MAINT unpin cython in CI config#27627
Conversation
| - blas[build=openblas] | ||
| - scipy=1.5.0 # min | ||
| - cython<3.0.0 | ||
| - cython=0.29.33 # min |
There was a problem hiding this comment.
Since other dependencies used min dependencies on this build I decided to follow this rule for cython as well.
The doc lock file generation works fine on my machine and takes around 30 seconds. |
|
@lesteve can you please push the results to this PR? |
|
Once the CI has completed. |
|
I am unable to reproduce the weird scipy-dev error locally. On the other hand I get plenty of errors locally because of deprecated things in Python 3.12 some of them are not inside scikit-learn e.g. joblib, dateutil (dateutil/dateutil#1284). Probably the easiest thing to do would be to stay on Python 3.11 for scipy-dev for now (there is a chance that it fixes the scipy-dev CI if I have to guess). scipy-dev error |
|
Let me push this. |
|
I pushed too early without waiting for the end of the PyPy tests of the previous build but they are still green after 1h55 of running, which is a net improvement compared to the last scheduled job that was crashing from the start: Still they are too slow to run to wait for the review of this PR. Let's see the results of the next nightly PyPy build after the merge of this PR. |
|
Yeah! the |
jjerphan
left a comment
There was a problem hiding this comment.
LGTM! Thank you for the recent work on investigating performance regression with Cython 3!
I think we can create "good first Cython issue" to replace interface inclusion which has been introduced in Cython 3 such as this one:
What do you think?
Co-authored-by: Loïc Estève <loic.esteve@ymail.com>
Co-authored-by: Loïc Estève <loic.esteve@ymail.com>
#27086 was closed as Cython 3.0.3 fixed all the Cython 3 related performance regressions we observed in the benchmarks.
So let's unpin Cython.
@lesteve I also updated the script to make it easy to display the commands and their outputs.
Note: the conda-lock command for
"doc"runs indefinitely on my host, so I had to interrupt it with ctrl-C. There is no error message. I don't know how to debug this.@lesteve or someone else could you please try to see if you can reproduce?