cuda-python
cuda-python copied to clipboard
Draft a version-pinning advice for downstream projects
We don't want people to pin any of our projects too tightly that results in version conflicts between packages. We need to give advices.
Current thought after discussing with @brycelelbach: For packages that depend on the compiler (NVCC, numba-cuda, or cutile-python in the future), they should agree on a lower bound of the compiler version, and the compiler should in turn loosely pin a set of packages that it needs. We might want to consider collectively bumping the lower bound when a new CUDA major (or minor?!) release is out. Let me raise this discussion in the Python wg.