[leburgel]

TensorKit.jl v0.17.3

A patch release adding VectorInterface v0.7 compat, batched SVD, Enzyme rules for all factorizations, and much faster GPU index manipulations.

Highlights

• Compat: VectorInterface v0.6 and v0.7; TensorKitSectors v0.3.10 with NamedSector/@NamedSector re-exported (#560, #566)
• Batched SVD: svd_full, svd_compact and svd_vals use MatrixAlgebraKit's batched drivers where supported, e.g. on the GPU (#567)
• GPU performance: add_transform! rewritten as KernelAbstractions kernels parallelizing over the output data, orders of magnitude faster for symmetric tensors; norm(t, Inf) now works on GPU (#533, #509, #586)
• Enzyme AD: forward and reverse rules for qr, lq, eig, eigh, svd, polar and project_(anti)hermitian; forward rules for permute/transpose/braid (#464, #573)
• CPU performance: cheaper FusionTreeBlock construction, BLAS-backed recoupling matrices, storagetype-specific TreeTransformer caches, and no copy of C in ChainRules rules when β = Zero() (#581, #556, #555)
• Fixes: promote_storagetype in the planar preprocessor, the twist Enzyme rule on Julia 1.10, and untested error paths found by JET (#577, #559, #570)

Full Changelog

See CHANGELOG for the complete list of changes.

[kshyatt]

Can we get #602 in as well?

[leburgel]

It would be great to have a patch release with support for VectorInterface.jl v0.7, to unblock Julia 1.13 users.

[kshyatt]

KK, we can always tag 0.17.4 soon, I guess

[leburgel]

It would be great to have a patch release with support for VectorInterface.jl v0.7, to unblock Julia 1.13 users.

I actually started typing this before I saw your comment. It's the same for me, but sooner rather than later would be very helpful for @VictorVanthilt.

[lkdvos]

It actually looks from the compat check that we might require the truncationstrategy PR to avoid ambiguity errors, so I guess that settles that question 🙃