Lazylinop news

2026-05-18

• Move lazylinop.butterfly.dct() and lazylinop.butterfly.dst() functions out of work-in-progress status.

• quantization module is compatible with KsmPermLazyLinOp now.

2026-04-21

lazylinop.butterfly.optimize() works with instance of Ksm*LazyLinOp classes.

2026-04-04

Add Fast-Noiselet-Transform to butterfly and signal modules.

2026-03-18

Move lazylinop.butterfly.optimize() function out of work-in-progress status.

2026-03-13

Add complex quantization to the quantization module.

2026-02-12

• Move quantization module out of work-in-progress status.

• Add bf16 support to CUDA version of ksmm kernel.

2025-12-03

Add work-in-progress quantization module.

2025-10-13

Add partial singular value decomposition linalg.svds() to linalg module.

2025-09-15

Redesign and cleanup of LazyLinOp core class and introduction of ArrayBasedLazyLinOp class specialization for more efficient operators based on arrays.

2025-07-23

• ksd() works with LazyLinOp arguments.

• ksm() works with CuPy arrays and torch tensors ks_values.

2025-06-27

Add PyTorch and CuPy compatibility to basicops and signal modules (work-in-progress).

2025-06-05

Move butterfly module out of work-in-progress.

2025-06-03

Default data type of the LazyLinOp is None now. It is only used for LazyLinops created from arrays L = aslazylinop(array), in which case L.dtype = array.dtype. All other lazylinops have dtype = None.