app.py

import argparse
import datetime
import hashlib
import json
import os
import sys
import time
import warnings

import gradio as gr
import spaces
import torch

from builder import load_pretrained_model
from llava.constants import IMAGE_TOKEN_INDEX
from llava.constants import LOGDIR
from llava.conversation import (default_conversation, conv_templates)
from llava.mm_utils import KeywordsStoppingCriteria, tokenizer_image_token
from llava.utils import (build_logger, violates_moderation, moderation_msg)
from taxonomy import wrap_taxonomy, default_taxonomy


def clear_conv(conv):
    conv.messages = []
    return conv


logger = build_logger("gradio_web_server", "gradio_web_server.log")

headers = {"User-Agent": "LLaVA Client"}

no_change_btn = gr.Button()
enable_btn = gr.Button(interactive=True)
disable_btn = gr.Button(interactive=False)

priority = {
    "LlavaGuard-7B": "aaaaaaa",
    "LlavaGuard-13B": "aaaaaab",
    "LlavaGuard-34B": "aaaaaac",
}


@spaces.GPU
def run_llava(prompt, pil_image, temperature, top_p, max_new_tokens):
    image_size = pil_image.size
    image_tensor = image_processor.preprocess(pil_image, return_tensors='pt')['pixel_values'].half().cuda()
    # image_tensor = image_tensor.to(model.device, dtype=torch.float16)
    input_ids = tokenizer_image_token(prompt, tokenizer, IMAGE_TOKEN_INDEX, return_tensors="pt")
    input_ids = input_ids.unsqueeze(0).cuda()
    with torch.inference_mode():
        output_ids = model.generate(
            input_ids,
            images=image_tensor,
            image_sizes=[image_size],
            do_sample=True,
            temperature=temperature,
            top_p=top_p,
            top_k=50,
            num_beams=2,
            max_new_tokens=max_new_tokens,
            use_cache=True,
            stopping_criteria=[KeywordsStoppingCriteria(['}'], tokenizer, input_ids)]
        )
    outputs = tokenizer.batch_decode(output_ids, skip_special_tokens=True)

    return outputs[0].strip()


def load_selected_model(model_path):
    model_name = model_path.split("/")[-1]
    global tokenizer, model, image_processor, context_len
    with warnings.catch_warnings(record=True) as w:
        warnings.simplefilter("always")
        tokenizer, model, image_processor, context_len = load_pretrained_model(model_path, None, model_name)
    for warning in w:
        if "vision" not in str(warning.message).lower():
            print(warning.message)
    model.config.tokenizer_model_max_length = 2048 * 2


def get_conv_log_filename():
    t = datetime.datetime.now()
    name = os.path.join(LOGDIR, f"{t.year}-{t.month:02d}-{t.day:02d}-conv.json")
    return name


def get_model_list():
    models = [
                 'AIML-TUDA/LlavaGuard-7B',
                 'AIML-TUDA/LlavaGuard-v1.1-7B-hf',
                 'AIML-TUDA/LlavaGuard-13B',
                 'AIML-TUDA/LlavaGuard-v1.1-13B-hf']
    return models


get_window_url_params = """
function() {
    const params = new URLSearchParams(window.location.search);
    url_params = Object.fromEntries(params);
    console.log(url_params);
    return url_params;
    }
"""


def load_demo(url_params, request: gr.Request):
    logger.info(f"load_demo. ip: {request.client.host}. params: {url_params}")

    dropdown_update = gr.Dropdown(visible=True)
    if "model" in url_params:
        model = url_params["model"]
        if model in models:
            dropdown_update = gr.Dropdown(value=model, visible=True)

    state = default_conversation.copy()
    return state, dropdown_update


def load_demo_refresh_model_list(request: gr.Request):
    logger.info(f"load_demo. ip: {request.client.host}")
    models = get_model_list()
    state = default_conversation.copy()
    dropdown_update = gr.Dropdown(
        choices=models,
        value=models[0] if len(models) > 0 else ""
    )
    return state, dropdown_update


