How to deploy Stable Diffusion via Docker container using modal client
I recommend using python docker container instead of using venv, since download-output.py does not work properly if the host machine's python version is 3.8 (default version of ubuntu20.04)
modal.com => LOG IN => SETTINGS => New Token => copy the command showed up below => close the window
(Do not lose the copied command, this will be required later...)
huggingface.co => Log In => settings => Access Tokens => New token => copy it
modal.com => LOG IN => SECRETS => Create new secret => Hugging Face => Paste the value (key : HUGGINGFACE_TOKEN) => next => set the secret name to "my-huggingface-secret" => create
Create this directory into your local machine (I named the root directory name stable-diffusion-modal-docker)
.
├── docker/
│ └── modal-cli/
│ └── Dockerfile
├── docker-compose.yml
└── modal-cli/
├── .modal.toml
└── workdir/
├── download-output.py
├── models/
│ ├── Lora/
│ ├── Stable-diffusion/
│ └── VAE/
├── outputs/
│ ├── txt2img-grids/
│ └── txt2img-images/
└── stable-diffusion-webui.py[default]
token_id = "aa-aaaaaa"
token_secret = "bb-bbbbbb"
Please replace aa-aaaaaa and bb-bbbbbb with your previously copied modal token id and secret
(modal token set --token-id aa-aaaaaa --token-secret bb-bbbbbb)
Please make sure any spaces is not included at the end of both id and secret
# If your logging-in user at host has a different User / Group ID,
# please change the both value for UID & GID `1000`
# to your desired ID number.
FROM python:3.10.6
ARG USERNAME=sd-webui
ARG GROUPNAME=sd-webui
ARG UID=1000
ARG GID=1000
RUN groupadd -g $GID $GROUPNAME && \
useradd -m -s /bin/bash -u $UID -g $GID $USERNAME
WORKDIR /home/sd-webui
RUN apt-get update && \
apt-get upgrade -y && \
apt-get install -y vim less tree jq
RUN pip install modal colorama pathlib
WORKDIR /home/sd-webui/workdir
# If you want to use the original `.modal.toml` file,
# located at your host machine's user home directory,
# please replace `source: ./modal-cli/.modal.toml`
# with `source: ~/.modal.toml`
services:
modal-cli:
build:
context: ./docker/modal-cli
dockerfile: Dockerfile
volumes:
- type: bind
source: ./modal-cli/workdir
target: /home/sd-webui/workdir
- type: bind
source: ./modal-cli/.modal.toml
target: /home/sd-webui/.modal.toml
user: sd-webui
tty: true
stdin_open: true
# This script is made referencing these awesome Japanese websites!
# https://zenn.dev/cp20/articles/stable-diffusion-webui-with-modal
# https://note.com/light_impala137/n/n99b014389a69
from colorama import Fore
from pathlib import Path
import modal
import shutil
import subprocess
import sys
import shlex
import os
# Variables definition related to Modal service
stub = modal.Stub("stable-diffusion-webui")
volume_main = modal.NetworkFileSystem.from_name("stable-diffusion-webui-main", create_if_missing=True)
# Paths definition
webui_dir = "/content/stable-diffusion-webui"
webui_model_dir = webui_dir + "/models/Stable-diffusion/"
# Model IDs on Hugging Face
model_ids = [
{
"repo_id": "XpucT/Deliberate",
"model_path": "Deliberate_v2.safetensors",
"config_file_path": "",
"model_name": "Deliberate_v2.safetensors",
},
{
"repo_id": "WarriorMama777/OrangeMixs",
"model_path": "Models/BloodOrangeMix/BloodNightOrangeMix.ckpt",
