The latest version released on the official website is 0.22.1 instead of 0.22.0?

[jestiny0]

I previously raised an issue about upgrading DJL to the latest version and upgrading PyTorch to 2.0.0. At that time, the version I received from the Maven repository was 0.22.0, and I also tested and upgraded all my production services to 0.22.0. However, I found that the latest version released on the DJL official website is 0.22.1 instead of 0.22.0.

My key questions are:

1. What are the key changes made in 0.22.1 compared to 0.22.0?
2. Is it necessary for me to upgrade to 0.22.1?

[frankfliu]

0.22.0 default use PyTorch 2.0.0, which we found quite a few issues. in 0.22.1 there are two critical updates:

1. downgrade default PyTorch version to 1.13.1 in DJL's BOM
2. Fixed bug in NeuronUtils which doesn't detect neuron core correctly.

If you are not using PyTorch 2.0.0, you should be fine.

[jestiny0]

0.22.0 default use PyTorch 2.0.0, which we found quite a few issues

What specific issues are there? We will evaluate and test whether it will affect our use case.

We upgraded to DJL 0.22.0-2.0.0 solely for the purpose of preparing for training models with PyTorch 2.0.0. We are currently planning to upgrade our training PyTorch version from 1.12.1 to 2.0.0. Are there any concerns with this upgrade?

Or do you think we should hold off on upgrading the PyTorch version for model training pipeline until a more stable release version of DJL is available in the future?

[frankfliu]

We observed cuda error when running multi-threading inference on GPU: https://github.com/pytorch/pytorch/issues/99372

If you are using mpi in training/inference in python, it will crash as well.

PyTorch 2.0.1 is right at the corner, we will test 2.0.1 and see if it's stable.

[jestiny0]

@frankfliu I saw that pytorch 2.0.1 has been officially released. Do you have a specific plan for when to release DJL that supports version 2.0.1?