def vote_last_response(state, vote_type, model_selector, request: gr.Request):
    with open(get_conv_log_filename(), "a") as fout:
        data = {
            "tstamp": round(time.time(), 4),
            "type": vote_type,
            "model": model_selector,
            "state": state.dict(),
            "ip": request.client.host,
        }
        fout.write(json.dumps(data) + "\n")


def upvote_last_response(state, model_selector, request: gr.Request):
    logger.info(f"upvote. ip: {request.client.host}")
    vote_last_response(state, "upvote", model_selector, request)
    return ("",) + (disable_btn,) * 3


def downvote_last_response(state, model_selector, request: gr.Request):
    logger.info(f"downvote. ip: {request.client.host}")
    vote_last_response(state, "downvote", model_selector, request)
    return ("",) + (disable_btn,) * 3


def flag_last_response(state, model_selector, request: gr.Request):
    logger.info(f"flag. ip: {request.client.host}")
    vote_last_response(state, "flag", model_selector, request)
    return ("",) + (disable_btn,) * 3


def regenerate(state, image_process_mode, request: gr.Request):
    logger.info(f"regenerate. ip: {request.client.host}")
    state.messages[-1][-1] = None
    prev_human_msg = state.messages[-2]
    if type(prev_human_msg[1]) in (tuple, list):
        prev_human_msg[1] = (*prev_human_msg[1][:2], image_process_mode)
    state.skip_next = False
    return (state, state.to_gradio_chatbot(), "", None) + (disable_btn,) * 5


def clear_history(request: gr.Request):
    logger.info(f"clear_history. ip: {request.client.host}")
    state = default_conversation.copy()
    return (state, state.to_gradio_chatbot(), "", None) + (disable_btn,) * 5


def add_text(state, text, image, image_process_mode, request: gr.Request):
    logger.info(f"add_text. ip: {request.client.host}. len: {len(text)}")
    if len(text) <= 0 or image is None:
        state.skip_next = True
        return (state, state.to_gradio_chatbot(), "", None) + (no_change_btn,) * 5
    if args.moderate:
        flagged = violates_moderation(text)
        if flagged:
            state.skip_next = True
            return (state, state.to_gradio_chatbot(), moderation_msg, None) + (
                no_change_btn,) * 5

    text = wrap_taxonomy(text)
    if image is not None:
        text = text  # Hard cut-off for images
        if '<image>' not in text:
            # text = '<Image><image></Image>' + text
            text = text + '\n<image>'
        text = (text, image, image_process_mode)
        state = default_conversation.copy()
    state = clear_conv(state)
    state.append_message(state.roles[0], text)
    state.append_message(state.roles[1], None)
    state.skip_next = False
    return (state, state.to_gradio_chatbot(), default_taxonomy, None) + (disable_btn,) * 5


def llava_bot(state, model_selector, temperature, top_p, max_new_tokens, request: gr.Request):
    start_tstamp = time.time()
    model_name = model_selector

    if state.skip_next:
        # This generate call is skipped due to invalid inputs
        yield (state, state.to_gradio_chatbot()) + (no_change_btn,) * 5
        return

