gradio_web_server.py

import argparse
import datetime
import json
import os
import time

import gradio as gr
import requests
from PIL import Image
import base64
from io import BytesIO

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

worker_addr = os.getenv('WORKER_ADDR')
apikey = os.getenv('AUTHORIZATION')


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

headers = {"Authorization": apikey}

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

priority = {
    "vicuna-13b": "aaaaaaa",
    "koala-13b": "aaaaaab",
}


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.update(visible=True)
    if "model" in url_params:
        model = url_params["model"]
        if model in models:
            dropdown_update = gr.Dropdown.update(
                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[:1536]  # Hard cut-off
    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>'
        text = (text, 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 convert_image_to_base64(image, format=None):
    # 如果未指定格式，则使用图像的原始格式
    if format is None:
        format = "jpeg"

    # 将图像保存到字节流
    buffered = BytesIO()
    image.save(buffered, format=format)

    # 编码为base64
    img_str = base64.b64encode(buffered.getvalue()).decode()

    # 格式化最终的字符串
    return f"data:image/{format.lower()};base64,{img_str}"

# 修改此处请求模型的逻辑
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 "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]
    
    image = convert_image_to_base64(all_images[0])

    # Make requests
    pload = {
        "image_path": image,
        "model": model_name,
        "messages": prompt,
        "stream": True,
        "max_tokens": 512
    }
    logger.info(f"==== request ====\n{json.dumps(pload)}")

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

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

    try:
        # Stream output
        response = requests.post(worker_addr,
            headers=headers, json=pload, timeout=60)
        
        output = json.loads(response.text)['message']['content']
        logger.info("the response is {output}")
        state.messages[-1][-1] = output
        yield (state, state.to_gradio_chatbot()) + (disable_btn, disable_btn, disable_btn, enable_btn, enable_btn)
    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]
    yield (state, state.to_gradio_chatbot()) + (enable_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 = ("""
# 🌋 LLaVA: Large Language and Vision Assistant
[[Project Page](https://llava-vl.github.io)] [[Code](https://github.com/haotian-liu/LLaVA)] [[Model](https://github.com/haotian-liu/LLaVA/blob/main/docs/MODEL_ZOO.md)] | 📚 [[LLaVA](https://arxiv.org/abs/2304.08485)] [[LLaVA-v1.5](https://arxiv.org/abs/2310.03744)]
""")

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%);
}

"""

def build_demo(embed_mode):
    textbox = gr.Textbox(show_label=False, placeholder="Enter text and press ENTER", container=False)
    with gr.Blocks(title="LLaVA", 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")
                image_process_mode = gr.Radio(
                    ["Crop", "Resize", "Pad", "Default"],
                    value="Default",
                    label="Preprocess for non-square image", visible=False)

                cur_dir = os.path.dirname(os.path.abspath(__file__))
                gr.Examples(examples=[
                    [f"{cur_dir}/examples/extreme_ironing.jpg", "What is unusual about this image?"],
                    [f"{cur_dir}/examples/waterview.jpg", "What are the things I should be cautious about when I visit here?"],
                ], inputs=[imagebox, textbox])

                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=0.7, 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="LLaVA Chatbot", height=550)
                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],
            queue=False
        )
        downvote_btn.click(
            downvote_last_response,
            [state, model_selector],
            [textbox, upvote_btn, downvote_btn, flag_btn],
            queue=False
        )
        flag_btn.click(
            flag_last_response,
            [state, model_selector],
            [textbox, upvote_btn, downvote_btn, flag_btn],
            queue=False
        )

        regenerate_btn.click(
            regenerate,
            [state, 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
        )

        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
        )

        submit_btn.click(
            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
        )

        if args.model_list_mode == "once":
            demo.load(
                load_demo,
                [url_params],
                [state, model_selector],
                _js=get_window_url_params,
                queue=False
            )
        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=10)
    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 = ["yi-34b-vl"]

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