add support for torch 2.0?

[yahui624]

hi, right now it has a limitation of satisfies the requirement torch<1.13.0,>=1.9.0 (from backpack-for-pytorch). Will you be adding support for torch2.0 soon?

[f-dangel]

Hi,

thanks for your question. To the best of my knowledge, torch 2.0 is fully backward-compatible with the latest 1.xx, but introduces a new compile function. So in principle, BackPACK should support 2.0. The reason we do not officially support 2.0 is that we have not tested how compile interferes with the forward and full backward hooks that BackPACK uses under the hood.

If you run into issues while using BackPACK with torch 2.0, please let us know.

Best, Felix

[yahui624]

Hi, i m experimenting my model on nvidia h100/sm 90, which requires pytorch to build from source. inside that conda environment, i do pip install pip install backpack-for-torch, it will auto degrade my torch version to 1.13, how can i stop that from happening?

[f-dangel]

Hi,

you can pip install backpack-for-pytorch first, then pip uninstall torch and install PyTorch by building locally.

[yahui624]

awesome thanks!

[hlzl]

@f-dangel I can confirm from my testing that it works with torch 2.0.1. However, using pip to install backpack now automatically downgrades your torch version to 1.13.0, which is kind of annoying. It would thus be great if you could change the requirement to torch>=1.9.0 or torch<2.1.0,>=1.9.0 as an intermediate solution.

[f-dangel]

Hi @hlzl,

if you install the latest version of BackPACK (1.6.0) it should work with torch>=2

Best, Felix

[hlzl]

Ah sorry, seems like I downloaded it just before the bump to 1.6.0. Thanks for letting me know!