    if len(state.messages) == state.offset + 2:
        # First round of conversation
        if "llava" in model_name.lower():
            if 'llama-2' in model_name.lower():
                template_name = "llava_llama_2"
            elif "mistral" in model_name.lower() or "mixtral" in model_name.lower():
                if 'orca' in model_name.lower():
                    template_name = "mistral_orca"
                elif 'hermes' in model_name.lower():
                    template_name = "chatml_direct"
                else:
                    template_name = "mistral_instruct"
            elif 'llava-v1.6-34b' in model_name.lower():
                template_name = "chatml_direct"
            elif "v1" in model_name.lower():
                if 'mmtag' in model_name.lower():
                    template_name = "v1_mmtag"
                elif 'plain' in model_name.lower() and 'finetune' not in model_name.lower():
                    template_name = "v1_mmtag"
                else:
                    template_name = "llava_v1"
            elif "mpt" in model_name.lower():
                template_name = "mpt"
            else:
                if 'mmtag' in model_name.lower():
                    template_name = "v0_mmtag"
                elif 'plain' in model_name.lower() and 'finetune' not in model_name.lower():
                    template_name = "v0_mmtag"
                else:
                    template_name = "llava_v0"
        elif "mpt" in model_name:
            template_name = "mpt_text"
        elif "llama-2" in model_name:
            template_name = "llama_2"
        else:
            template_name = "vicuna_v1"
        new_state = conv_templates[template_name].copy()
        new_state.append_message(new_state.roles[0], state.messages[-2][1])
        new_state.append_message(new_state.roles[1], None)
        state = new_state

    # Construct prompt
    prompt = state.get_prompt()

    all_images = state.get_images(return_pil=True)
    all_image_hash = [hashlib.md5(image.tobytes()).hexdigest() for image in all_images]
    for image, hash in zip(all_images, all_image_hash):
        t = datetime.datetime.now()
        filename = os.path.join(LOGDIR, "serve_images", f"{t.year}-{t.month:02d}-{t.day:02d}", f"{hash}.jpg")
        if not os.path.isfile(filename):
            os.makedirs(os.path.dirname(filename), exist_ok=True)
            image.save(filename)

    output = run_llava(prompt, all_images[0], temperature, top_p, max_new_tokens)

    state.messages[-1][-1] = output

    yield (state, state.to_gradio_chatbot()) + (disable_btn,) * 5

    finish_tstamp = time.time()
    logger.info(f"{output}")

    with open(get_conv_log_filename(), "a") as fout:
        data = {
            "tstamp": round(finish_tstamp, 4),
            "type": "chat",
            "model": model_name,
            "start": round(start_tstamp, 4),
            "finish": round(finish_tstamp, 4),
            "state": state.dict(),
            "images": all_image_hash,
            "ip": request.client.host,
        }
        fout.write(json.dumps(data) + "\n")


title_markdown = ("""
# LLAVAGUARD: VLM-based Safeguard for Vision Dataset Curation and Safety Assessment
[[Project Page](https://ml-research.github.io/human-centered-genai/projects/llavaguard/index.html)]
[[Code](https://github.com/ml-research/LlavaGuard)]
[[Model](https://huggingface.co/collections/AIML-TUDA/llavaguard-665b42e89803408ee8ec1086)]
[[Dataset](https://huggingface.co/datasets/aiml-tuda/llavaguard)]
[[LavaGuard](https://arxiv.org/abs/2406.05113)]
""")

tos_markdown = ("""
### Terms of use
By using this service, users are required to agree to the following terms:
The service is a research preview intended for non-commercial use only. It only provides limited safety measures and may generate offensive content. It must not be used for any illegal, harmful, violent, racist, or sexual purposes. The service may collect user dialogue data for future research.
Please click the "Flag" button if you get any inappropriate answer! We will collect those to keep improving our moderator.
For an optimal experience, please use desktop computers for this demo, as mobile devices may compromise its quality.
""")

learn_more_markdown = ("""
### License
The service is a research preview intended for non-commercial use only, subject to the model [License](https://github.com/facebookresearch/llama/blob/main/MODEL_CARD.md) of LLaMA, [Terms of Use](https://openai.com/policies/terms-of-use) of the data generated by OpenAI, and [Privacy Practices](https://chrome.google.com/webstore/detail/sharegpt-share-your-chatg/daiacboceoaocpibfodeljbdfacokfjb) of ShareGPT. Please contact us if you find any potential violation.
""")

block_css = """

