app.py

import os
import logging
import re
import gradio as gr
import base64
import io
from PIL import Image
from typing import Iterator

from gateway import request_generation

# Setup logging
logging.basicConfig(level=logging.INFO)

# CONSTANTS
# Get max new tokens from environment variable, if it is not set, default to 2048
MAX_NEW_TOKENS: int = int(os.getenv("MAX_NEW_TOKENS", 2048))

# Get max number of images to be passed in the prompt
MAX_NUM_IMAGES: int = int(os.getenv("MAX_NUM_IMAGES"))
if not MAX_NUM_IMAGES:
    raise EnvironmentError("MAX_NUM_IMAGES is not set. Please set it to 1 or more.")

# Validate environment variables
CLOUD_GATEWAY_API = os.getenv("API_ENDPOINT")
if not CLOUD_GATEWAY_API:
    raise EnvironmentError("API_ENDPOINT is not set.")

MODEL_NAME: str = os.getenv("MODEL_NAME")
if not MODEL_NAME:
    raise EnvironmentError("MODEL_NAME is not set.")

# Get API Key
API_KEY = os.getenv("API_KEY")
if not API_KEY:  # simple check to validate API Key
    raise Exception("API Key not valid.")

# Create a header, avoid declaring multiple times
HEADER = {"x-api-key": f"{API_KEY}"}


def validate_media(message: str, chat_history: list = None) -> bool:
    """Validate the number of image files in the new message.
    Args:
        message (str): input message from the user
        chat_history (list[tuple[str, str]]): entire chat history of the session
    Returns:
        bool: True if the number of image files is less than or equal to MAX_NUM_IMAGES, False otherwise
    """
    image_count = sum(1 for path in message["files"])
    # Check if there are <image> tags in the prompt and add count
    image_count += message["text"].count("<image>")

    if image_count > MAX_NUM_IMAGES:
        gr.Warning(f"You can upload up to {MAX_NUM_IMAGES} images at a time.")
        return False

    # If there are files, check if they are images
    if not all(
        file.lower().endswith((".png", ".jpg", ".jpeg")) for file in message["files"]
    ):
        gr.Warning("Only images are allowed. Format available: PNG, JPG, JPEG")
        return False
    return True


def encode_pil_to_base64(pil_image: Image.Image, format: str) -> str:
    """Encode a PIL image to base64 string.
    Args:
        pil_image (Image.Image): PIL image object
        format (str): format to save the image, defaults to JPEG
    Returns:
        str: base64 encoded string of the image
    """
    buffered = io.BytesIO()

    # Handle potential transparency issues for JPEG or JPG
    if format == "JPEG" and pil_image.mode in ("RGBA", "LA", "P"):
        # Convert to RGB
        pil_image = pil_image.convert("RGB")

    # Define save arguments, including quality for JPEG
    save_kwargs = {"format": format}
    if format == "JPEG":
        save_kwargs["quality"] = 85  # Adjust quality as needed (0-100)

    try:
        pil_image.save(buffered, **save_kwargs)
    except Exception as e:
        print(f"Error saving image to buffer with format {format}: {e}")

    img_str = base64.b64encode(buffered.getvalue()).decode("utf-8")

    # Determine the MIME type based on the format
    mime_format_part = format.lower()
    if mime_format_part == "jpeg":
        mime_type = "image/jpeg"
    elif mime_format_part == "png":
        mime_type = "image/png"
    else:
        gr.Error(f"Unsupported image format: {format}")
        return None

    return f"data:{mime_type};base64,{img_str}"


def process_images(message: list) -> list[dict]:
    """Process images in the message.
    Args:
        message (list): message list containing text and files
    Returns:
        list[dict]: list of dictionaries containing text and image content
    """
    content = []

