Cresset: The One Template to Train Them All

TL;DR

A new MLOps system for deep learning development using Docker Compose with the aim of providing reproducible and easy-to-use interactive development environments for deep learning practitioners. Hopefully, the methods presented here will become best practice in both academia and industry.

Introductory Video (In English)

Installation on a New Host

If this is your first time using this project, follow these steps:

1. Install the NVIDIA CUDA Driver appropriate for the target host and NVIDIA GPU. If the driver has already been installed, check that the installed version is compatible with the target CUDA version. CUDA driver version mismatch is the single most common issue for new users. See the compatibility matrix for compatible versions of the CUDA driver and CUDA Toolkit.

2. Install Docker (v23.0+ is recommended) or update to a recent version compatible with Docker Compose V2. Docker incompatibility with Docker Compose V2 is another common issue for new users. Note that Windows users may use WSL (Windows Subsystem for Linux). Cresset has been tested on Windows 11 WSL2 with the Windows CUDA driver using Docker Desktop for Windows. There is no need to install a separate WSL CUDA driver or Docker for Linux inside WSL. Note that only Docker Desktop is under a commercial EULA and Docker Engine (for Linux) and Lima Docker (for Mac) are still both open-source. N.B. Windows Security real-time protection causes significant slowdown if enabled. Disable any active antivirus programs on Windows for best performance. N.B. Linux hosts may also install via this repo.

3. Run make install-compose to install Docker Compose V2 for Linux hosts. Installation does not require root permissions. Visit the documentation for the latest installation information. Note that Docker Compose V2 is available for WSL users with Docker Desktop by default.

4. Run make env SERVICE=(train|devel|ngc|hub|simple) on the terminal at project root to create a basic .env file. The .env file provides environment variables for docker-compose.yaml, allowing different users and machines to set their own variables as required. The Makefile has also been configured to read values from the .env file if it exists, allowing non-default values to be specified only once. Each host should have a separate .env file for host-specific configurations.

5. Run make over to create a docker-compose.override.yaml file. Add configurations that should not be shared via source control there. For example, volume-mount pairs specific to each host machine.

6. If Cresset is being placed within a pre-existing project's subdirectory, change the volume pairing from .:${PROJECT_ROOT} to ..:${PROJECT_ROOT}. All commands in Cresset assume that they are being run at project root but this can be changed easily.

Explanation of services

Different Docker Compose services are organized to serve different needs.

• train, the default service, should be used when compiled dependencies are necessary or when PyTorch needs to be compiled from source due to Compute Capability issues, etc.
• devel is designed for PyTorch CUDA/C++ developers who need to recompile frequently and have many complex dependencies.
• ngc is derived from the official NVIDIA PyTorch NGC images with the option to install additional packages. It is recommended for users who wish to base their projects on the NGC images provided by NVIDIA. Note that the NGC images change between different releases and that configurations for one release may not work for another one.
• hub is derived from the official PyTorch Docker Hub image and serves a similar function as the ngc service described above. However, the PyTorch Docker images have a more stable interface than the NGC images.
• simple is derived from the Official Ubuntu Linux image by default as some corporations restrict the use of Docker images not officially verified by Docker. It installs all packages via conda by default and can optionally install highly reproducible environments via conda-lock. Note that pip packages can also be installed via conda. Also, the base image can be configured to use images other than the Official Linux Docker images by specifying the BASE_IMAGE argument directly in the .env file. PyTorch runtime performance may be superior in official NVIDIA CUDA images under certain circumstances. Use the tests to benchmark runtime speeds. The simple service is recommended for users without compiled dependencies.

The Makefile has been configured to take values specified in the .env file if the .env file exists. Therefore, all make commands will automatically use the ${SERVICE} specified by make env SERVICE=${SERVICE} after the .env file is created.

Notes for Rootless Users

Many institutions forbid the use of Docker because it requires root permissions, compromising security. For users without Docker root access, using rootless Docker link is recommended.

While installing rootless Docker requires root permissions on the host, root permissions are not necessary after the initial installation.