#buttons button {
    min-width: min(120px,100%);
}

"""

taxonomies = ["Default", "Modified w/ O1 non-violating", "Default message 3"]


def build_demo(embed_mode, cur_dir=None, concurrency_count=10):
    with gr.Accordion("Safety Risk Taxonomy", open=False) as accordion:
        textbox = gr.Textbox(
            label="Safety Risk Taxonomy",
            show_label=True,
            placeholder="Enter your safety policy here",
            container=True,
            value=default_taxonomy,
            lines=50)
    with gr.Blocks(title="LlavaGuard", theme=gr.themes.Default(), css=block_css) as demo:
        state = gr.State()

        if not embed_mode:
            gr.Markdown(title_markdown)

        with gr.Row():
            with gr.Column(scale=3):
                with gr.Row(elem_id="model_selector_row"):
                    model_selector = gr.Dropdown(
                        choices=models,
                        value=models[0] if len(models) > 0 else "",
                        interactive=True,
                        show_label=False,
                        container=False)

                imagebox = gr.Image(type="pil", label="Image", container=False)
                image_process_mode = gr.Radio(
                    ["Crop", "Resize", "Pad", "Default"],
                    value="Default",
                    label="Preprocess for non-square image", visible=False)

                if cur_dir is None:
                    cur_dir = os.path.dirname(os.path.abspath(__file__))
                gr.Examples(examples=[
                    [f"{cur_dir}/examples/image{i}.png"] for i in range(1, 6)
                ], inputs=imagebox)

                with gr.Accordion("Parameters", open=False) as parameter_row:
                    temperature = gr.Slider(minimum=0.0, maximum=1.0, value=0.2, step=0.1, interactive=True,
                                            label="Temperature", )
                    top_p = gr.Slider(minimum=0.0, maximum=1.0, value=0.95, step=0.1, interactive=True, label="Top P", )
                    max_output_tokens = gr.Slider(minimum=0, maximum=1024, value=512, step=64, interactive=True,
                                                  label="Max output tokens", )

            with gr.Column(scale=8):
                chatbot = gr.Chatbot(
                    elem_id="chatbot",
                    label="LLavaGuard Safety Assessment",
                    height=650,
                    layout="panel",
                )
                with gr.Row():
                    with gr.Column(scale=8):
                        textbox.render()
                    with gr.Column(scale=1, min_width=50):
                        submit_btn = gr.Button(value="Send", variant="primary")
                with gr.Row(elem_id="buttons") as button_row:
                    upvote_btn = gr.Button(value="👍  Upvote", interactive=False)
                    downvote_btn = gr.Button(value="👎  Downvote", interactive=False)
                    flag_btn = gr.Button(value="⚠️  Flag", interactive=False)
                    # stop_btn = gr.Button(value="⏹️  Stop Generation", interactive=False)
                    regenerate_btn = gr.Button(value="🔄  Regenerate", interactive=False)
                    clear_btn = gr.Button(value="🗑️  Clear", interactive=False)

        if not embed_mode:
            gr.Markdown(tos_markdown)
            gr.Markdown(learn_more_markdown)
        url_params = gr.JSON(visible=False)

        # Register listeners
        btn_list = [upvote_btn, downvote_btn, flag_btn, regenerate_btn, clear_btn]
        upvote_btn.click(
            upvote_last_response,
            [state, model_selector],
            [textbox, upvote_btn, downvote_btn, flag_btn]
        )
        downvote_btn.click(
            downvote_last_response,
            [state, model_selector],
            [textbox, upvote_btn, downvote_btn, flag_btn]
        )
        flag_btn.click(
            flag_last_response,
            [state, model_selector],
            [textbox, upvote_btn, downvote_btn, flag_btn]
        )

