VL / app.py
zyliu's picture
update app.py
2a9ac22
raw
history blame
30.4 kB
import spaces
import argparse
from ast import parse
import datetime
import json
import os
import time
import hashlib
import re
import gradio as gr
import requests
import random
from filelock import FileLock
from io import BytesIO
from PIL import Image, ImageDraw, ImageFont
from constants import LOGDIR
from utils import (
build_logger,
server_error_msg,
violates_moderation,
moderation_msg,
load_image_from_base64,
get_log_filename,
)
from conversation import Conversation
logger = build_logger("gradio_web_server", "gradio_web_server.log")
headers = {"User-Agent": "InternVL-Chat Client"}
no_change_btn = gr.Button()
enable_btn = gr.Button(interactive=True)
disable_btn = gr.Button(interactive=False)
@spaces.GPU(duration=10)
def make_zerogpu_happy():
pass
def write2file(path, content):
lock = FileLock(f"{path}.lock")
with lock:
with open(path, "a") as fout:
fout.write(content)
def sort_models(models):
def custom_sort_key(model_name):
# InternVL-Chat-V1-5 should be the first item
if model_name == "InternVL2-Pro":
return (2, model_name) # 2 indicates highest precedence
elif model_name.startswith("InternVL2-8B"):
return (1, model_name) # 0 indicates highest precedence
else:
return (0, model_name) # 0 indicates normal order
models.sort(key=custom_sort_key, reverse=True)
# try: # We have five InternVL-Chat-V1-5 models, randomly choose one to be the first
# first_three = models[:4]
# random.shuffle(first_three)
# models[:4] = first_three
# except:
# pass
return models
def get_model_list():
logger.info(f"Call `get_model_list`")
ret = requests.post(args.controller_url + "/refresh_all_workers")
logger.info(f"status_code from `get_model_list`: {ret.status_code}")
assert ret.status_code == 200
ret = requests.post(args.controller_url + "/list_models")
logger.info(f"status_code from `list_models`: {ret.status_code}")
models = ret.json()["models"]
models = sort_models(models)
logger.info(f"Models (from {args.controller_url}): {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 init_state(state=None):
if state is not None:
del state
return Conversation()
def find_bounding_boxes(state, response):
pattern = re.compile(r"<ref>\s*(.*?)\s*</ref>\s*<box>\s*(\[\[.*?\]\])\s*</box>")
matches = pattern.findall(response)
results = []
for match in matches:
results.append((match[0], eval(match[1])))
returned_image = None
latest_image = state.get_images(source=state.USER)[-1]
returned_image = latest_image.copy()
width, height = returned_image.size
draw = ImageDraw.Draw(returned_image)
for result in results:
line_width = max(1, int(min(width, height) / 200))
random_color = (
random.randint(0, 128),
random.randint(0, 128),
random.randint(0, 128),
)
category_name, coordinates = result
coordinates = [
(
float(x[0]) / 1000,
float(x[1]) / 1000,
float(x[2]) / 1000,
float(x[3]) / 1000,
)
for x in coordinates
]
coordinates = [
(
int(x[0] * width),
int(x[1] * height),
int(x[2] * width),
int(x[3] * height),
)
for x in coordinates
]
for box in coordinates:
draw.rectangle(box, outline=random_color, width=line_width)
font = ImageFont.truetype("assets/SimHei.ttf", int(20 * line_width / 2))
text_size = font.getbbox(category_name)
text_width, text_height = (
text_size[2] - text_size[0],
text_size[3] - text_size[1],
)
text_position = (box[0], max(0, box[1] - text_height))
draw.rectangle(
[
text_position,
(text_position[0] + text_width, text_position[1] + text_height),
],
fill=random_color,
)
draw.text(text_position, category_name, fill="white", font=font)
return returned_image if len(matches) > 0 else None
def query_image_generation(response, sd_worker_url, timeout=15):
if not sd_worker_url:
return None
sd_worker_url = f"{sd_worker_url}/generate_image/"
pattern = r"```drawing-instruction\n(.*?)\n```"
match = re.search(pattern, response, re.DOTALL)
if match:
payload = {"caption": match.group(1)}
print("drawing-instruction:", payload)
response = requests.post(sd_worker_url, json=payload, timeout=timeout)
response.raise_for_status() # 检查HTTP请求是否成功
image = Image.open(BytesIO(response.content))
return image
else:
return None
def load_demo(url_params, request: gr.Request = None):
