llava/serve/gradio_web_server.py

import argparse
import datetime
import json
import os
import time

import gradio as gr
import requests

from llava.conversation import (default_conversation, conv_templates,
                                   SeparatorStyle)
from llava.constants import LOGDIR
from llava.utils import (build_logger, server_error_msg,
    violates_moderation, moderation_msg)
import hashlib


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

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

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

priority = {
    "vicuna-13b": "aaaaaaa",
    "koala-13b": "aaaaaab",
}
from PIL import Image
import io
import base64


def resize_image(image, default_width=(1344, 896), request_width=None):
    # 如果 request 中指定了 width，则使用传入的值
    if request_width:
        default_width = request_width

    original_width, original_height = image.size

    print("Original size:", original_width, original_height)

    # 根据宽高比决定 resize 逻辑
    if original_width >= original_height:
        # 根据 width 的值进行 resize
        new_width = default_width[0]
        resize_scale = new_width / original_width
        new_height = round(original_height * resize_scale)
    else:
        # 根据 width 的值进行 resize
        new_width = default_width[1]
        resize_scale = new_width / original_width
        new_height = round(original_height * resize_scale)

    # 调整图像大小
    resized_image = image.resize((new_width, new_height))
    print("After initial resize:", new_width, new_height)

    # 如果高度仍然超过 2016，则将图片固定调整为 896x2016
    if new_height > 2016:
        new_width, new_height = 672, 2016
        resized_image = resized_image.resize((new_width, new_height))
        print("Adjusted to fixed size:", new_width, new_height)

    return resized_image


from PIL import Image, ImageDraw


def draw_circle_on_image(image, x, y, radius=20, color=(255, 0, 0)):
    # 获取图片的宽度和高度
    img_width, img_height = image.size

    # 判断 x 坐标是否在图片范围内
    if not (0 <= x <= img_width):
        print(f"x 坐标 {x} 不在图片宽度范围内，直接返回原图。")
        return image

    # 判断 y 坐标是否在图片范围内
    if not (0 <= y <= img_height):
        print(f"y 坐标 {y} 超出了图片高度范围，尝试减去 224。")
        y -= 224
        # 如果调整后的 y 坐标仍然超出范围，返回原图
        if not (0 <= y <= img_height):
            print(f"调整后的 y 坐标 {y} 仍然超出了图片范围，直接返回原图。")
            return image

    # 创建一个可以在图片上绘制的对象
    draw = ImageDraw.Draw(image)

    # 定义圆圈的外接矩形框
    left_up_point = (x - radius, y - radius)
    right_down_point = (x + radius, y + radius)

    # 绘制圆圈 (outline 参数设置圆圈的颜色，width 设置线条粗细)
    draw.ellipse([left_up_point, right_down_point], outline=color, width=5)

    return image,(x,y)

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():
    ret = requests.post(args.controller_url + "/refresh_all_workers")
    assert ret.status_code == 200
    ret = requests.post(args.controller_url + "/list_models")
    models = ret.json()["models"]
    models.sort(key=lambda x: priority.get(x, x))
    logger.info(f"Models: {models}")
    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 and 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 = text[:500]  # Hard cut-off
    text=f"In the screenshot, where are the pixel coordinates (x, y) of the element corresponding to \"{text}\"?"

    if image is not None:
        text = text[:1200]  # Hard cut-off for images
        if '<image>' not in text:
            # text = '<Image><image></Image>' + text
            text = text + '\n<image>'
        resized_image = resize_image(image)
        text = (text, resized_image, image_process_mode)
        if len(state.get_images(return_pil=True)) > 0:
            state = default_conversation.copy()
    state.append_message(state.roles[0], text)
    state.append_message(state.roles[1], None)
    state.skip_next = False
    return (state, state.to_gradio_chatbot(), "", None) + (disable_btn,) * 5


def http_bot(state, model_selector, temperature, top_p, max_new_tokens, request: gr.Request):
    logger.info(f"http_bot. ip: {request.client.host}")
    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

    # Query worker address
    controller_url = args.controller_url
    ret = requests.post(controller_url + "/get_worker_address",
            json={"model": model_name})
    worker_addr = ret.json()["address"]
    logger.info(f"model_name: {model_name}, worker_addr: {worker_addr}")

    # No available worker
    if worker_addr == "":
        state.messages[-1][-1] = server_error_msg
        yield (state, state.to_gradio_chatbot(), disable_btn, disable_btn, disable_btn, enable_btn, enable_btn)
        return

    # 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)

    # Make requests
    pload = {
        "model": model_name,
        "prompt": prompt,
        "temperature": float(temperature),
        "top_p": float(top_p),
        "max_new_tokens": min(int(max_new_tokens), 1536),
        "stop": state.sep if state.sep_style in [SeparatorStyle.SINGLE, SeparatorStyle.MPT] else state.sep2,
        "images": f'List of {len(state.get_images())} images: {all_image_hash}',
    }
    logger.info(f"==== request ====\n{pload}")

    pload['images'] = state.get_images()

    state.messages[-1][-1] = "▌"
    yield (state, state.to_gradio_chatbot()) + (disable_btn,) * 5