When using rootless Docker, it is most convenient to set ADD_USER=exclude in the .env file as the root user will be the host user in rootless Docker.

Project Configuration

1. To build PyTorch from source, set BUILD_MODE=include and the CUDA Compute Capability (CCA) of the target NVIDIA GPU in the .env file. Visit the NVIDIA website to find compute capabilities of NVIDIA GPUs. Visit the documentation for an explanation of compute capability and its relevance. Note that the Docker cache will save previously built binaries if the given configurations are identical.

2. Read the docker-compose.yaml file to fill in extra variables in .env. Also, feel free to edit docker-compose.yaml as necessary by changing session names, hostnames, etc. for different projects and configurations. The docker-compose.yaml file provides reasonable default values but these can be overridden by values specified in the .env file. An important configuration is ipc: host, which allows the container to access the shared memory of the host. This is required for multiprocessing, e.g., to use num_workers in the PyTorch DataLoader class. Disable this configuration on WSL and specify shm_size: instead as WSL cannot use host IPC as of the time of writing.

3. Edit requirements in reqs/apt-train.requirements.txt and reqs/train-environment.yaml. These contain project package dependencies. The apt requirements are designed to resemble an ordinary Python requirements.txt file.

4. Edit the volumes section of a service to include external directories in the container environment. Run make over to create a docker-compose.override.yaml file to add custom volumes and configurations. The docker-compose.override.yaml file is excluded from version control to allow per-user and per-server settings.

5. (Advanced) If an external file must be included in the Docker image build process, edit the .dockerignore file to allow the Docker context to find the external file. By default, all files except requirements files are excluded from the Docker build context.

Example .env file for user with username USERNAME, group name GROUPNAME, user id 1000, group id 1000 on service train. Use the simple service if no dependencies need to be compiled and requirements can either be downloaded or installed via apt, conda, or pip.

# Generated automatically by `make env`.
GID=1000
UID=1000
GRP=GROUPNAME
USR=USERNAME
PROJECT=train-username             # `PROJECT` must be in lowercase.
SERVICE=train
COMMAND=/bin/zsh                   # Command to execute on starting the container.
IMAGE_NAME=cresset:train-username  # `IMAGE_NAME` is also converted to lowercase.
PROJECT_ROOT=/opt/project

# [[Optional]]: Fill in these configurations manually if the defaults do not suffice.

# NVIDIA GPU Compute Capability (CCA) values may be found at https://developer.nvidia.com/cuda-gpus
CCA=8.6              # Compute capability. CCA=8.6 for RTX3090.
# CCA='8.6+PTX'      # The '+PTX' enables forward compatibility. Multiple CCAs can also be specified.
# CCA='7.5 8.6+PTX'  # Visit https://pytorch.org/docs/stable/cpp_extension.html for details.

# Used only if building PyTorch from source (`BUILD_MODE=include`).
# The `*_TAG` variables are used only if `BUILD_MODE=include`. No effect otherwise.
BUILD_MODE=exclude               # Whether to build PyTorch from source.
PYTORCH_VERSION_TAG=v2.0.0       # Any `git` tag can be used (but not just any commit hash).
TORCHVISION_VERSION_TAG=v0.15.1

# General environment configurations.
LINUX_DISTRO=ubuntu   # Visit the NVIDIA Docker Hub repo for available base images.
DISTRO_VERSION=22.04  # https://hub.docker.com/r/nvidia/cuda/tags
CUDA_VERSION=11.8.0   # Must be compatible with hardware and CUDA driver.
CUDNN_VERSION=8       # Only major version specifications are available.
PYTHON_VERSION=3.10   # Specify the Python version.
MKL_MODE=include      # Enable MKL for Intel CPUs.
TZ=Asia/Seoul         # Set the container timezone.

# Advanced Usage.
TARGET_STAGE=train    # Target Dockerfile stage. The `*.whl` files are available in `train-builds`.
ADD_USER=include      # Whether to create a new user (include) or use `root` user (exclude).