    # Iterate through the files in the message
    for path in message:
        pil_image = Image.open(path)
        # Get the image format
        image_format = pil_image.format.upper()
        if image_format == "JPG":
            image_format = "JPEG"

        if image_format in ["JPEG", "PNG"]:
            # Converting image to base64
            base64_image_data = encode_pil_to_base64(pil_image, format=image_format)
            content.append(
                {"type": "image_url", "image_url": {"url": base64_image_data}}
            )

    return content


def extract_image_urls_from_tags(message):
    """Extract image URLs from the <image> tags in the message text.
    Args:
        message (str): message text containing <image> tags
    Returns:
        list[str]: list of image URLs extracted from the <image> tags
    """
    # Extract all <image> tags from the message text using regex
    image_urls = re.findall(r"<image>(.*?)</image>", message, re.IGNORECASE | re.DOTALL)
    # Basic cleanup: strip whitespace from found URLs
    image_urls = [url.strip() for url in image_urls]
    return image_urls


def process_new_user_message(message: dict) -> list[dict]:
    """Process the new user message and return a list of dictionaries containing text and image content.
    Args:
        message (dict): message dictionary containing text and files
    Returns:
        list[dict]: list of dictionaries containing text and image content
    """
    # Create the content list messages
    messages = []

    if message["text"]:
        # Remove the <image> tags from the message text
        prompt = re.sub(
            r"<image>.*?</image>", "", message["text"], flags=re.DOTALL | re.IGNORECASE
        ).strip()
        # If the message text is empty after removing <image> tags, return an empty list
        if not prompt:
            gr.Warning("Please insert a prompt.")
            return []
        # If the message text is not empty, append it to the content list
        messages.append({"type": "text", "text": prompt})

        # processing image urls within tags
        image_urls = extract_image_urls_from_tags(message["text"])
        for url in image_urls:
            if not url or not url.lower().startswith(("http://", "https://")):
                continue
            # Append the image URL to the content list
            messages.append({"type": "image_url", "image_url": {"url": url}})

        if message["files"]:
            # If there are files, process the images
            image_content = process_images(message["files"])
            # Append the image content to the messages list
            messages.extend(image_content)

        return messages
    else:
        # If there are no text parts, throw a gr.Warning to insert prompt and return nothing
        gr.Warning("Please insert a prompt.")
        return []


def run(
    message: str,
    chat_history: list,
    system_prompt: str,
    max_new_tokens: int = 1024,
    temperature: float = 0.6,
    frequency_penalty: float = 0.0,
    presence_penalty: float = 0.0,
) -> Iterator[str]:
    """Send a request to backend, fetch the streaming responses and emit to the UI.

    Args:
        message (str): input message from the user
        chat_history (list[tuple[str, str]]): entire chat history of the session
        system_prompt (str): system prompt
        max_new_tokens (int, optional): maximum number of tokens to generate, ignoring the number of tokens in the
                                        prompt. Defaults to 1024.
        temperature (float, optional): the value used to module the next token probabilities. Defaults to 0.6.
        top_p (float, optional): if set to float<1, only the smallest set of most probable tokens with probabilities
                                    that add up to top_p or higher are kept for generation. Defaults to 0.9.
        top_k (int, optional): the number of highest probability vocabulary tokens to keep for top-k-filtering.
                                Defaults to 50.
        repetition_penalty (float, optional): the parameter for repetition penalty. 1.0 means no penalty.
                                Defaults to 1.2.

    Yields:
        Iterator[str]: Streaming responses to the UI
    """
    if not validate_media(message):
        # If the number of image files is not valid, return an empty string
        yield ""
        return

    messages = []
    if system_prompt:
        messages.append(
            {"role": "system", "content": [{"type": "text", "text": system_prompt}]}
        )

    # Append the new user message if it returns anything other than empty string
    content = process_new_user_message(message)
    if content:
        # Append the new user message to the messages list
        messages.append({"role": "user", "content": content})
    else:
        # If the content is empty, return an empty string
        yield ""
        return

