Conversation
|
Hmm I'm not sure why the docs build is failing. Otherwise lgtm |
|
Actually, even after unpin, the CI seems to be running with 0.2.19 because that's the latest version available in Py3.8, which is what the CI uses. So, the CI passing here is meaningless. I'm getting some weird cc @Balandat |
|
Removing |
|
Ok, CI is also happy (except for docs, which seems unrelated) and running with latest jaxtyping now |
| pip install torch==1.11+cpu -f https://download.pytorch.org/whl/torch_stable.html | ||
| pip install "numpy<2" |
There was a problem hiding this comment.
why is numpy < 2 needed here?
There was a problem hiding this comment.
torch=1.11 is not compatible with numpy>2 but it does not explicitly require <2 either, so we end up with CI failing with RuntimeError: Numpy is not available on tensor.numpy() calls despite numpy==2.2.2 being installed. See https://github.com/cornellius-gp/linear_operator/actions/runs/12961267021/job/36156321493
There was a problem hiding this comment.
I see. torch 1.11 is rather ancient but I guess this is only relevant for the CI testing, so this is fine by me.
3 participants
The required jaxtyping version conflicts with the requirements of other libraries. The code appears to be compatible with newer versions, so attempting to unpin to see if anything breaks.