Test more WMMA configurations #171
Conversation
thomasfaingnaert
force-pushed
the
tf/test-more-configs
branch
3 times, most recently
from
07ff68b
to
4078daa
Compare
thomasfaingnaert
force-pushed
the
tf/test-more-configs
branch
from
4078daa
to
d06271d
Compare
| catch err | ||
| # Count tests with config errors as "broken". | ||
| if isa(err, GemmKernels.ConfigError) | ||
| @test true skip=true |
There was a problem hiding this comment.
@maleadt For now, I've marked unsupported configurations as "Broken". The term is not really appropriate, as broken tests should be tests that ought to pass, but do not currently. We could also just mark these as "pass", but I quite like that they are reported separately so we can easily see how many configurations are skipped due to ConfigErrors. WDYT?
There was a problem hiding this comment.
Ideally we hard-code whether a configuration is unsupported, i.e., which configuration we know is supposed to throw a ConfigError. That way the broken would work as intended, resulting in test failures when it anything starts failing/passing without updating the tests. However, I guess that would be hard to do, given how we now generate configurations using multiple loops. Maybe we could maintain a separate list of known-broken configurations? I guess it would also depend on the device though (because of the shmem limitation), so maybe that's not feasible either...
thomasfaingnaert
force-pushed
the
tf/test-more-configs
branch
from
fa2b56a
to
4b435ec
Compare
|
Remaining CI failure is due to illegal memory access during benchmarking, which should be fixed in CUDA 12.3 Update 2. |
No description provided.