Better model security with new PyTorch version

[mayankjobanputra]

When loading PyTorch models from the modelhub, arbitrary code can be executed. See here.

Seems like PyTorch already solved this:

• Issue on the PyTorch Repo: pickle is a security issue pytorch/pytorch#52596
• By default, this is not available in torch==1.13. TORCH_FORCE_WEIGHTS_ONLY_LOAD needs to be set to True which allows global override to safe only model loading via.

Test
Test haystack tests with new flags

Solution:

• Bump up PyTorch version to 1.13

Behaviour

• Secure by default TORCH_FORCE_WEIGHTS_ONLY_LOAD always set to true when Haystack loads any models

[mayankjobanputra]

Something that I have been struggling for a couple of days: python-poetry/poetry#6409

[LLukas22]

@mayankjobanputra I also tried to get this working. This seams to be a bug on pytorchs side as mentioned in pytorch/pytorch#89924. It will probably get fixes in the 1.13.1 release which will be released in few days. I'll take another look when they fixed the marker problem in pytorch.

[mayankjobanputra]

@LLukas22 thanks a lot :)

I am also tracking this PyTorch issue and looking forward to upgrading PyTorch version in Haystack!

[mayankjobanputra]

pytorch/pytorch#52181 (comment)

Discussion is still going on about enabling TORCH_FORCE_WEIGHTS_ONLY_LOAD by default.