Mario Lezcano Casado

Can you open a new issue with a self-contained repro (if possible, that doesn't use external libraries) to track this one?

I see there is a PR that was open but never landed that would fix this https://github.com/pytorch/pytorch/pull/108206. @mikaylagawarecki would you mind resubmitting that PR with the changes you had in...

It still does not render well in my browser:  The difference I see is that now the background is in a slightly darker grey, and the formulae in double...

Yes, sorry, I wrote that in both the issues. All I meant to write here is that the spacing between the end of `.. warning::` and the end of `.....

Oh, In my head correct would be to leave an space the size of that just before `..code::` so that it does not look crammed. this space would have the...

cc @ThomasRaoux @Jokeren for visibility.

IIRC, PyTorch does implement a stable algorithm as well, but I don't know how fine does this algorithm need to be. cc @peterbell10 who I think implemented this algorithm. What...

As for the particular case... I don't think this should be tested. If you are making a case for testing this, you'll be able to make similar cases for all...

The roll error is fine, but that pdist error I'll have to look into. I'll spin up an A100 box next week and make sure the fix I submit works...

I thought the issue was coming from the use of index_select and its decomposition dispatching to advanced indexing, which sometime synchronizes