ROCm / apex

A PyTorch Extension: Tools for easy mixed precision and distributed training in Pytorch
BSD 3-Clause "New" or "Revised" License
19 stars 17 forks source link

Skip the unit tests #41

Closed lcskrishna closed 3 years ago

lcskrishna commented 3 years ago

Skipping unit tests for stability of CI.

lcskrishna commented 3 years ago

jenkins: retest this please

lcskrishna commented 3 years ago

These are the unit tests that got regressed when using the latest ROCm versions.

test_fuzz (Mulit-tensor-axpy)
test_fuzz (Multi_tensor_scale)
test_bfloat16( FUsedAdagrad)
test_float (FusedAdagrad)
test_multi_params (FusedAdagrad)
test_float (FusedAdam)
test_bfloat16 (FusedAdam)
sunway513 commented 3 years ago

@lcskrishna can you confirm if the tests can pass using the public ROCm3.7 docker container?

lcskrishna commented 3 years ago

@sunway513 Public docker containers doesn't have hipify changes. With the hipify changes from pytorch installed on ROCm3.7 I don't see any issues locally.

sunway513 commented 3 years ago

Thanks, that’s helpful. LGTM, please go ahead merge it.

lcskrishna commented 3 years ago

Thanks!