if not 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 = init_state()
return state, dropdown_update
def load_demo_refresh_model_list(request: gr.Request = None):
if not request:
logger.info(f"load_demo. ip: {request.client.host}")
models = get_model_list()
state = init_state()
dropdown_update = gr.Dropdown(
choices=models, value=models[0] if len(models) > 0 else ""
)
return state, dropdown_update
def vote_last_response(state, liked, model_selector, request: gr.Request):
conv_data = {
"tstamp": round(time.time(), 4),
"like": liked,
"model": model_selector,
"state": state.dict(),
"ip": request.client.host,
}
write2file(get_log_filename(), json.dumps(conv_data) + "\n")
def upvote_last_response(state, model_selector, request: gr.Request):
logger.info(f"upvote. ip: {request.client.host}")
vote_last_response(state, True, model_selector, request)
textbox = gr.MultimodalTextbox(value=None, interactive=True)
return (textbox,) + (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, False, model_selector, request)
textbox = gr.MultimodalTextbox(value=None, interactive=True)
return (textbox,) + (disable_btn,) * 3
def vote_selected_response(
state, model_selector, request: gr.Request, data: gr.LikeData
):
logger.info(
f"Vote: {data.liked}, index: {data.index}, value: {data.value} , ip: {request.client.host}"
)
conv_data = {
"tstamp": round(time.time(), 4),
"like": data.liked,
"index": data.index,
"model": model_selector,
"state": state.dict(),
"ip": request.client.host,
}
write2file(get_log_filename(), json.dumps(conv_data) + "\n")
return
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)
textbox = gr.MultimodalTextbox(value=None, interactive=True)
return (textbox,) + (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
state.update_message(Conversation.ASSISTANT, None, -1)
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
textbox = gr.MultimodalTextbox(value=None, interactive=True)
return (state, state.to_gradio_chatbot(), textbox) + (disable_btn,) * 5
def clear_history(request: gr.Request):
logger.info(f"clear_history. ip: {request.client.host}")
state = init_state()
textbox = gr.MultimodalTextbox(value=None, interactive=True)
return (state, state.to_gradio_chatbot(), textbox) + (disable_btn,) * 5
def change_system_prompt(state, system_prompt, request: gr.Request):
logger.info(f"Change system prompt. ip: {request.client.host}")
state.set_system_message(system_prompt)
return state
def add_text(state, message, system_prompt, model_selector, request: gr.Request):
print(f"state: {state}")
if not state:
state = init_state()
images = message.get("files", [])
text = message.get("text", "").strip()
logger.info(f"add_text. ip: {request.client.host}. len: {len(text)}")
# import pdb; pdb.set_trace()
textbox = gr.MultimodalTextbox(value=None, interactive=False)
if len(text) <= 0 and len(images) == 0:
state.skip_next = True
return (state, state.to_gradio_chatbot(), textbox) + (no_change_btn,) * 5
if args.moderate:
flagged = violates_moderation(text)
if flagged:
state.skip_next = True
textbox = gr.MultimodalTextbox(
value={"text": moderation_msg}, interactive=True
)
return (state, state.to_gradio_chatbot(), textbox) + (no_change_btn,) * 5
images = [Image.open(path).convert("RGB") for path in images]
if len(images) > 0 and len(state.get_images(source=state.USER)) > 0:
state = init_state(state)
state.set_system_message(system_prompt)
state.append_message(Conversation.USER, text, images)
state.skip_next = False
return (state, state.to_gradio_chatbot(), textbox, model_selector) + (
disable_btn,
) * 5
def http_bot(
state,
model_selector,
temperature,
top_p,
repetition_penalty,
max_new_tokens,
max_input_tiles,
# bbox_threshold,
# mask_threshold,
request: gr.Request,
):
logger.info(f"http_bot. ip: {request.client.host}")
start_tstamp = time.time()
model_name = model_selector
if hasattr(state, "skip_next") and state.skip_next:
# This generate call is skipped due to invalid inputs
yield (
state,
state.to_gradio_chatbot(),
gr.MultimodalTextbox(interactive=False),
) + (no_change_btn,) * 5
return
# 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"]
if worker_addr.startswith("http://0.0.0.0") and args.worker_ip:
worker_addr = worker_addr.replace("0.0.0.0", args.worker_ip)
logger.info(f"model_name: {model_name}, worker_addr: {worker_addr}")
# No available worker
if worker_addr == "":
# state.messages[-1][-1] = server_error_msg
state.update_message(Conversation.ASSISTANT, server_error_msg)
yield (
state,
state.to_gradio_chatbot(),
gr.MultimodalTextbox(interactive=False),
disable_btn,
disable_btn,
disable_btn,
enable_btn,
enable_btn,
)
return
all_images = state.get_images(source=state.USER)
all_image_paths = [state.save_image(image) for image in all_images]
# Make requests
pload = {
"model": model_name,
"prompt": state.get_prompt(),
"temperature": float(temperature),
"top_p": float(top_p),
"max_new_tokens": max_new_tokens,
"max_input_tiles": max_input_tiles,
# "bbox_threshold": bbox_threshold,
# "mask_threshold": mask_threshold,
"repetition_penalty": repetition_penalty,
"images": f"List of {len(all_images)} images: {all_image_paths}",
}
logger.info(f"==== request ====\n{pload}")
pload.pop("images")
pload["prompt"] = state.get_prompt(inlude_image=True)
state.append_message(Conversation.ASSISTANT, state.streaming_placeholder)
yield (
state,
state.to_gradio_chatbot(),
gr.MultimodalTextbox(interactive=False),
) + (disable_btn,) * 5
try:
# Stream output
response = requests.post(
worker_addr + "/worker_generate_stream",
headers=headers,
json=pload,
stream=True,
timeout=20,
)
for chunk in response.iter_lines(decode_unicode=False, delimiter=b"\0"):
if chunk:
data = json.loads(chunk.decode())
if data["error_code"] == 0:
if "text" in data:
output = data["text"].strip()
output += state.streaming_placeholder
image = None
if "image" in data:
image = load_image_from_base64(data["image"])
_ = state.save_image(image)
state.update_message(Conversation.ASSISTANT, output, image)
yield (
state,
state.to_gradio_chatbot(),
gr.MultimodalTextbox(interactive=False),
) + (disable_btn,) * 5
else:
output = (
f"**{data['text']}**" + f" (error_code: {data['error_code']})"
)
state.update_message(Conversation.ASSISTANT, output, None)
yield (
state,
state.to_gradio_chatbot(),
gr.MultimodalTextbox(interactive=True),
) + (
disable_btn,
disable_btn,
disable_btn,
enable_btn,
enable_btn,
)
return
except requests.exceptions.RequestException as e:
state.update_message(Conversation.ASSISTANT, server_error_msg, None)
yield (
state,
state.to_gradio_chatbot(),
gr.MultimodalTextbox(interactive=True),
) + (
disable_btn,
disable_btn,
disable_btn,
enable_btn,
enable_btn,
)
return
ai_response = state.return_last_message()
if "<ref>" in ai_response:
returned_image = find_bounding_boxes(state, ai_response)
returned_image = [returned_image] if returned_image else []
state.update_message(Conversation.ASSISTANT, ai_response, returned_image)
if "```drawing-instruction" in ai_response:
returned_image = query_image_generation(
ai_response, sd_worker_url=sd_worker_url
)
returned_image = [returned_image] if returned_image else []
state.update_message(Conversation.ASSISTANT, ai_response, returned_image)
state.end_of_current_turn()
yield (
state,
state.to_gradio_chatbot(),
gr.MultimodalTextbox(interactive=True),
) + (enable_btn,) * 5
finish_tstamp = time.time()
logger.info(f"{output}")
data = {
"tstamp": round(finish_tstamp, 4),
"like": None,
"model": model_name,
"start": round(start_tstamp, 4),
"finish": round(start_tstamp, 4),
"state": state.dict(),
"images": all_image_paths,
"ip": request.client.host,
}
write2file(get_log_filename(), json.dumps(data) + "\n")
title_html = """
<h2> <span class="gradient-text" id="text">InternVL2</span><span class="plain-text">: Better than the Best—Expanding Performance Boundaries of Open-Source Multimodal Models with the Progressive Scaling Strategy</span></h2>
<a href="https://internvl.github.io/blog/2024-07-02-InternVL-2.0/">[📜 InternVL2 Blog]</a>
<a href="https://huggingface.co/spaces/OpenGVLab/InternVL">[🤗 HF Demo]</a>
<a href="https://github.com/OpenGVLab/InternVL?tab=readme-ov-file#quick-start-with-huggingface">[🚀 Quick Start]</a>
<a href="https://github.com/OpenGVLab/InternVL/blob/main/document/How_to_use_InternVL_API.md">[🌐 API]</a>
"""
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.