General Usage After Initial Installation and Configuration

1. Run make build to build the image from the Dockerfile and start the service. The make commands are defined in the Makefile and target the train service by default. Run make up if the image has already been built and rebuilding the image from the Dockerfile is not necessary.
2. Run make exec to enter the interactive container environment. Using tmux inside the container is recommended.
3. There is no step 3. Just start coding. Check out the documentation or create an issue if anything goes wrong.

Makefile Instructions

The Makefile contains shortcuts for common docker compose commands. Please read the Makefile to see the exact commands.

1. make build builds the Docker image from the Dockerfile regardless of whether the image already exists. This will reinstall packages to the updated requirements files, and then recreate the container.
2. make up creates a fresh container from the image, undoing any changes to the container made by the user. Allows changing container settings as network ports, mounted volumes, shared memory configurations, etc. Recommended method for using this project.
3. make exec enters the interactive terminal of the container created by make build or make up.
4. make down stops Compose containers and deletes networks. Necessary for service teardown.
5. make start restarts a stopped container without recreating it. Similar to make up but does not delete the current container. Not recommended unless data saved in container are absolutely necessary.
6. make ls shows all Docker Compose services, both active and inactive.
7. make run is used for debugging. Containers are removed on exit. If a service fails to start, use this to find the error.
8. make build-only builds the Docker image from the Dockerfile without starting the service. It exists to help publish images to container registries.

Tips

• The PROJECT, SERVICE, and COMMAND variables in the Makefile use variables specified in the .env file if available.
• If something does not work, first try make down to remove the current container and then make up to create a new container from the image. Explicitly tearing the container down is often necessary when something happens to the host.
• If the service startup stalls during make up, check docker system df to see if there is space left on the host machine.
• make up is akin to rebooting a computer. The current container is removed and a new container is created from the current image.
• make build is akin to resetting/formatting a computer. The current image, if present, is removed and a new image is built from the Dockerfile, after which a container is created from the resulting image. In contrast, make up only creates an image from source if the specified image is not present.
• make exec is akin to logging into a computer. It is the most important command and allows the user to access the container's terminal interactively.
• Configurations such as connected volumes and network ports cannot be changed in a running container, requiring a new container to be created.
• Docker automatically caches all builds up to defaultKeepStorage. Builds use caches from previous builds by default, greatly speeding up later builds by only building modified layers.
• If the build fails during git clone, try make build again with a stable internet connection.
• If the build fails during pip install, check the PyPI mirror URLs and package requirements.
• If any networking issues arise, check docker network ls and check for conflicts. Most networking and SSH problems can be solved by running docker network prune.

Project Overview

The main components of the project are as follows. The other files are utilities.

1. Dockerfile
2. docker-compose.yaml
3. docker-compose.override.yaml
4. reqs/(*requirements.txt|*environment.yaml)
5. .env

When the user inputs make up or another make command, commands specified in the Makefile are executed. The Makefile is used to specify shorthand commands and variables.

When a command related to Docker Compose (e.g., make build) is executed, The docker-compose.yaml file and the .env file are read by Docker Compose. The docker-compose.yaml file specifies reasonable default values but users may wish to change them as per their needs. The values specified in the .env file take precedence over the defaults specified in the docker-compose.yaml file. Environment variables specified in the shell take precedence over those in the .env file. The .env file is deliberately excluded from source control to allow different users and machines to use different configurations.

The docker-compose.yaml file manages configurations, builds, runs, etc. using the Dockerfile. Visit the Docker Compose Specification and Reference for details.

The docker-compose.override.yaml is read by the docker-compose.yaml file during the setup phase. Add configurations specific to each host that should not be shared via source control such as volume mounts for host-specific paths.

The Dockerfile is configured to read only requirements files in the reqs directory. Edit reqs/pip-train.requirements.txt to specify Python package requirements. Edit reqs/apt-train.requirements.txt to specify Ubuntu package requirements. Users must edit the .dockerignore file to COPY other files into the Docker build, for example, when building from private code during the Docker build.