"config_file_path": "",
"model_name": "BloodNightOrangeMix.ckpt",
},
]
@stub.function(
# For forcing the docker image to rebuild
# https://modal.com/docs/guide/custom-container#forcing-an-image-to-rebuild
# image=modal.Image.from_registry("python:3.10.6-slim", force_build=True)
image=modal.Image.from_registry("python:3.10.6-slim")
.apt_install(
"git",
"libgl1-mesa-dev",
"libglib2.0-0",
"libsm6",
"libxrender1",
"libxext6",
"gcc",
"libcairo2-dev",
"aria2",
)
.run_commands(
"pip install -e git+https://github.com/CompVis/taming-transformers.git@master#egg=taming-transformers"
)
.pip_install(
"blendmodes==2022",
"transformers==4.30.2",
"accelerate==0.21.0",
"basicsr==1.4.2",
"gfpgan==1.3.8",
"gradio==3.41.2",
"numpy==1.23.5",
"Pillow==9.5.0",
"realesrgan==0.3.0",
"torch",
"omegaconf==2.2.3",
"pytorch_lightning==1.9.4",
"scikit-image==0.21.0",
"fonts",
"font-roboto",
"timm==0.9.2",
"piexif==1.1.3",
"einops==0.4.1",
"jsonmerge==1.8.0",
"clean-fid==0.1.35",
"resize-right==0.0.2",
"torchdiffeq==0.2.3",
"kornia==0.6.7",
"lark==1.1.2",
"inflection==0.5.1",
"GitPython==3.1.32",
"torchsde==0.2.5",
"safetensors==0.3.1",
"httpcore==0.15",
"tensorboard==2.9.1",
"taming-transformers==0.0.1",
"clip",
"xformers",
"test-tube",
"diffusers",
"invisible-watermark",
"pyngrok",
"xformers==0.0.16rc425",
"gdown",
"huggingface_hub",
"colorama",
"torchmetrics==0.11.4",
"fastapi==0.94.0",
"open-clip-torch==2.20.0",
"psutil==5.9.5",
"tomesd==0.1.3",
"httpx==0.24.1",
)
.pip_install(
"git+https://github.com/mlfoundations/open_clip.git@bb6e834e9c70d9c27d0dc3ecedeebeaeb1ffad6b"
),
secret=modal.Secret.from_name("my-huggingface-secret"),
network_file_systems={webui_dir: volume_main},
# Designate the target GPU
gpu="A10G",
# gpu=modal.gpu.A10G(count=2),
# gpu=modal.gpu.T4(count=2),
timeout=12000,
)
async def run_stable_diffusion_webui():
print(Fore.CYAN + "\n---------- Start setting up for all models ----------\n")
webui_dir_path = Path(webui_model_dir)
if not webui_dir_path.exists():
# If you encountered RPC failure related error upon git cloning
# subprocess.run("git config --global http.postBuffer 200M", shell=True)
subprocess.run(
f"git clone -b v2.7 https://github.com/camenduru/stable-diffusion-webui {webui_dir}",
shell=True,
)
# Function definition used for downloading files from Hugging face
def download_hf_file(repo_id, filename):
from huggingface_hub import hf_hub_download
download_dir = hf_hub_download(repo_id=repo_id, filename=filename)
return download_dir
for model_id in model_ids:
print(Fore.GREEN + model_id["repo_id"] + " : Start setting up....")
if not Path(webui_model_dir + model_id["model_name"]).exists():
# Download and copy for model files
model_downloaded_dir = download_hf_file(
model_id["repo_id"],
model_id["model_path"],
)
shutil.copy(
model_downloaded_dir,
webui_model_dir + os.path.basename(model_id["model_path"]),
)
if "config_file_path" not in model_id:
continue
if not Path(webui_model_dir + model_id["config_file_path"]).exists():
# Download and copy for config files
config_downloaded_dir = download_hf_file(
model_id["repo_id"], model_id["config_file_path"]
)
shutil.copy(
config_downloaded_dir,
webui_model_dir + os.path.basename(model_id["config_file_path"]),
)
print(Fore.GREEN + model_id["repo_id"] + " : Finished setting up!")
