It would be awesome if there were batched versions of potrf, potrs, potri, and qr. The batch Cholesky functions would be useful in MultivariateNormal, which currently uses a for loop to compute a batch of Cholesky decompositions.

It looks like Magma has batched versions of these functions. Lapack does not, but the Lapack version of this function could always use a for-loop internally.

Open to a PR?
/cc @vishwakftw @ssnl @jacobrgardner

Edit: and batched matrix inverse
Edit [vishwakftw]: Added a checklist

• solve a.k.a. gesv (Add batched linear solver to torch.gesv() #6100)
• inverse ([RFC, ready] Batched Inverse #9949)
• cholesky_solve a.k.a potrs ([Ready] Make potrs batched #13453)
• cholesky a.k.a potrf ([Ready] Batched cholesky decomposition #14017)
• lu a.k.a btrifact ([ready for review] torch.btrifact for tensors with greater than 3 dimensions #14964)
• triangular_solve a.k.a trtrs (Add batched version of trtrs #18025)
• qr (Enable batched QR decomposition and add a some option #20689)
• lstsq a.k.a gels
• lu_solve a.k.a btrisolve (Enable broadcasting of batch dimensions RHS and LHS tensors for lu_solve #24333)
• svd (Port SVD to ATen, enable batching for matrix inputs #21588)
• cholesky_inverse a.k.a potri (Port cholesky_inverse from TH/THC to ATen #31275)
• det / logdet / slogdet (Allowing batching for det/logdet/slogdet operations #22909)
• symeig (Port symeig to ATen and enable batching of inputs #21858)
• pinverse (Enable batching for pinverse #26095)