Closed djhoese closed 5 months ago
All modified and coverable lines are covered by tests :white_check_mark:
Comparison is base (
f892115
) 94.06% compared to head (fa3771b
) 93.99%. Report is 5 commits behind head on main.
:umbrella: View full report in Codecov by Sentry.
:loudspeaker: Have feedback on the report? Share it here.
Let's see if I can get all my profiling results for gradient search organized. These are all doing ABI C02 (500m resolution) full disk to a large eqc
target area and the final compute is going to nowhere so there is no to-disk writing involved in this. Only 4 dask workers involved. Chunk size set via PYTROLL_CHUNK_SIZE to 1356 but I think that still rounds up.
My indexing changes are negligible in this use case. They were maybe a bit faster (~2 seconds of 47) but to me that's nothing. Both newer Cython 3 versions (the first 2 results below) are about the same as main
on Cython 2. I'll try more workers in a second to see how it goes.
In reverse chronological/development order:
Note: I had to force Cython <3
to be used here otherwise it took 6+ minutes to run because it was hitting the GIL a ton. We need to get this PR merged and released as soon as possible.
Yeah no differences with 8 workers over 4 workers.
As mentioned on slack the unstable CI environment is failing due to upstream libraries like shapely and rasterio not being compatible with newest numpy 2 changes. I'll merge this PR as-is and then make another PR to test against my own custom fork of these libraries with fixes.
Just realized that I never updated CI to build wheels against numpy 2. I'll do that in another PR.
numpy.import_array()
for all Cython extensions to fix compatibility with Numpy 2.0TODO: I should update wheel building to use numpy 2 when creating wheels so future wheels will already be compatible.
git diff origin/main **/*py | flake8 --diff