The Dockerfile uses Docker BuildKit and a multi-stage build where control flow is specified via stage names and build-time environment variables given via docker-compose.yaml. See the Docker BuildKit Syntax for more information on Docker BuildKit. The train service specified in the docker-compose.yaml file uses the train stage specified in the Dockerfile, which assumes an Ubuntu image.

Raison d'tre

The purpose of this section is to introduce a new paradigm for deep learning development. The hope is that Cresset, or at least the ideas behind it, will eventually become best practice for small to medium-scale deep learning research and development.

Developing in local environments with conda or pip is commonplace in the deep learning community. However, this risks rendering the development environment, and the code meant to run on it, unreproducible. This state of affairs is a serious detriment to scientific progress that many readers of this article will have experienced at first-hand.

Docker containers are the standard method for providing reproducible programs across different computing environments. They create isolated environments where programs can run without interference from the host or from one another. For details, see the documentation.

But in practice, Docker containers are often misused. Containers are meant to be transient and best practice dictates that a new container be created for each run. However, this is very inconvenient for development, especially for deep learning applications, where new libraries must constantly be installed and bugs are often only evident at runtime. This leads many researchers to develop inside interactive containers. Docker users often have run.sh files with commands such as docker run -v my_data:/mnt/data -p 8080:22 -t my_container my_image:latest /bin/bash (look familiar, anyone?) and use SSH to connect to running containers. VSCode even provides a remote development mode to code inside containers.

The problem with this approach is that these interactive containers become just as unreproducible as local development environments. A running container cannot connect to a new port or attach a new volume. But if the computing environment within the container was created over several months of installs and builds, the only way to keep it is to save the container as an image and create a new container from the saved image. After a few iterations of this process, the resulting images become bloated and no less scrambled than the local environments that they were meant to replace.

Problems become even more evident when preparing for deployment. MLOps, defined as a set of practices that aims to deploy and maintain machine learning models reliably and efficiently, has gained enormous popularity of late as many practitioners have come to realize the importance of continuously maintaining ML systems long after the initial development phase ends.

However, bad practices such as those mentioned above mean that much coffee has been spilled turning research code into anything resembling a production-ready product. Often, even the original developers cannot recreate the same model after a few months. Many firms thus have entire teams dedicated to model translation, a huge expenditure.

To alleviate these problems, Docker Compose is proposed as a simple MLOps solution. Using Docker and Docker Compose, the entire training environment can be reproduced. Compose has not yet caught on in the deep learning community, possibly because it is usually advertised as a multi-container solution. This is a misunderstanding as it can be used for single-container development just as well.

A docker-compose.yaml file is provided for easy management of containers. Using the provided docker-compose.yaml file will create an interactive environment, providing a programming experience very similar to using a terminal on a remote server. Integrations with popular IDEs (PyCharm, VSCode) are also available.

Moreover, it also allows the user to specify settings for both build and run, removing the need to manage the environment with custom shell scripts. Connecting a new volume or port is as simple as removing the current container, adding a line in the docker-compose.yaml file, then running make up to create a new container from the same image.

Build caches allow new images to be built very quickly, removing another barrier to Docker adoption, the long initial build time. For more information on Compose, visit the documentation.

Docker Compose can also be used for deployment, which is useful for small to medium-sized deployments. If and when large-scale deployments using container orchestration such as Kubernetes becomes necessary, using reproducible Docker environments from the very beginning will accelerate the development process and smooth the path to MLOps adoption. Accelerating time-to-market by streamlining the development process is a competitive edge for any firm, whether lean startup or tech titan.

With luck, the techniques proposed here will enable the deep learning community to "write once, train anywhere". But even if most users are not persuaded of the merits of this method, Many a hapless grad student may be spared from the sisyphean labor of setting up their conda environment, only to have it crash and burn right before their paper submission is due.

Compose as Best Practice

Docker Compose is superior to using custom shell scripts for each environment. Not only does it gather all variables and commands for both build and run into a single file, but its native integration with Docker means that it makes complicated Docker build/run setups simple to implement and use.

Using Docker Compose this way is a general-purpose technique that does not depend on anything about this project. The other services available in the project emphasize this point.

