Hi. So there are several things related to logs that have been fixed (or were supposed to) :

1. You do not have to invent names that start with brian2 to make use of Brian's infrastructure. E.g., previously brian2cuda.cuda_generator displayed as brian2.codegen.generators.cuda_generator, which didn't correspond to an actual package. This is no longer necessary.
2. In Brian2CUDA's tests, you can ask catch_logs to only consider logs from brian2cuda, so we don't have to worry about Brian2 adding/removing warnings.
3. In Brian2's tests, only logs from brian2 are considered, so having a warning in Brian2CUDA doesn't break tests

Points 1 and 2 work fine, point 3 is unfortunately broken in 2.6.0. (but already fixed in master). For now, this simply means that we need to keep the workarounds that were already in place, i.e. the ignore_brian2_tests feature in brian2cuda.tests.__init__, and the duplicated test_neurongroup.py::test_semantics_floor_division test.

I didn't follow the run_args implementation in detail, but I'm assuming its mostly copy-paste since its running all in host code, right?

Yes, it's ugly, but it is exactly the same ugliness as in C++ standalone ;)

And are you running the test suite locally with GPU access? Or should I do that (I would have to set things up first since our cluster infrastructure has changed and I haven't used it since).

I've been running the tests on my laptop with a GPU – not fast, but good enough for now.

I forgot to fix a test (for a silly thing, see brian-team/brian2#1520), but now all tests pass on my machine 👍 This also means it is time to do a new release! If I am not mistaken, this simply means creating a new tag (v1.0a4?), so this should be straightforward.