### Acknowledgement
This demo is modified from LLaVA's demo. Thanks for their awesome work!
"""
# .gradio-container {margin: 5px 10px 0 10px !important};
block_css = """
.gradio-container {margin: 0.1% 1% 0 1% !important; max-width: 98% !important;};
#buttons button {
min-width: min(120px,100%);
}
.gradient-text {
font-size: 28px;
width: auto;
font-weight: bold;
background: linear-gradient(45deg, red, orange, yellow, green, blue, indigo, violet);
background-clip: text;
-webkit-background-clip: text;
color: transparent;
}
.plain-text {
font-size: 22px;
width: auto;
font-weight: bold;
}
"""
js = """
function createWaveAnimation() {
const text = document.getElementById('text');
var i = 0;
setInterval(function() {
const colors = [
'red, orange, yellow, green, blue, indigo, violet, purple',
'orange, yellow, green, blue, indigo, violet, purple, red',
'yellow, green, blue, indigo, violet, purple, red, orange',
'green, blue, indigo, violet, purple, red, orange, yellow',
'blue, indigo, violet, purple, red, orange, yellow, green',
'indigo, violet, purple, red, orange, yellow, green, blue',
'violet, purple, red, orange, yellow, green, blue, indigo',
'purple, red, orange, yellow, green, blue, indigo, violet',
];
const angle = 45;
const colorIndex = i % colors.length;
text.style.background = `linear-gradient(${angle}deg, ${colors[colorIndex]})`;
text.style.webkitBackgroundClip = 'text';
text.style.backgroundClip = 'text';
text.style.color = 'transparent';
text.style.fontSize = '28px';
text.style.width = 'auto';
text.textContent = 'InternVL2';
text.style.fontWeight = 'bold';
i += 1;
}, 200);
const params = new URLSearchParams(window.location.search);
url_params = Object.fromEntries(params);
// console.log(url_params);
// console.log('hello world...');
// console.log(window.location.search);
// console.log('hello world...');
// alert(window.location.search)
// alert(url_params);
return url_params;
}
"""
def build_demo(embed_mode):
textbox = gr.MultimodalTextbox(
interactive=True,
file_types=["image", "video"],
placeholder="Enter message or upload file...",
show_label=False,
)
with gr.Blocks(
title="InternVL-Chat",
theme=gr.themes.Default(),
css=block_css,
) as demo:
models = get_model_list()
state = gr.State()
if not embed_mode:
# gr.Markdown(title_markdown)
gr.HTML(title_html)
with gr.Row():
with gr.Column(scale=2):
with gr.Row(elem_id="model_selector_row"):
model_selector = gr.Dropdown(
choices=models,
value=models[0] if len(models) > 0 else "",
# value="InternVL-Chat-V1-5",
interactive=True,
show_label=False,
container=False,
)
with gr.Accordion("System Prompt", open=False) as system_prompt_row:
system_prompt = gr.Textbox(
value="请尽可能详细地回答用户的问题。",
label="System Prompt",
interactive=True,
)
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",
)
repetition_penalty = gr.Slider(
minimum=1.0,
maximum=1.5,
value=1.1,
step=0.02,
interactive=True,
label="Repetition penalty",
)
max_output_tokens = gr.Slider(
minimum=0,
maximum=4096,
value=1024,
step=64,
interactive=True,
label="Max output tokens",
)
max_input_tiles = gr.Slider(
minimum=1,
maximum=32,
value=12,
step=1,
interactive=True,
label="Max input tiles (control the image size)",
)
examples = gr.Examples(