print(Fore.CYAN + "\n---------- Finished setting up for all models ----------\n")
# Activate WebUI
sys.path.append(webui_dir)
sys.argv += shlex.split("--skip-install --xformers")
os.chdir(webui_dir)
from launch import start, prepare_environment
prepare_environment()
# Note that the first argument will be ignored
sys.argv = shlex.split("--a --gradio-debug --share --xformers --skip-version-check")
start()
@stub.local_entrypoint()
def main():
run_stable_diffusion_webui.remote()
# This script is made referencing these awesome Japanese websites!
# https://zenn.dev/cp20/articles/stable-diffusion-webui-with-modal
import os
import modal
import subprocess
from concurrent import futures
stub = modal.Stub("stable-diffusion-webui-download-output")
volume_key = "stable-diffusion-webui-main"
volume = modal.NetworkFileSystem.from_name(volume_key)
webui_dir = "/content/stable-diffusion-webui/"
remote_outputs_dir = "outputs"
output_dir = "./outputs"
@stub.function(
network_file_systems={webui_dir: volume},
)
def list_output_image_path(cache: list[str]):
absolute_remote_outputs_dir = os.path.join(webui_dir, remote_outputs_dir)
image_path_list = []
for root, dirs, files in os.walk(top=absolute_remote_outputs_dir):
for file in files:
if not file.lower().endswith((".png", ".jpg", ".jpeg")):
continue
absolutefilePath = os.path.join(root, file)
relativeFilePath = absolutefilePath[(len(absolute_remote_outputs_dir)) :]
if not relativeFilePath in cache:
image_path_list.append(relativeFilePath.lstrip("/"))
return image_path_list
def download_image_using_modal(image_path: str):
download_dest = os.path.dirname(os.path.join(output_dir, image_path))
os.makedirs(download_dest, exist_ok=True)
subprocess.run(
f"modal nfs get {volume_key} {os.path.join(remote_outputs_dir, image_path)} {download_dest}",
shell=True,
)
@stub.local_entrypoint()
def main():
cache = []
for root, dirs, files in os.walk(top=output_dir):
for file in files:
relativeFilePath = os.path.join(root, file)[len(output_dir) :]
cache.append(relativeFilePath)
image_path_list = list_output_image_path.remote(cache)
print(f"\nTotal of {len(image_path_list)} files are now being downloaded....\n")
future_list = []
with futures.ThreadPoolExecutor(max_workers=10) as executor:
for image_path in image_path_list:
future = executor.submit(download_image_using_modal, image_path=image_path)
future_list.append(future)
_ = futures.as_completed(fs=future_list)
print(f"\nDownload completed!\n")
@project's root directory
docker compose up -d
docker exec -it <container name> bash
modal run stable-diffusion-webui.py
modal run download-output.py
Put the LoRA files into Lora directory and execute the following command
modal nfs put stable-diffusion-webui-main models/Lora/<lora file name> models/Lora/
Alternatively, you can upload whole directory into Modal server - for multiple LoRA files at a time
(this procedure will overwrite the previously located any Lora files)
modal nfs put stable-diffusion-webui-main models/Lora/ models/Lora/
Put the VAE files into VAE directory and execute the following command
modal nfs put stable-diffusion-webui-main models/VAE/<vae file name> models/VAE/
Alternatively, you can upload whole directory into Modal server - for multiple VAE files at a time
(this procedure will overwrite the previously located any VAE files)
modal nfs put stable-diffusion-webui-main models/VAE/ models/VAE/
Put any model files (including base and refiner models) into Stable-diffusion directory and execute the following command
modal nfs put stable-diffusion-webui-main models/Stable-diffusion/<model file name> models/Stable-diffusion/
or, whole directory at once
(this procedure will overwrite the previously located any models)
modal nfs put stable-diffusion-webui-main models/Stable-diffusion/ models/Stable-diffusion/
Put any embedding files into workdir/embeddings/ folder then,
modal nfs put stable-diffusion-webui-main embeddings/<embedding file name> embeddings/
Sometimes upon executing download-output.py, the previously downloaded files are downloaded again, to prevent this, please execute the following command prior to creating any new pictures with stable diffusion webui
modal nfs rm -r stable-diffusion-webui-main outputs/
If you want to rebuild docker image, such as for updating to the latest modules or packages, please refer to here
stable-diffusion-webui.py
image=modal.Image.from_registry("python:3.10.6-slim", force_build=True)
https://qiita.com/fkgw/items/eaa431b974af20b57179
https://zenn.dev/cp20/articles/stable-diffusion-webui-with-modal
https://note.com/light_impala137/n/n99b014389a69
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent