[ANN] TensorKit.jl v0.13

Package Announcements
, , , ,
1

Hi everyone,

Together with @juthohaegeman, I’m excited to announce the release of TensorKit.jl v0.13, a Julia package designed for large-scale tensor computations with a touch of category theory. This release brings a number of performance enhancements and simplifications to the data structures, particularly for symmetric tensors involving complex fusion trees. We expect a significant reduction in overhead in such cases.

Breaking Changes:

  • Scalar Type Parameter: The scalar type (eltype) is now an explicit parameter in the AbstractTensorMap type, and appears in the first position. Consequently, you should now use TensorMap{T}(undef, codomain ← domain) to create a TensorMap with uninitialized data of scalar type T.

  • Updated Tensor Constructors: The previous form TensorMap(randn, T, codomain ← domain) is being replaced by randn(T, codomain ← domain). We now overload rand, randn, and randexp from the Random module, mimicking the constructors for Array. You can also pass a custom random number generator as the first argument. Note that T (the scalar type) is optional and defaults to Float64. The old constructors are still available but in deprecation mode and will be removed in the 1.0 release.

  • TensorMap Structure: The TensorMap data structure has been simplified, with all tensor data now stored in a single <:DenseVector. While this does not change the interface, TensorMap objects saved with TensorKit.jl v0.12.7 or earlier (e.g., using JLD2.jl) are not compatible with this new version. We provide a script to help migrate data for v0.13 compatibility.

Major Non-Breaking Changes:

  • Support for TensorOperations.jl v5: TensorKit.jl now supports TensorOperations.jl v5, which introduces new backend and allocator support in the @tensor macro. This includes using Bumper.jl for managing intermediate tensors.

  • Separation of Sector Types: The Sector type hierarchy and its corresponding interface have been moved to TensorKitSectors.jl. This package is now a direct dependency of TensorKit.jl and is re-exported to ensure backward compatibility while allowing other packages to use it as a lightweight dependency.

  • fusiontrees Iterator Update: The fusiontrees iterator now processes FusionTree objects in a different order, improving the performance of certain future operations. It also accepts a ProductSpace as input, rather than just a tuple of Sector objects, which also affects the ordering of data in TensorMap objects.

  • Improved Structural Information Handling: The structural information for TensorMap objects (linked to their HomSpace) is now cached globally, eliminating the need for recalculation when creating new TensorMap instances. This change reduces overhead, particularly in scenarios where such computations are frequent.

We hope these updates improve your experience with TensorKit.jl, especially when working with symmetric tensors. As always, we welcome any feedback, issues, comments or feature requests.

5 Likes