-
Notifications
You must be signed in to change notification settings - Fork 87
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
test: update test_2799 for Numba 0.59.0 #2998
Conversation
This is just the sort of thing that comes out of nowhere and stops a release:
Is it saying that we need just one more thing in CMakeLists.txt? If so, why did this never come up before, and why only on one version of one build? Why, o why? |
In fact, awkward/awkward-cpp/CMakeLists.txt Lines 1 to 9 in b0322b5
|
@jpivarski no, that's a symptom that the build environment on Windows is not working -- we need to discover the C++ compilers to be able to proceed. The OS is I'm spitballin' as they say, but I don't think it's necessarily a problem. We can check to see whether the windows image has changed, for example. |
|
I'll try |
OK, it seems unlikely that the image has changed. I note that this fails later when trying to build llvmlite. I suspect this is because Numba no longer releases wheels for this platform - I know that they've just released a new version. But, I think that would be a red herring - we should see the same error in |
This is ultimately being caused by pip trying to build an sdist of llvmlite because PyPI doesn't have a 32-bit wheel. I am content not to get to the bottom of why this happens, but I suspect the traceback is caused by llvmlite's internal detection logic (https://github.com/numba/llvmlite/blob/f22420ad768a31dd15ae45d551bb4e73907a27fa/ffi/build.py#L56). This follows from the fact that an awkward-cpp build succeeds in the same job (although it might use a different CMake from PyPI, that's irrelevant). Off-hand I can't remember @jpivarski when we plan to drop 32-bit windows. It won't be hard to keep it running for now; we just need disable numba on win (32bit) |
That's a good idea—dropping Numba on the 32-bit Windows test. I'll do that. The Windows 32-bit test was historically very valuable in helping us eliminate cases in which we wrongly assumed that You've also answered the "why now?" question: because Numba released 0.59.0 so recently. |
Numba 0.59.0 is now available (on PyPI), so test_2799 was failing because its first line was
I think you were waiting on information about how Numba would be dealing with this issue when 0.59.0 comes out. Your expectation that
flattened.dtype
would beint8
doesn't seem to be right. Experimentally, it returnsint64
. Even for pure NumPy (no Awkward Array):I've updated the test with this correction (what I believe to be a correction).
#2996 and #2997 both need it to be merged, and they need to be merged to release Awkward 2.6.0. (In fact, any subsequent runs of the CI tests will need this, since they get Numba through PyPI and that's 0.59.0 now.)
Two other-library updates came out on the same day, both of which complicate our tests! (pytest 8.0.0 came out, which breaks pytest-asyncio, and Numba 0.59.0 came out.)