        # model_selector.change(
        #     load_selected_model,
        #     [model_selector],
        # )

        regenerate_btn.click(
            regenerate,
            [state, image_process_mode],
            [state, chatbot, textbox, imagebox] + btn_list
        ).then(
            llava_bot,
            [state, model_selector, temperature, top_p, max_output_tokens],
            [state, chatbot] + btn_list,
            concurrency_limit=concurrency_count
        )

        clear_btn.click(
            clear_history,
            None,
            [state, chatbot, textbox, imagebox] + btn_list,
            queue=False
        )

        textbox.submit(
            add_text,
            [state, textbox, imagebox, image_process_mode],
            [state, chatbot, textbox, imagebox] + btn_list,
            queue=False
        ).then(
            llava_bot,
            [state, model_selector, temperature, top_p, max_output_tokens],
            [state, chatbot] + btn_list,
            concurrency_limit=concurrency_count
        )

        submit_btn.click(
            add_text,
            [state, textbox, imagebox, image_process_mode],
            [state, chatbot, textbox, imagebox] + btn_list
        ).then(
            llava_bot,
            [state, model_selector, temperature, top_p, max_output_tokens],
            [state, chatbot] + btn_list,
            concurrency_limit=concurrency_count
        )

        if args.model_list_mode == "once":
            demo.load(
                load_demo,
                [url_params],
                [state, model_selector],
                js=get_window_url_params
            )
        elif args.model_list_mode == "reload":
            demo.load(
                load_demo_refresh_model_list,
                None,
                [state, model_selector],
                queue=False
            )
        else:
            raise ValueError(f"Unknown model list mode: {args.model_list_mode}")

    return demo


if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument("--host", type=str, default="0.0.0.0")
    parser.add_argument("--port", type=int)
    parser.add_argument("--controller-url", type=str, default="http://localhost:10000")
    parser.add_argument("--concurrency-count", type=int, default=5)
    parser.add_argument("--model-list-mode", type=str, default="reload", choices=["once", "reload"])
    parser.add_argument("--share", action="store_true")
    parser.add_argument("--moderate", action="store_true")
    parser.add_argument("--embed", action="store_true")
    args = parser.parse_args()
    models = []

    title_markdown += """

ONLY WORKS WITH GPU!

Set the environment variable `model` to change the model:
['AIML-TUDA/LlavaGuard-7B'](https://huggingface.co/AIML-TUDA/LlavaGuard-7B),
['AIML-TUDA/LlavaGuard-13B'](https://huggingface.co/AIML-TUDA/LlavaGuard-13B),
['AIML-TUDA/LlavaGuard-34B'](https://huggingface.co/AIML-TUDA/LlavaGuard-34B),
"""
    print(f"args: {args}")
    concurrency_count = int(os.getenv("concurrency_count", 5))
    api_key = os.getenv("token")


    models = get_model_list()
    bits = int(os.getenv("bits", 16))
    model = os.getenv("model", models[0])
    available_devices = os.getenv("CUDA_VISIBLE_DEVICES", "0")
    model_path, model_name = model, model.split("/")[0]
    if api_key:
        cmd = f"huggingface-cli login --token {api_key} --add-to-git-credential"
        os.system(cmd)
    else:
        if '/workspace' not in sys.path:
            sys.path.append('/workspace')
        from llavaguard.hf_utils import set_up_env_and_token
        api_key = set_up_env_and_token(read=True, write=False)
        model_path = '/common-repos/LlavaGuard/models/LlavaGuard-v1.1-7b-full/smid_and_crawled_v2_with_augmented_policies/json-v16/llava'

    print(f"Loading model {model_path}")
    load_selected_model(model_path)
    model.config.tokenizer_model_max_length = 2048 * 2

    exit_status = 0
    try:
        demo = build_demo(embed_mode=False, cur_dir='./', concurrency_count=concurrency_count)
        demo.queue(
            status_update_rate=10,
            api_open=False
        ).launch(
            server_name=args.host,
            server_port=args.port,
            share=args.share
        )

    except Exception as e:
        print(e)
        exit_status = 1
    finally:
        sys.exit(exit_status)