cupy has support for detecting duplicate installations. If done poorly this can slow down imports substantially, so we should be careful, but the last set of improvements I made to cupy's approach should make it fast enough to be suitable for RAPIDS. This would protect against coinstallation of e.g. pylibcudf-cu12 and pylibcudf-cu13 wheels, or the pylibcudf conda package.
cupy has support for detecting duplicate installations. If done poorly this can slow down imports substantially, so we should be careful, but the last set of improvements I made to cupy's approach should make it fast enough to be suitable for RAPIDS. This would protect against coinstallation of e.g. pylibcudf-cu12 and pylibcudf-cu13 wheels, or the pylibcudf conda package.