MAINT: SciPy 1.5.0rc2 backports #12345
Conversation
It looks like a bug in a particular OpenBLAS version that we use in python3.7-dbg in CI.
Using np.prod(..., dtype=int) can overflow on 32-bit platforms. Use scipy._lib._util.prod instead.
Cobyla is not threadsafe and can segfault when called from more than one thread. The problem is in f2py. Until the problem is fixed upstream in NumPy (f2py) we must synchronize the calls to cobyla with a lock.
* update the release notes for SciPy `1.5.0rc2` backports
|
Yep, looks like you'll need #12335, which I just now marked as backport-candidate (sorry about that!) |
* add scipygh-12335 to SciPy 1.5.0 release notes prior to rc2 release
|
I think I need a backport for pydata/sparse related fix in Travis CI re: Python 3.7 + NumPy dev wheels cc @hameerabbasi I thought we fixed that somewhere--I'll try to dig it up, should be fairly recent |
|
It's #12123. |
|
It must be something else--that PR is already in the maintenance branch. I think Pauli did another more recently. |
|
Ah yes, Pauli added a skip on the pydata/sparse install from NumPy pre-wheels in our CI here: #12285 So I'll probably want to backport that |
The run is too slow for the test suite to complete, so split it into two batches.
Numpy prerelease appears to cause problems for pydata/sparse, so avoid running tests concerning it.
* update `1.5.0` release notes to include scipygh-12285
|
That 32-bit Windows Azure job/mingw DL time is highly problematic; I'll give it one more restart.. |
|
Ok, I restarted the 32-bit Windows Azure CI job 3 or 4 times and it never finished on time. However, the most recent run clearly shows that the test suite did pass and time simply ran out in the subsequent cleanup stages, so this should be safe to merge and I'm behind schedule, so in it goes. |
Backports included for those PRs with
1.6.0milestone and a current backport label:test_heequb, it fails intermittently. #12276Rotationmultiplication #12317I have a lingering suspicion I may also need to backport a CircleCI sphinx version pinning for CI to pass as well, though no such PR was labelled as above--I guess we'll see what happens.
Release notes also updated based on above backports.