ltt
light-the-torch
pypi i light-the-torch
ltt

light-the-torch

Install PyTorch distributions with computation backend auto-detection

by Philip Meier

0.6.0 (see all)License:BSD-3-Clause
pypi i light-the-torch
Readme

light-the-torch

BSD-3-Clause License Project Status: WIP Code coverage via codecov.io

light-the-torch is a small utility that wraps pip to ease the installation process for PyTorch distributions like torch, torchvision, torchaudio, and so on as well as third-party packages that depend on them. It auto-detects compatible CUDA versions from the local setup and installs the correct PyTorch binaries without user interference.

Why do I need it?

PyTorch distributions like torch, torchvision, torchaudio, and so on are fully pip install'able, but PyPI, the default pip search index, has some limitations:

  1. PyPI regularly only allows binaries up to a size of approximately 60 MB. One can request a file size limit increase (and the PyTorch team probably does that for every release), but it is still not enough: although PyTorch has pre-built binaries for Windows with CUDA, they cannot be installed through PyPI due to their size.
  2. PyTorch uses local version specifiers to indicate for which computation backend the binary was compiled, for example torch==1.11.0+cpu. Unfortunately, local specifiers are not allowed on PyPI. Thus, only the binaries compiled with one CUDA version are uploaded without an indication of the CUDA version. If you do not have a CUDA capable GPU, downloading this is only a waste of bandwidth and disk capacity. If on the other hand your NVIDIA driver version simply doesn't support the CUDA version the binary was compiled with, you can't use any of the GPU features.

To overcome this, PyTorch also hosts most1 binaries on their own package indices. To access PyTorch's package indices, you can still use pip install, but some additional options are needed:

pip install torch --extra-index-url https://download.pytorch.org/whl/cu113

While this is certainly an improvement, it still has a few downsides:

  1. You need to know what computation backend, e.g. CUDA 11.3 (cu113), is supported on your local machine. This can be quite challenging for new users and at least tedious for more experienced ones.
  2. Besides the stable binaries, PyTorch also offers nightly, test, and long-time support (LTS) ones. To install them, you need a different --extra-index-url for each.
  3. For the nightly and test channel you also need to supply the --pre option. Failing to do so, will pull the stable binary from PyPI even if the rest of the installation command is correct.
  4. When installing from the LTS channel, you need to pin the exact version, since pip prefers newer releases from PyPI. Thus, it is not possible to automatically get the latest LTS release.

In case you only want to install PyTorch distributions, point 3. and 4. above can be resolved by using --index-url instead and completely disabling installing from PyPI. But of course this means it is not possible to install any package not hosted by PyTorch, but that depends on it.

If any of these points don't sound appealing to you, and you just want to have the same user experience as pip install for PyTorch distributions, light-the-torch was made for you.

How do I install it?

Installing light-the-torch is as easy as

pip install light-the-torch

Since it depends on pip and it might be upgraded during installation, Windows users should install it with

py -m pip install light-the-torch

How do I use it?

After light-the-torch is installed you can use its CLI interface ltt as drop-in replacement for pip:

ltt install torch

In fact, ltt is pip with a few added options:

  • By default, ltt uses the local NVIDIA driver version to select the correct binary for you. You can pass the --pytorch-computation-backend option to manually specify the computation backend you want to use:

    ltt install --pytorch-computation-backend=cu102 torch torchvision torchaudio
    

    Borrowing from the mutex packages that PyTorch provides for conda installations, --cpuonly is available as shorthand for --pytorch-computation-backend=cu102.

    In addition, the computation backend to be installed can also be set through the LTT_PYTORCH_COMPUTATION_BACKEND environment variable. It will only be honored in case no CLI option for the computation backend is specified.

  • By default, ltt installs stable PyTorch binaries. To install binaries from the nightly, test, or LTS channels pass the --pytorch-channel option:

    ltt install --pytorch-channel=nightly torch torchvision torchaudio
    

    If --pytorch-channel is not passed, using pip's builtin --pre option will install PyTorch test binaries.

Of course, you are not limited to install only PyTorch distributions. Everything shown above also works if you install packages that depend on PyTorch:

ltt install --pytorch-computation-backend=cpu --pytorch-channel=nightly pystiche

How does it work?

The authors of pip do not condone the use of pip internals as they might break without warning. As a results of this, pip has no capability for plugins to hook into specific tasks.

light-the-torch works by monkey-patching pip internals at runtime:

  • While searching for a download link for a PyTorch distribution, light-the-torch replaces the default search index with an official PyTorch download link. This is equivalent to calling pip install with the --extra-index-url option only for PyTorch distributions.
  • While evaluating possible PyTorch installation candidates, light-the-torch culls binaries incompatible with the hardware.

How do I contribute?

Thanks a lot for your interest to contribute to light-the-torch! All contributions are appreciated, be it code or not. Especially in a project like this, we rely on user reports for edge cases we didn't anticipate. Please feel free to open an issue if you encounter anything that you think should be working but doesn't.

If you want to contribute code, check out our contributing guidelines to learn more about the workflow.


  1. Some distributions are not compiled against a specific computation backend and thus hosting them on PyPI is sufficient since they work in every environment.
VersionTagPublished
0.6.0
1mo ago
0.5.0
3mos ago
0.5.0rc0
3mos ago
0.4.2
6mos ago
No alternatives found
No tutorials found
Add a tutorial
No dependencies found

Rate & Review

100
No reviews found
Be the first to rate