    # sample method to yield responses from the llm model
    outputs = []
    for text in request_generation(
        header=HEADER,
        messages=messages,
        max_new_tokens=max_new_tokens,
        temperature=temperature,
        presence_penalty=presence_penalty,
        frequency_penalty=frequency_penalty,
        cloud_gateway_api=CLOUD_GATEWAY_API,
        model_name=MODEL_NAME,
    ):
        outputs.append(text)
        yield "".join(outputs)


examples = [
    ["Plan a three-day trip to Washington DC for Cherry Blossom Festival."],
    ["How many hours does it take a man to eat a Helicopter?"],
    [
        {
            "text": "Write the matplotlib code to generate the same bar chart.",
            "files": ["assets/sample-images/barchart.png"],
        }
    ],
    [
        {
            "text": "Describe the atmosphere of the scene.",
            "files": ["assets/sample-images/06-1.png"],
        }
    ],
    [
        {
            "text": "Write a short story about what might have happened in this house.",
            "files": ["assets/sample-images/08.png"],
        }
    ],
    [
        {
            "text": "Describe the creatures that would live in this world.",
            "files": ["assets/sample-images/10.png"],
        }
    ],
    [
        {
            "text": "Read text in the image.",
            "files": ["assets/sample-images/1.png"],
        }
    ],
    [
        {
            "text": "When is this ticket dated and how much did it cost?",
            "files": ["assets/sample-images/2.png"],
        }
    ],
    [
        {
            "text": "Read the text in the image into markdown.",
            "files": ["assets/sample-images/3.png"],
        }
    ],
    [
        {
            "text": "Evaluate this integral.",
            "files": ["assets/sample-images/4.png"],
        }
    ],
    [
        {
            "text": "Caption this image",
            "files": ["assets/sample-images/01.png"],
        }
    ],
    [
        {
            "text": "What's the sign says?",
            "files": ["assets/sample-images/02.png"],
        }
    ],
    [
        {
            "text": "Compare and contrast the two images.",
            "files": ["assets/sample-images/03.png"],
        }
    ],
    [
        {
            "text": "List all the objects in the image and their colors.",
            "files": ["assets/sample-images/04.png"],
        }
    ],
]

description = f"""
This Space is an Alpha release that demonstrates [Llama-4-Maverick](https://huggingface.co/meta-llama/Llama-4-Maverick-17B-128E-Instruct) model running on AMD MI300 infrastructure. The space is built with Meta Llama 4 [License](https://www.llama.com/llama4/license/). Feel free to play with it!
"""


demo = gr.ChatInterface(
    fn=run,
    type="messages",
    chatbot=gr.Chatbot(type="messages", scale=1, allow_tags=["image"]),
    textbox=gr.MultimodalTextbox(
        file_types=["image"],
        file_count="single" if MAX_NUM_IMAGES == 1 else "multiple",
        autofocus=True,
        placeholder="Type message, drop PNG/JPEG or use <image>URL</image>...",
    ),
    multimodal=True,
    additional_inputs=[
        gr.Textbox(
            label="System prompt",
            # value="You are a highly capable AI assistant. Provide accurate, concise, and fact-based responses that are directly relevant to the user's query. Avoid speculation, ensure logical consistency, and maintain clarity in longer outputs.",
            value="",
            lines=3,
        ),
        gr.Slider(
            label="Max New Tokens",
            minimum=1,
            maximum=MAX_NEW_TOKENS,
            step=1,
            value=2048,
        ),
        gr.Slider(
            label="Temperature",
            minimum=0.1,
            maximum=4.0,
            step=0.1,
            value=0.3,
        ),
        gr.Slider(
            label="Frequency penalty",
            minimum=-2.0,
            maximum=2.0,
            step=0.1,
            value=0.0,
        ),
        gr.Slider(
            label="Presence penalty",
            minimum=-2.0,
            maximum=2.0,
            step=0.1,
            value=0.0,
        ),
    ],
    stop_btn=False,
    title="Llama-4 Maverick Instruct",
    description=description,
    fill_height=True,
    run_examples_on_click=False,
    examples=examples,
    css_paths="style.css",
    cache_examples=False,
)


if __name__ == "__main__":
    demo.queue(
        max_size=int(os.getenv("QUEUE")),
        default_concurrency_limit=int(os.getenv("CONCURRENCY_LIMIT")),
    ).launch()