examples=[
[
{
"files": [
"gallery/prod_9.jpg",
],
"text": "What's at the far end of the image?",
}
],
[
{
"files": [
"gallery/astro_on_unicorn.png",
],
"text": "What does this image mean?",
}
],
[
{
"files": [
"gallery/prod_12.png",
],
"text": "What are the consequences of the easy decisions shown in this image?",
}
],
[
{
"files": [
"gallery/water.jpg",
],
"text": "Please describe this image.",
}
],
],
inputs=[textbox],
)
with gr.Column(scale=8):
chatbot = gr.Chatbot(
elem_id="chatbot",
label="InternVL2",
height=580,
show_copy_button=True,
show_share_button=True,
avatar_images=[
"assets/human.png",
"assets/assistant.png",
],
bubble_full_width=False,
)
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],
)
chatbot.like(
vote_selected_response,
[state, model_selector],
[],
)
flag_btn.click(
flag_last_response,
[state, model_selector],
[textbox, upvote_btn, downvote_btn, flag_btn],
)
regenerate_btn.click(
regenerate,
[state, system_prompt],
[state, chatbot, textbox] + btn_list,
).then(
http_bot,
[
state,
model_selector,
temperature,
top_p,
repetition_penalty,
max_output_tokens,
max_input_tiles,
# bbox_threshold,
# mask_threshold,
],
[state, chatbot, textbox] + btn_list,
)
clear_btn.click(clear_history, None, [state, chatbot, textbox] + btn_list)
textbox.submit(
add_text,
[state, textbox, system_prompt, model_selector],
[state, chatbot, textbox, model_selector] + btn_list,
).then(
http_bot,
[
state,
model_selector,
temperature,
top_p,
repetition_penalty,
max_output_tokens,
max_input_tiles,
# bbox_threshold,
# mask_threshold,
],
[state, chatbot, textbox] + btn_list,
)
submit_btn.click(
add_text,
[state, textbox, system_prompt, model_selector],
[state, chatbot, textbox, model_selector] + btn_list,
).then(
http_bot,
[
state,
model_selector,
temperature,
top_p,
repetition_penalty,
max_output_tokens,
max_input_tiles,
# bbox_threshold,
# mask_threshold,
],
[state, chatbot, textbox] + btn_list,
)
# NOTE: The following code will be not triggered when deployed on HF space.
# It's very strange. I don't know why.
"""
if args.model_list_mode == "once":
demo.load(
load_demo,
[url_params],
[state, model_selector],
js=js,
)
elif args.model_list_mode == "reload":
demo.load(
load_demo_refresh_model_list,
None,
[state, model_selector],
js=js,
)
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, default=7860)
parser.add_argument("--controller-url", type=str, default=None)
parser.add_argument("--worker-ip", type=str, default=None)
parser.add_argument("--concurrency-count", type=int, default=10)
parser.add_argument(
"--model-list-mode", type=str, default="reload", choices=["once", "reload"]
)
parser.add_argument("--sd-worker-url", type=str, default=None)
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}")
if not args.controller_url:
args.controller_url = os.environ.get("CONTROLLER_URL", None)
if not args.controller_url:
raise ValueError("controller-url is required.")
if not args.worker_ip:
args.worker_ip = os.environ.get("WORKER_IP", None)
sd_worker_url = args.sd_worker_url
logger.info(args)
demo = build_demo(args.embed)
demo.queue(api_open=False).launch(
server_name=args.host,
server_port=args.port,
share=args.share,
max_threads=args.concurrency_count,
)