@ethanluoyc thanks for the contribution! I can review this after you're done with it; feel free to ping with any questions;

@zou3519 one very general question though.
Ideally, there will be a bunch of batched operators. (batched gesv, cholesky etc).
How do you suggest we encourage reuse the code for batching when we implement these operators?

I implemented batch gesv a while ago as a part of torch.gesv: #6100

For API cleanliness, it would be nice for the batched operations and unbatched operations to be a part of the same function, like torch.gesv.

What is the status of this PR?

@zou3519 I am having some issue with passing the tests. What I currently have is a simple wrapper around dpotrf_. I can't find out why this is giving me the wrong result. Would you be able to take a breif look and see if there are obvious mistakes?

Hi.
I accidentally wrote out a batch potrf without being aware of this PR.
On the up side, I think I have a working batch potrf forward based on top of @vishwakftw batched inverse #9949 , and also batched tril and triu to go with it (which I think you would have hit soonish). I am currently testing the backward (mainly working around no batched vector->diagonal matrix function and batchifying ops).

My apologies for not checking before going this far.

This can be closed since batched cholesky is merged into master.