Spaces:
qnguyen3
/
Running on Zero

nanoLLaVA / app.py
qnguyen3's picture
Update app.py
0533f81 verified
raw
history blame
20.1 kB
import argparse
import datetime
import json
import os
import time
import gradio as gr
import requests
import hashlib
import pypandoc
import base64
import sys
import spaces
from io import BytesIO
from serve.conversation import (default_conversation, conv_templates, SeparatorStyle)
from serve.constants import LOGDIR
from serve.utils import (build_logger, server_error_msg, violates_moderation, moderation_msg)
import subprocess
subprocess.run('pip install flash-attn --no-build-isolation', env={'FLASH_ATTENTION_SKIP_CUDA_BUILD': "TRUE"}, shell=True)
#logger = build_#logger("gradio_web_server", "gradio_web_server.log")
headers = {"User-Agent": "Bunny Client"}
no_change_btn = gr.update()
enable_btn = gr.update(interactive=True)
disable_btn = gr.update(interactive=False)
priority = {
"Bunny": "aaaaaaa",
}
def start_controller():
print("Starting the controller")
controller_command = [
sys.executable,
"serve/controller.py",
"--host",
"0.0.0.0",
"--port",
"10000",
]
print(controller_command)
return subprocess.Popen(controller_command)
@spaces.GPU
def start_worker(model_path: str):
print(f"Starting the model worker for the model {model_path}")
model_path = 'qnguyen3/nanoLLaVA'
worker_command = [
sys.executable,
"serve/model_worker.py",
"--host",
"0.0.0.0",
"--controller",
"http://localhost:10000",
"--port",
"40000",
"worker",
"http://localhost:40000",
"--model-path",
model_path,
"--model-type",
"qwen1.5-0.5b",
"--use-flash-attn",
]
print(worker_command)
return subprocess.Popen(worker_command)
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.update(visible=True)
if "model" in url_params:
model = url_params["model"]
if model in models:
dropdown_update = gr.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.update(
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 save_conversation(conversation):
print("save_conversation_wrapper is called")
html_content = "<html><body>"
for role, message in conversation.messages:
if isinstance(message, str): # only text
html_content += f"<p><b>{role}</b>: {message}</p>"
elif isinstance(message, tuple): # text+image
text, image_obj, _ = message
# add text
if text:
html_content += f"<p><b>{role}</b>: {text}</p>"
# add image
buffered = BytesIO()
image_obj.save(buffered, format="PNG")
encoded_image = base64.b64encode(buffered.getvalue()).decode()
html_content += f'<img src="data:image/png;base64,{encoded_image}" /><br>'
html_content += "</body></html>"
doc_path = "./conversation.docx"
pypandoc.convert_text(html_content, 'docx', format='html', outputfile=doc_path,
extra_args=["-M2GB", "+RTS", "-K64m", "-RTS"])
return doc_path
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()
#logger.info(f"Input Text: {text}")
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:
template_name = "bunny"
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
#logger.info(f"Processed Input Text: {state.messages[-2][1]}")
# 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(), enable_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": '<|im_end|>', #state.sep if state.sep_style in [SeparatorStyle.PLAIN, ] 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()
print('=========> get_images')
state.messages[-1][-1] = "โ–Œ"
yield (state, state.to_gradio_chatbot()) + (disable_btn,) * 5
print('=========> state', state.messages[-1][-1])
try:
# Stream output
response = requests.post(worker_addr + "/worker_generate_stream",
headers=headers, json=pload, stream=True, timeout=1000)
print("====> response ok")
print("====> response dir", dir(response))
print("====> response", response)
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 + "โ–Œ"
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()) + (enable_btn, enable_btn, enable_btn)
return
time.sleep(0.03)
except requests.exceptions.RequestException as e:
state.messages[-1][-1] = server_error_msg
yield (state, state.to_gradio_chatbot()) + (enable_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 = ("""
# ๐Ÿฐ Bunny: A family of lightweight multimodal models
[๐Ÿ“–[Technical report](https://arxiv.org/abs/2402.11530)] | [๐Ÿ [Code](https://github.com/BAAI-DCAI/Bunny)] | [๐Ÿค—[Model](https://huggingface.co/BAAI/Bunny-v1_0-3B)]
""")
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
This project utilizes certain datasets and checkpoints that are subject to their respective original licenses. Users must comply with all terms and conditions of these original licenses. The content of this project itself is licensed under the Apache license 2.0.
""")
block_css = """
.centered {
text-align: center;
}
#buttons button {
min-width: min(120px,100%);
}
#file-downloader {
min-width: min(120px,100%);
height: 50px;
}
"""
def trigger_download(doc_path):
return doc_path
def build_demo(embed_mode):
textbox = gr.Textbox(show_label=False, placeholder="Enter text and press ENTER", container=False)
with gr.Blocks(title="Bunny", theme=gr.themes.Default(primary_hue="blue", secondary_hue="lime"),
css=block_css) as demo:
state = gr.State()
if not embed_mode:
gr.Markdown(title_markdown)
with gr.Row():
with gr.Column(scale=4):
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,
allow_custom_value=True
)
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/example_1.png", "What is the astronaut holding in his hand?"],
[f"{cur_dir}/examples/example_2.png", "Why is the image funny?"],
], 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", )
file_output = gr.components.File(label="Download Document", visible=True, elem_id="file-downloader")
with gr.Column(scale=8):
chatbot = gr.Chatbot(elem_id="chatbot", label="Bunny Chatbot",
avatar_images=[f"{cur_dir}/examples/user.png", f"{cur_dir}/examples/icon.jpg"],
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="๐Ÿ‘ Upvote", interactive=False)
downvote_btn = gr.Button(value="๐Ÿ‘Ž Downvote", interactive=False)
# stop_btn = gr.Button(value="โน๏ธ Stop Generation", interactive=False)
regenerate_btn = gr.Button(value="๐Ÿ” Regenerate", interactive=False)
clear_btn = gr.Button(value="๐Ÿšฎ Clear", interactive=False)
save_conversation_btn = gr.Button(value="๐Ÿ—ƒ๏ธ Save", 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, regenerate_btn, clear_btn, save_conversation_btn]
upvote_btn.click(
upvote_last_response,
[state, model_selector],
[textbox, upvote_btn, downvote_btn]
)
downvote_btn.click(
downvote_last_response,
[state, model_selector],
[textbox, upvote_btn, downvote_btn]
)
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
)
save_conversation_btn.click(
save_conversation,
inputs=[state],
outputs=file_output
)
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="127.0.0.1")
parser.add_argument("--port", type=int)
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("--controller-url", type=str, default="http://localhost:10000")
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 = ['qnguyen3/nanoLLaVA']
#logger.info(args)
concurrency_count = int(os.getenv("concurrency_count", 5))
controller_proc = start_controller()
model_path = 'qnguyen3/nanoLLaVA'
worker_proc = start_worker(model_path)
time.sleep(10)
exit_status = 0
demo = build_demo(args.embed)
demo.launch(
server_name=args.host,
server_port=args.port,
share=args.share,
debug=True,
max_threads=10)
# )
# except Exception as e:
# print(e)
# exit_status = 1
# finally:
# worker_proc.kill()
# controller_proc.kill()
# sys.exit(exit_status)