    try:
        # Stream output
        full_output = ""
        response = requests.post(worker_addr + "/worker_generate_stream",
            headers=headers, json=pload, stream=True, timeout=10)
        for chunk in response.iter_lines(decode_unicode=False, delimiter=b"\0"):
            if chunk:
                data = json.loads(chunk.decode())
                if data["error_code"] == 0:
                    output = data["text"][len(prompt):].strip()
                    state.messages[-1][-1] = output + "▌"
                    # full_output += output
                    yield (state, state.to_gradio_chatbot()) + (disable_btn,) * 5
                else:
                    output = data["text"] + f" (error_code: {data['error_code']})"
                    state.messages[-1][-1] = output
                    yield (state, state.to_gradio_chatbot()) + (disable_btn, disable_btn, disable_btn, enable_btn, enable_btn)
                    return
                time.sleep(0.03)
        # full_output=state.messages[-1][-1]
        # if "▌" in full_output:
        #     full_output=full_output[:-1]
    except requests.exceptions.RequestException as e:
        state.messages[-1][-1] = server_error_msg
        yield (state, state.to_gradio_chatbot()) + (disable_btn, disable_btn, disable_btn, enable_btn, enable_btn)
        return

    state.messages[-1][-1] = state.messages[-1][-1][:-1]
    full_output=state.messages[-1][-1][:-1]
    yield (state, state.to_gradio_chatbot()) + (enable_btn,) * 5

    # print(f"Complete output: {full_output}")
    # logger.info(f"Complete output: {full_output}")

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

    print(f"Complete output: {full_output}")
    logger.info(f"Complete output: {full_output}")
    full_output=output
    logger.info(f"{output}")

    print(f"Complete output: {full_output}")
    logger.info(f"Complete output: {full_output}")

    original_coord=(0,0)
    try:
        original_coord = eval(full_output)
        logger.info(f"successfully get {original_coord}")
    except Exception as e:
        logger.info(f"{e}")

    if len(all_images) > 0:
        # 假设我们对第一张图片进行 resize 并展示

        resized_image,coordinates = draw_circle_on_image(resize_image(all_images[0]),original_coord[0],original_coord[1])
        # state.append_message(state.roles[1], ("", resized_image,"Default"))
        yield (state, state.to_gradio_chatbot(resized_image,coordinates)) + (enable_btn,) * 5

    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 = ("""
# UGround: Universal Visual Grounding for GUI Agents
[[🏠Project Homepage](https://osu-nlp-group.github.io/UGround/)] [[Code](https://github.com/OSU-NLP-Group/UGround)] [[😊Model](https://huggingface.co/osunlp/UGround)][[📚Paper](https://arxiv.org/abs/2410.05243)]
""")

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. Please contact us if you find any potential violation.
""")

block_css = """

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

#chatbot img {
    max-width: 80%;    /* 宽图片根据宽度调整 */
    max-height: 80vh;  /* 高图片根据视口高度调整 */
    width: auto;        /* 保持宽度自适应 */
    height: auto;       /* 保持高度自适应 */
    object-fit: contain; /* 保持图片宽高比，不失真 */
}

"""

def build_demo(embed_mode, cur_dir=None, concurrency_count=1):
    textbox = gr.Textbox(show_label=False, placeholder="Enter an element description (referring expression) and press ENTER", container=False)
    with gr.Blocks(title="UGround", 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)
                # model_selector="llava-UGround-v1-4bit"

                imagebox = gr.Image(type="pil")
                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}/amazon.jpg",f"Search bar at the top of the page"],
                    [f"{cur_dir}/shopping.jpg", f"delete button for the second item in the cart list"],
                    [f"{cur_dir}/ios.jpg", f"Open Maps"],
                    [f"{cur_dir}/toggle.jpg", f"toggle button labeled by VPN"],
                    [f"{cur_dir}/semantic.jpg", f"Home"],
                    [f"{cur_dir}/accweather.jpg", f"Select May"],
                    [f"{cur_dir}/arxiv.jpg", f"Home"],
                    [f"{cur_dir}/arxiv.jpg", f"Edit the page"],
                    [f"{cur_dir}/ios.jpg", f"icon at the top right corner"],
                    [f"{cur_dir}/health.jpg", f"text labeled by 2023/11/26"],

                    # toggle button labeled by VPN
                    # Button labeled by 2023/11/26
                    # [f"{cur_dir}/examples/waterview.jpg", "What are the things I should be cautious about when I visit here?"],
                ], inputs=[imagebox, textbox],examples_per_page=4)
                # temperature=0
                # top_p=0.7
                # max_output_tokens=16384
                #
                with gr.Accordion("Parameters", open=False) as parameter_row:
                    temperature = gr.Slider(minimum=0.0, maximum=1.0, value=0, step=0.1, interactive=True, label="Temperature",)
                    top_p = gr.Slider(minimum=0.0, maximum=1.0, value=0, 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="UGround Chatbot",
                    height=650,
                    # min_width=400,
                    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]
        )

        regenerate_btn.click(
            regenerate,
            [state, image_process_mode],
            [state, chatbot, textbox, imagebox] + btn_list
        ).then(
            http_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(
            http_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(
            http_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:21001")
    parser.add_argument("--concurrency-count", type=int, default=2)
    parser.add_argument("--model-list-mode", type=str, default="once",
        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()
    logger.info(f"args: {args}")

    models = get_model_list()

    logger.info(args)
    demo = build_demo(args.embed, concurrency_count=args.concurrency_count)
    demo.queue(
        api_open=False
    ).launch(
        server_name=args.host,
        server_port=args.port,
        share=args.share
    )