You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
[DISCUSSION] Consider making cupy>=14 a dependency #85
Starting cupy 14 wheels, have the option on including a ctk, by installing like cupy-cuda12x[ctk]" or "cupy-cuda13x[ctk]" . This makes cupy installed via pip to not depend on a system cuda toolkit., which happens in several systems.
When we install via conda this is not an issue, as conda includes it already.
cupy-xarray, can't work without cupy, at the moment in main cupy is not a dependency, I wonder if the situation described above, makes it easier to implement a reliable path to workable dependencies.
@leofang Maybe you have a suggestion, on how to handle this, for a bit more context the dependency was removed because the different builds, i.e cupy, cupy-cuda12x and cupy-cuda13x.
How do other libraries handle this? See extended discussion in #66 (comment)
Activity
ncclementi commentedon Mar 13, 2026
@leofang Maybe you have a suggestion, on how to handle this, for a bit more context the dependency was removed because the different builds, i.e cupy, cupy-cuda12x and cupy-cuda13x.
How do other libraries handle this? See extended discussion in #66 (comment)
weiji14 commentedon May 23, 2026
Proper solution would be to wait for stabilization of wheel variants (PEP 825) and for cupy to support it. See upstream thread at cupy/cupy#9258
Alternative solution I mentioned at #66 (comment) is to wait for PEP 771 – Default Extras to be stabilized, and pick a default cuda version.
These are both part of the WheelNext initiative. Some tools like
uvare now variant-enabled, and one could install from a special index (e.g. https://wheelnext.github.io/variants-index-test/v0.0.3/), but I'm not quite sure how this should be enabled for downstream libraries likecupy-xarrayyet.