Using Compose with PyCharm and VSCode

The Docker Compose container environment can be used with popular Python IDEs, not just in the terminal. PyCharm and Visual Studio Code, both very popular in the deep learning community, are compatible with Docker Compose.

PyCharm (Professional only)

Both Docker and Docker Compose are natively available as Python interpreters. See tutorials for Docker and Compose for details. JetBrains Gateway can also be used to connect to running containers.

N.B. PyCharm Professional and other JetBrains IDEs are available free of charge to anyone with a valid university e-mail address.

VSCode

Install the Remote Development extension pack. See tutorial for details.

VSCode Tips

VSCode may fail to start up when accessing remote containers created by Cresset because of the ${HOME}/.vscode-server volume mounted in the docker-compose.yaml file, which is used to preserve the .vscode-server directory between separate containers.

The reason for VSCode connection failure is that if any host directory specified as a volume does not exist, Docker will automatically create the specified host directory with the directory owner set to root. Directories that already exist retain their directory ownership. When the .vscode-server directory is created by Docker this way, VSCode is unable to install any files in the .vscode-server directory.

This has been fixed in the Makefile but problems related to the .vscode-server directory occur frequently. To solve this problem, simply change the directory ownership to the user with sudo chown -R $(id -u):$(id -g) ${HOME}/.vscode-server. This command can be run either on the host or inside the container, which is useful if sudo permissions are unavailable on the host.

Also, when one user switches between multiple Cresset-based containers on a single machine, VSCode may not be able to find the container workspace. This is because the docker-compose.yaml file mounts the host's ~/.vscode-server directory to the /home/${USR}/.vscode-server directory of all containers to preserve VSCode extensions between containers. To fix this issue, create a new directory on the host to mount the containers' .vscode-server directories. For example, one can set volume pairs as ${HOME}/.vscode-project1:/home/${USR}/.vscode-server for project1 and ${HOME}/.vscode-project2:/home/${USR}/.vscode-server for project2. Do not forget to create ${HOME}/.vscode-project1 and ${HOME}/.vscode-project2 on the host first. Otherwise, the directory will be owned by root, which will cause VSCode to stall indefinitely due to permission issues.

For other VSCode problems, try deleting ~/.vscode-server on the host.

Known Issues

1. Connecting to a running container by ssh will remove all variables set by ENV. This is because sshd starts a new environment, deleting all previous variables. Using docker/docker compose to enter containers is strongly recommended.

2. pip install package[option] will fail on the terminal because of Z-shell globbing. Characters such as [,],*, etc. will be interpreted by Z-shell as special commands. Use string literals, e.g., pip install 'package[option]', for cross-shell consistency.

3. If the build fails during git clone, simply try make build again. Most of the build will be cached. Failure is probably due to networking issues during installation. Updating git submodules is not fail-safe.

4. torch.cuda.is_available() will return a ... UserWarning: CUDA initialization:... error or the image will simply not start if the host CUDA driver is incompatible with the CUDA version on the Docker image. Either upgrade the host CUDA driver or downgrade the CUDA version of the image. Check the compatibility matrix to see if the host CUDA driver is compatible with the desired version of CUDA. Also, check if the CUDA driver has been configured correctly on the host. The CUDA driver version can be found using the nvidia-smi command.

5. Docker Compose V2 will silently fail if the installed Docker engine version is too low on Linux hosts. Update Docker to the latest version (23.0+) to use Docker Compose V2.

Desiderata

1. MORE STARS. No Contribution Without Appreciation!

2. A method of building Magma from source would be appreciated. Currently, Cresset depends on the magma-cudaXXX package provided in the PyTorch channel of Anaconda.

3. Bug reports are welcome. Only the latest versions have been tested rigorously. Please raise an issue if there are any versions that do not build properly. However, please check that your host Docker, Docker Compose, and especially NVIDIA Driver are up-to-date before doing so.

4. Translations into other languages and updates to existing translations are welcome. Please create a separate LANG.README.md file and make a pull request.