Feature
Since we have switched to GPU runners, tests can be rewritten:
- run all tests on GPU by default
- FA2 + optimum quanto are installed in
make test workflow, s.t. no test is skipped
For 1., one could add a fixture that returns the device (usual "cuda" if torch.cuda ...) and update the tests accordingly. It is still useful to have the possibility to run tests on cpu for debugging purposes on non gpu machines, i.e. no hardocding of devices.
Additionally, one could add slower integration tests (e.g. test ruler performance @50% CR) that are triggered from time to time (see e.g. here for an example). This would probably go into another .yml.
Motivation
- Better GPU utilization
- GPU tests are relevant for this library
Feature
Since we have switched to GPU runners, tests can be rewritten:
make testworkflow, s.t. no test is skippedFor 1., one could add a fixture that returns the device (usual "cuda" if torch.cuda ...) and update the tests accordingly. It is still useful to have the possibility to run tests on cpu for debugging purposes on non gpu machines, i.e. no hardocding of devices.
Additionally, one could add slower integration tests (e.g. test ruler performance @50% CR) that are triggered from time to time (see e.g. here for an example). This would probably go into another .yml.
Motivation