Replace abandoned scikit-cuda #63
Description
Looks like scikit-cuda has been abandoned, which is unfortunate. Currently I can't even install it without manually updating the source code.
Looks like scikit-cuda has been abandoned, which is unfortunate. Currently I can't even install it without manually updating the source code.
Activity
johnh2o2 commentedon Jun 11, 2026
Status update — v1.0 (branch
v1.0, tip 39d3367) ships two mitigations:cuvarbase/_skcuda_compat.py): restores the aliases scikit-cuda 0.5.3 references (np.typeDict,np.float/np.int/np.complex,np.sctypes) before anyskcudaimport, so cuvarbase installs and runs against numpy ≥ 1.24 / 2.x without hand-patching scikit-cuda.__getattr__incuvarbase/__init__.py):import cuvarbaseand the BLS / CE / PDM paths no longer touch scikit-cuda at all — only the Lomb-Scargle/NFFT modules (which need cuFFT) import it, and they apply the shim first.This is verified in CI by a build-wheel → clean-venv → import smoke test that runs with only numpy+scipy installed.
Full replacement is still planned post-1.0: scikit-cuda is only used for cuFFT access, so the candidates are CuPy's
cupy.cuda.cufftor a direct cuFFT binding alongside the existing PyCUDA streams. Keeping this issue open to track that.johnh2o2 commentedon Jun 12, 2026
v1.0 release note (deferral confirmed): the packed v1.0 release ships with scikit-cuda 0.5.3 still backing the Lomb-Scargle/NFFT cuFFT path. What shipped instead of the migration:
cuvarbase/_skcuda_compat.py) applied automatically before skcuda loads;import cuvarbaseand the BLS/CE/PDM paths no longer touch skcuda at all.The actual replacement (cupy.cuda.cufft or a direct cuFFT binding) remains committed post-1.0 as this issue describes; the v1.0 CHANGELOG carries the same statement under Known limitations and deferred work.
B2: in-house cuFFT ctypes binding, drop the scikit-cuda import (#63)
GPU batch 2 (A5000): validate B2/B3/C1; complete B2 by dropping sciki…
johnh2o2 commentedon Jun 18, 2026
Hey @astrobatty , as usual thank you!! I made some decent progress on v1.0 in the 2 days that Fable 5 was available (you can surely tell the above was from Fable); I'm planning to wait until its back for a full audit and then manual review; v1.0 should be production ready soon!
astrobatty commentedon Jun 19, 2026
Expect some changes in BLS, but otherwise sounds good!
johnh2o2 commentedon Sep 28, 2026
Release triage for PR #69 (cuvarbase 1.0.1, source
6b6c48945160).This issue is implemented in PR #69 and will close automatically when that PR merges into master. Keeping it open until the merge.
Lomb–Scargle/NFFT now use the in-house ctypes cuFFT binding. scikit-cuda is absent from runtime dependencies and imports, and the old NumPy compatibility shim is gone. The prior June comments describing replacement as deferred are superseded by this implementation; the original comments remain as history.
Evidence: direct lazy cuFFT binding · NFFT call-site migration · current package dependencies