light-the-torch

• add support for pip==23.2 (#139)

A lot of Nvidia packages as well as fsspec made it onto the PyTorch indices. light-the-torch now handles them correctly (#137)

• add support for pip==23.1 (#131)

• fix local installations and add CI job (#124)

• Add support for pip==23.0 (#122)

• Add support for CUDA 12 @jamesobutler (#110)
• Fix minimum required driver versions for CUDA 11 (#111)
• Mitigate risk of supply chain attacks (#115, #118)
• Drop support for PyTorch LTS (#116)

• Add support for CUDA 11.7 and 11.8 by @lassoan (#108)
• Add support for pip==22.3.1 (#105)

This release is needed since pip==22.3 breaks previous light-the-torch versions. Compatibility was fixed in #99.

Apart from that, this release brings a few other quality-of-life improvements:

• use PyPI as fallback if stable binaries are not on PyTorch indices (#93)
• improve CLI help messages (#94)
• add option to specify computation backend through environment variable (#102)

• use PyPI as fallback if stable binaries are not on PyTorch indices (#93)
• improve CLI help messages (#94)
• Update to pip==22.3 by (#99)
• add option to specify computation backend through environment variable by (#102)

If you want to install this version of light-the-torch, use pip install --pre light-the-torch

• Expand test matrix to also install different computation backends #80

• exclude torchrec from patched distributions #85
• add support for pip==22.2 #89

As noted in #40, the supported pip versions before this release were quite outdated. A major blocker was the dependency resolver that pip==21 introduced. It changed a lot of the internals that we relied on. Thus, we needed to completely overhaul the project.

And we did in #60 🎉 As a result, we now support every pip version starting from 22 with a few exceptions. Of course, given that we need to rely on pip internals without any BC guarantees, there will be some breakages in future version. But we are confident, that the changes that needed to happen in light-the-torch should be small.

In the process of rewriting, we also added support for other PyTorch distributions to be installed through light-the-torch. Before it could only handle

• torch
• torchvision
• torchtext
• torchaudio

but now in addition we support:

• torch
• torch_model_archiver
• torch_tb_profiler
• torcharrow
• torchcsprng
• torchdata
• torchdistx
• torchserve

Unfortunately, overhauling the project also brings some BC breaking changes. Namely, ltt extract and ltt find and their API counterparts light_the_torch.extract_dists and light_the_torch.find_links were removed. This processing in stages, i.e. extracting the PyTorch distributions from the installation command, finding the correct link for them and only then installing them in a subsequent command is no longer possible with the new pip dependency resolver.

Shout out to our contributors: @pmeier @tonyfast @jamesobutler

Adds the following PyTorch distributions to be patched by light-the-torch:

• torch_model_archiver
• torch_tb_profiler
• torcharrow
• torchcsprng
• torchdata
• torchdistx
• torchserve

• add naive torch install test (#58)
• use extra index rathe than find links for link patching (#59)

If you want to install this version of light-the-torch, use pip install --pre light-the-torch.

• relax ROCm deselection even further (#57)

If you want to install this version of light-the-torch, use pip install --pre light-the-torch.

• fix ROCm deselection (#56)

If you want to install this version of light-the-torch, use pip install --pre light-the-torch.

• disable ROCm wheels (#54)

If you want to install this version of light-the-torch, use pip install --pre light-the-torch.

• fix candidate selection #52
• fix module entrypoint #50
• rewrite README #45 (@jamesobutler)

If you want to install this version of light-the-torch, use pip install --pre light-the-torch.

Cleans up the code and fixes some manual found bugs.

If you want to install this version of light-the-torch, use pip install --pre light-the-torch.

This is the first step towards supporting newer versions of pip in the backend. The major change is that instead of using pip internals, light-the-torch now actually wraps pip.

If you want to install this version of light-the-torch, use pip install --pre light-the-torch.