torch-max-mem

This package provides decorators for memory utilization maximization with PyTorch and CUDA by starting with a maximum parameter size and applying successive halving until no more out-of-memory exception occurs.

💪 Getting Started

Assume you have a function for batched computation of nearest neighbors using brute-force distance calculation.

import torch

def knn(x, y, batch_size, k: int = 3):
    return torch.cat(
        [
            torch.cdist(x[start : start + batch_size], y).topk(k=k, dim=1, largest=False).indices
            for start in range(0, x.shape[0], batch_size)
        ],
        dim=0,
    )

With torch_max_mem you can decorate this function to reduce the batch size until no more out-of-memory error occurs.

import torch
from torch_max_mem import maximize_memory_utilization


@maximize_memory_utilization()
def knn(x, y, batch_size, k: int = 3):
    return torch.cat(
        [
            torch.cdist(x[start : start + batch_size], y).topk(k=k, dim=1, largest=False).indices
            for start in range(0, x.shape[0], batch_size)
        ],
        dim=0,
    )

In the code, you can now always pass the largest sensible batch size, e.g.,

x = torch.rand(100, 100, device="cuda")
y = torch.rand(200, 100, device="cuda")
knn(x, y, batch_size=x.shape[0])

🚀 Installation

The most recent release can be installed from PyPI with uv:

uv pip install torch_max_mem

or with pip:

python3 -m pip install torch_max_mem

The most recent code and data can be installed directly from GitHub with uv:

uv pip install git+https://github.com/mberr/torch-max-mem.git

or with pip:

python3 -m pip install git+https://github.com/mberr/torch-max-mem.git

👐 Contributing

Contributions, whether filing an issue, making a pull request, or forking, are appreciated. See CONTRIBUTING.md for more information on getting involved.

👋 Attribution

Parts of the logic have been developed with Laurent Vermue for PyKEEN.

⚖️ License

The code in this package is licensed under the MIT License.

🍪 Cookiecutter

This package was created with @audreyfeldroy's cookiecutter package using @cthoyt's cookiecutter-snekpack template.

🛠️ For Developers

See developer instructions

The final section of the README is for if you want to get involved by making a code contribution.

Development Installation

To install in development mode, use the following:

git clone git+https://github.com/mberr/torch-max-mem.git
cd snekpack-demo
uv pip install -e .

Alternatively, install using pip:

python3 -m pip install -e .

Updating Package Boilerplate

This project uses cruft to keep boilerplate (i.e., configuration, contribution guidelines, documentation configuration) up-to-date with the upstream cookiecutter package. Install cruft with either uv tool install cruft or python3 -m pip install cruft then run:

cruft update

More info on Cruft's update command is available here.

🥼 Testing

After cloning the repository and installing tox with uv tool install tox --with tox-uv or python3 -m pip install tox tox-uv, the unit tests in the tests/ folder can be run reproducibly with:

tox -e py

Additionally, these tests are automatically re-run with each commit in a GitHub Action.

📖 Building the Documentation

The documentation can be built locally using the following:

git clone git+https://github.com/mberr/torch-max-mem.git
cd snekpack-demo
tox -e docs
open docs/build/html/index.html

The documentation automatically installs the package as well as the docs extra specified in the pyproject.toml. sphinx plugins like texext can be added there. Additionally, they need to be added to the extensions list in docs/source/conf.py.

The documentation can be deployed to ReadTheDocs using this guide. The .readthedocs.yml YAML file contains all the configuration you'll need. You can also set up continuous integration on GitHub to check not only that Sphinx can build the documentation in an isolated environment (i.e., with tox -e docs-test) but also that ReadTheDocs can build it too.

Configuring ReadTheDocs

1. Log in to ReadTheDocs with your GitHub account to install the integration at https://readthedocs.org/accounts/login/?next=/dashboard/
2. Import your project by navigating to https://readthedocs.org/dashboard/import then clicking the plus icon next to your repository
3. You can rename the repository on the next screen using a more stylized name (i.e., with spaces and capital letters)
4. Click next, and you're good to go!

📦 Making a Release

Configuring Zenodo

Zenodo is a long-term archival system that assigns a DOI to each release of your package.

1. Log in to Zenodo via GitHub with this link: https://zenodo.org/oauth/login/github/?next=%2F. This brings you to a page that lists all of your organizations and asks you to approve installing the Zenodo app on GitHub. Click "grant" next to any organizations you want to enable the integration for, then click the big green "approve" button. This step only needs to be done once.
2. Navigate to https://zenodo.org/account/settings/github/, which lists all of your GitHub repositories (both in your username and any organizations you enabled). Click the on/off toggle for any relevant repositories. When you make a new repository, you'll have to come back to this

After these steps, you're ready to go! After you make "release" on GitHub (steps for this are below), you can navigate to https://zenodo.org/account/settings/github/repository/mberr/torch-max-mem to see the DOI for the release and link to the Zenodo record for it.

Registering with the Python Package Index (PyPI)

You only have to do the following steps once.

1. Register for an account on the Python Package Index (PyPI)
2. Navigate to https://pypi.org/manage/account and make sure you have verified your email address. A verification email might not have been sent by default, so you might have to click the "options" dropdown next to your address to get to the "re-send verification email" button
3. 2-Factor authentication is required for PyPI since the end of 2023 (see this blog post from PyPI). This means you have to first issue account recovery codes, then set up 2-factor authentication
4. Issue an API token from https://pypi.org/manage/account/token

Configuring your machine's connection to PyPI

You have to do the following steps once per machine.

uv tool install keyring
keyring set https://upload.pypi.org/legacy/ __token__
keyring set https://test.pypi.org/legacy/ __token__

Note that this deprecates previous workflows using .pypirc.

Uploading to PyPI

After installing the package in development mode and installing tox with uv tool install tox --with tox-uv or python3 -m pip install tox tox-uv, run the following from the console:

tox -e finish

This script does the following:

1. Uses bump-my-version to switch the version number in the pyproject.toml, CITATION.cff, src/torch_max_mem/version.py, and docs/source/conf.py to not have the -dev suffix
2. Packages the code in both a tar archive and a wheel using uv build
3. Uploads to PyPI using uv publish.
4. Push to GitHub. You'll need to make a release going with the commit where the version was bumped.
5. Bump the version to the next patch. If you made big changes and want to bump the version by minor, you can use tox -e bumpversion -- minor after.

Releasing on GitHub

1. Navigate to https://github.com/mberr/torch-max-mem/releases/new to draft a new release
2. Click the "Choose a Tag" dropdown and select the tag corresponding to the release you just made
3. Click the "Generate Release Notes" button to get a quick outline of recent changes. Modify the title and description as you see fit
4. Click the big green "Publish Release" button

This will trigger Zenodo to assign a DOI to your release as well.