-
Notifications
You must be signed in to change notification settings - Fork 24.7k
Description
🐛 Describe the bug
For example, when #122503 was landed, it accidentally switched to using AVX2 only runners, which resulted in some test failing with being unable to compile a correct code, see https://hud.pytorch.org/pytorch/pytorch/pull/122503?sha=c599c89babdcd98e7048286eecd3ec27ee8e20eb
which were eventually fixed by #122608
But it proves that AVX2 is not tested in CI
Eager AVX2 and default dispatch are tested using
auto envar = std::getenv("ATEN_CPU_CAPABILITY"); |
While torch.compile SIMD selection can be controlled using
pytorch/torch/_inductor/codecache.py
Line 1223 in e3d80f2
if config.cpp.simdlen is None: |
but it would be ideal if the same environment variable could be used for both eager and compile
Also, it perhaps worth shifting some of the sanity testing from an integration to simple unittesting (though see #123219 , unittests are only good if they are run in CI)
Versions
CI
cc @seemethere @pytorch/pytorch-dev-infra @ezyang @msaroufim @bdhirsh @anijain2305 @chauhang