Conversation
|
I was wrong to assume that this would be simple. In order for Cython 3 to work for us we need to also install numpy 1.26 (as older numpys do not work with Cython 3). This has a number of issues:
@JDBetteridge, you were completely right about this being a hard problem. I vote that we wait for numpy 2.0 to land when the build system is sane again. However, I hope it's at least helpful that I've fixed |
| # Particularly important for debugging petsc fails. | ||
| with environment(**blas): | ||
| pipargs = ["-vvv"] + pipargs | ||
| pipargs = ["-v"] + pipargs |
|
@Ig-dolci FYI I am having another stab at this now. If we use the release candidate for numpy 2.0 then we can hopefully land Cython 3 and numpy at the same time. |
|
This is quite the pain, now loopy is also having issues. |
| run_pip_install(["typing_extensions"]) | ||
| extra_index_url = ["--extra-index-url", "https://download.pytorch.org/whl/cpu"] if args.torch == "cpu" else [] | ||
| run_pip_install(["torch"] + extra_index_url) | ||
| # FIXME: use nightly to see if that makes numpy 2.0 happy |
There was a problem hiding this comment.
Should really be
pip install torch==2.3.0 --index-url https://download.pytorch.org/whl/test/cpu
as detailed in https://dev-discuss.pytorch.org/t/pytorch-release-2-3-0-final-rc-is-available/1995
2.3.0 is supposed to be released 24/04.
|
Just to double check: are the numpy changes here backward compatible with 1.x, or not? cython 3.x is somewhat widespread now, while it will still take a while to get numpy 2.x widely used (at least, on installations that do not use firedrake-install) |
I think so. The errors fixed here were mostly due to numpy dropping support for deprecated API. |
|
Will this be closed in favor of #3546? |
Yes. |
3 participants
Description
Fixes #3081
It will also allow us to update our PETSc fork as petsc4py now requires cython 3.