@janosh Turns out there isn't much speed advantage for setup-uv over pip install uv, the first time running setup-uv is even slower. Perhaps we just keep things simple (also easy to maintain) and still use pip install uv for now.

Turns out there isn't much speed advantage for setup-uv over pip install uv, the first time running setup-uv is even slower.

good to know. i wasn't expecting it to be faster on a cold cache though. mainly hoping setup-uv would have good caching and be faster on subsequent runs

i wasn't expecting it to be faster on a cold cache though. mainly hoping setup-uv would have good caching and be faster on subsequent runs

Tests on second runs (native mamba + native uv (second run)) didn't show obvious advantage. Probably because our installation stage is not bottlenecked by dependency download, see extracted debug info below:

pip install torch (70% of the total time):

2024-09-13 02:33:35.316  torch install starts (`Collecting torch==2.2.1 ...`)
2024-09-13 02:34:05.962  All downloading finished (`Installing collected packages: ...`)
2024-09-13 02:34:42.070 Install finished (`Successfully installed ...`)

uv pip install -e .[extras]:

2024-09-13 02:34:42.691 pymatgen dependency install starts (`DEBUG Found static `pyproject.toml` for: pymatgen`)
2024-09-13 02:34:52.577 DEBUG Building: pymatgen @ file:///home/runner/work/pymatgen/pymatgen
2024-09-13 02:35:10.553 DEBUG Finished building: pymatgen @ file:///home/runner/work/pymatgen/pymatgen

thanks so much for your benchmarking! ❤️ that's super useful for the decision at hand and later reference. i agree seems there's not much point in adopting setup-uv.

merging now. thanks for cleaning up the needless build-time deps install

that's super useful for the decision at hand and later reference

No problem! That's what I was hoping too. Looks like torch is the huge bottleneck, hoping #3826 could have a solution soon :)