from e

        print(f"Error during inference: {e}")
        response += f"\nError: {str(e)}"
        yield response
    except Exception as e:
        print(f"Error during inference: {e}")
        response += f"\nError: {str(e)}"
        yield response

    print("Completed response generation.")

# GRADIO UI
with gr.Blocks(theme="Nymbo/Nymbo_Theme") as demo:
    with gr.Row(elem_id="oauth-row"):
        login_button = gr.LoginButton()
        auth_status = gr.Markdown(render_profile(None), elem_id="oauth-status")

    # Create the chatbot component
    chatbot = gr.Chatbot(
        height=600, 
        show_copy_button=True, 
        placeholder="Select a model and begin chatting. Now supports multimodal inputs.",
        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"]
    )

    # 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)"
                )
        
        # 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 = [
            "meta-llama/Llama-3.2-11B-Vision-Instruct",
            "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-3.2-11B-Vision-Instruct",  # Default to a multimodal model
            interactive=True
        )
        
        gr.Markdown("[View all Text-to-Text models](https://huggingface.co/models?inference_provider=all&pipeline_tag=text-generation&sort=trending) | [View all multimodal models](https://huggingface.co/models?inference_provider=all&pipeline_tag=image-text-to-text&sort=trending)")

    # Chat history state
    chat_history = gr.State([])
    oauth_token_state = gr.State(None)
    
    # 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
        
        # Prepare multimodal message format
        text_content = user_message.get("text", "").strip()
        files = user_message.get("files", [])
        
        print(f"Text content: {text_content}")
        print(f"Files: {files}")
        
        # If both text and files are empty, skip
        if not text_content and not files:
            print("No content to display")
            return history
        
        # Add message with images to history
        if files and len(files) > 0:
            # Add text message first if it exists
            if text_content:
                # Add a separate text message
                print(f"Adding text message: {text_content}")
                history.append([text_content, None])
            
            # Then add each image file separately
            for file_path in files:
                if file_path and isinstance(file_path, str):
                    print(f"Adding image: {file_path}")
                    # Add image as a separate message with no text
                    history.append([f"![Image]({file_path})", None])
            
            return history
        else:
            # For text-only messages
            print(f"Adding text-only message: {text_content}")
            history.append([text_content, None])
            return history
    
    # Define bot response function
    def bot(
        history,
        system_msg,
        max_tokens,
        temperature,
        top_p,
        freq_penalty,
        seed,
        custom_model,
        search_term,
        selected_model,
        oauth_token_obj: Optional[gr.OAuthToken] = None,
        *_extra,
    ):
        # Check if history is valid
        if not history or len(history) == 0:
            print("No history to process")
            return history
        
        # Get the most recent message and detect if it's an image
        user_message = history[-1][0]
        print(f"Processing user message: {user_message}")
        
        is_image = False
        image_path = None
        text_content = user_message
        
        # Check if this is an image message (marked with ![Image])
        if isinstance(user_message, str) and user_message.startswith("![Image]("):
            is_image = True
            # Extract image path from markdown format ![Image](path)
            image_path = user_message.replace("![Image](", "").replace(")", "")
            print(f"Image detected: {image_path}")
            text_content = ""  # No text for image-only messages
        
        # Look back for text context if this is an image
        text_context = ""
        if is_image and len(history) > 1:
            # Use the previous message as context if it's text
            prev_message = history[-2][0]
            if isinstance(prev_message, str) and not prev_message.startswith("![Image]("):
                text_context = prev_message
                print(f"Using text context from previous message: {text_context}")
        
        # Process message through respond function
        history[-1][1] = ""
        
        # Use either the image or text for the API
        if is_image:
            # For image messages
            for response in respond(
                text_context,           # Text context from previous message if any
                [image_path],           # Current image
                history[:-1],           # Previous history
                system_msg,
                max_tokens,
                temperature,
                top_p,
                freq_penalty,
                seed,
                custom_model,
                search_term,
                selected_model,
                oauth_token=oauth_token_obj,
            ):
                history[-1][1] = response
                yield history
        else:
            # For text-only messages
            for response in respond(
                text_content,           # Text message
                None,                   # No image
                history[:-1],           # Previous history
                system_msg,
                max_tokens,
                temperature,
                top_p,
                freq_penalty,
                seed,
                custom_model,
                search_term,
                selected_model,
                oauth_token=oauth_token_obj,
            ):
                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, custom_model_box, 
         model_search_box, featured_model_radio, oauth_token_state],
        [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.")

    login_button.click(
        refresh_auth_info,
        inputs=None,
        outputs=[oauth_token_state, auth_status],
    )

    demo.load(refresh_auth_info, inputs=None, outputs=[oauth_token_state, auth_status])

print("Gradio interface initialized.")

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