📦 agi-pack

A Dockerfile builder for Machine Learning developers.

📦 agi-pack allows you to define your Dockerfiles using a simple YAML format, and then generate images from them trivially using Jinja2 templates and Pydantic-based validation. It's a simple tool that aims to simplify the process of building Docker images for machine learning (ML).

Goals 🎯

• 😇 Simplicity: Make it easy to define and build docker images for ML.
• 📦 Best-practices: Bring best-practices to building docker images for ML -- good base images, multi-stage builds, minimal image sizes, etc.
• ⚡️ Fast: Make it lightning-fast to build and re-build docker images with out-of-the-box caching for apt, conda and pip packages.
• 🧩 Modular, Re-usable, Composable: Define base, dev and prod targets with multi-stage builds, and re-use them wherever possible.
• 👩‍💻 Extensible: Make the YAML / DSL easily hackable and extensible to support the ML ecosystem, as more libraries, drivers, HW vendors, come into the market.
• ☁️ Vendor-agnostic: agi-pack is not intended to be built for any specific vendor -- I need this tool for internal purposes, but I decided to build it in the open and keep it simple.

Installation 📦

pip install agi-pack

For shell completion, you can install them via:

agi-pack --install-completion <bash|zsh|fish|powershell|pwsh>

Go through the examples and the corresponding examples/generated directory to see a few examples of what agi-pack can do. If you're interested in checking out a CUDA / CUDNN example, check out examples/agibuild.base-cu118.yaml.

Quickstart 🛠

1. Create a simple YAML configuration file called agibuild.yaml. You can use agi-pack init to generate a sample configuration file.

   agi-pack init

2. Edit agibuild.yaml to define your custom system and python packages.

   images:
     sklearn-base:
       base: debian:buster-slim
       system:
       - wget
       - build-essential
       python: "3.8.10"
       pip:
       - loguru
       - typer
       - scikit-learn

   Let's break this down:

   • sklearn-base: name of the target you want to build. Usually, these could be variants like *-base, *-dev, *-prod, *-test etc.
   • base: base image to build from.
   • system: system packages to install via apt-get install.
   • python: specific python version to install via miniconda.
   • pip: python packages to install via pip install.

3. Generate the Dockerfile using agi-pack generate

   agi-pack generate -c agibuild.yaml

   You should see the following output:

   $ agi-pack generate -c agibuild.yaml
   📦 sklearn-base
   └── 🎉 Successfully generated Dockerfile (target=sklearn-base, filename=Dockerfile).
       └── `docker build -f Dockerfile --target sklearn-base .`

That's it! Here's the generated Dockerfile -- use it to run docker build and build the image directly.

Rationale 🤔

Docker has become the standard for building and managing isolated environments for ML. However, any one who has gone down this rabbit-hole knows how broken ML development is, especially when you need to experiment and re-configure your environments constantly. Production is another nightmare -- large docker images (10GB+), bloated docker images with model weights that are ~5-10GB in size, 10+ minute long docker build times, sloppy package management to name just a few.

What makes Dockerfiles painful? If you've ever tried to roll your own Dockerfiles with all the best-practices while fully understanding their internals, you'll still find yourself building, and re-building, and re-building these images across a whole host of use-cases. Having to build Dockerfile(s) for dev, prod, and test all turn out to be a nightmare when you add the complexity of hardware targets (CPUs, GPUs, TPUs etc), drivers, python, virtual environments, build and runtime dependencies.

agi-pack aims to simplify this by allowing developers to define Dockerfiles in a concise YAML format and then generate them based on your environment needs (i.e. python version, system packages, conda/pip dependencies, GPU drivers etc).

For example, you should be able to easily configure your dev environment for local development, and have a separate prod environment where you'll only need the runtime dependencies avoiding any bloat.

agi-pack hopes to also standardize the base images, so that we can really build on top of giants.

More Complex Example 📚

Now imagine you want to build a more complex image that has multiple stages, and you want to build a base image that has all the basic dependencies, and a dev image that has additional build-time dependencies.

images:
  base-cpu:
    name: agi
    base: debian:buster-slim
    system:
        - wget
    python: "3.8.10"
    pip:
        - scikit-learn
    run:
        - echo "Hello, world!"

  dev-cpu:
    base: base-cpu
    system:
    - build-essential

Once you've defined this agibuild.yaml, running agi-pack generate will generate the following output:

$ agi-pack generate -c agibuild.yaml
📦 base-cpu
└── 🎉 Successfully generated Dockerfile (target=base-cpu, filename=Dockerfile).
    └── `docker build -f Dockerfile --target base-cpu .`
📦 dev-cpu
└── 🎉 Successfully generated Dockerfile (target=dev-cpu, filename=Dockerfile).
    └── `docker build -f Dockerfile --target dev-cpu .`

As you can see, agi-pack will generate a single Dockerfile for each of the targets defined in the YAML file. You can then build the individual images from the same Dockerfile using docker targets: docker build -f Dockerfile --target <target> . where <target> is the name of the image target you want to build.

Here's the corresponding Dockerfile that was generated.

Why the name? 🤷‍♂️

agi-pack is very much intended to be tongue-in-cheek -- we are soon going to be living in a world full of quasi-AGI agents orchestrated via ML containers. At the very least, agi-pack should provide the building blocks for us to build a more modular, re-usable, and distribution-friendly container format for "AGI".

Inspiration and Attribution 🌟

TL;DR agi-pack was inspired by a combination of Replicate's cog, Baseten's truss, skaffold, and Docker Compose Services. I wanted a standalone project without any added cruft/dependencies of vendors and services.

📦 agi-pack is simply a weekend project I hacked together, that started with a conversation with ChatGPT / GPT-4.

ChatGPT Prompt

---

Prompt: I'm building a Dockerfile generator and builder to simplify machine learning infrastructure. I'd like for the Dockerfile to be dynamically generated (using Jinja templates) with the following parametrizations:

# Sample YAML file
images:
  base-gpu:
    base: nvidia/cuda:11.8.0-base-ubuntu22.04
    system:
    - gnupg2
    - build-essential
    - git
    python: "3.8.10"
    pip:
    - torch==2.0.1

I'd like for this yaml file to generate a Dockerfile via agi-pack generate -c <name>.yaml. You are an expert in Docker and Python programming, how would I implement this builder in Python. Use Jinja2 templating and miniconda python environments wherever possible. I'd like an elegant and concise implementation that I can share on PyPI.

Contributing 🤝

Contributions are welcome! Please read the CONTRIBUTING guide for more information.

License 📄

This project is licensed under the MIT License. See the LICENSE file for details.