CI: 9 tests failing with PyTorch 1.10(dev)

[Borda]

Discussed in #9273

^{Originally posted by daniellepintz September 2, 2021}
I noticed in recent commits these 9 tests are consistently failing in the Conda 1.10 environment

(https://github.com/PyTorchLightning/pytorch-lightning/runs/3483032531)

FAILED tests/callbacks/test_quantization.py::test_quantization[True-True-average]
FAILED tests/callbacks/test_quantization.py::test_quantization[True-False-average]
FAILED tests/callbacks/test_quantization.py::test_quantization[False-True-average]
FAILED tests/callbacks/test_quantization.py::test_quantization[False-False-average]
FAILED tests/models/test_amp.py::test_amp_cpus[1-bf16-None] - TypeError: __in...
FAILED tests/models/test_amp.py::test_amp_cpus[1-bf16-ddp_spawn] - TypeError:...
FAILED tests/models/test_amp.py::test_amp_cpus[2-bf16-None] - torch.multiproc...
FAILED tests/models/test_amp.py::test_amp_cpus[2-bf16-ddp_spawn] - torch.mult...
FAILED tests/models/test_amp.py::test_cpu_model_with_amp - AssertionError: Re...

Starting this thread to track this issue and try to resolve. Does anyone have any ideas? cc @Borda

[daniellepintz]

I am a bit confused because this test seems to be testing Pytorch 1.10 but according to https://pytorch.org/docs/versions.html v1.10 is not released yet

[Borda]

I am a bit confused because this test seems to be testing Pytorch 1.10 but according to https://pytorch.org/docs/versions.html v1.10 is not released yet

correct, we are building against the dev version to avoid any big breaking when the next version comes.
it is the same as FB build from PL master, but we are more conservative 🐰

[daniellepintz]

@Borda I see, thanks!

Update: after #9345 these are remaining failing tests:

FAILED tests/callbacks/test_quantization.py::test_quantization[True-True-average]
FAILED tests/callbacks/test_quantization.py::test_quantization[True-False-average]
FAILED tests/callbacks/test_quantization.py::test_quantization[False-True-average]
FAILED tests/callbacks/test_quantization.py::test_quantization[False-False-average]

[daniellepintz]

Stack trace for test_quantization[False-True-average]:

         quant_score = torch.mean(torch.tensor([mape(qmodel(x), y) for x, y in dm.test_dataloader()]))
         # test that the test score is almost the same as with pure training
>       assert torch.allclose(org_score, quant_score, atol=0.45)
E       assert False
E        +  where False = <built-in method allclose of type object at 0x7f66d3b6a780>(tensor(0.3262), tensor(0.9988), atol=0.45)
E        +    where <built-in method allclose of type object at 0x7f66d3b6a780> = torch.allclose

tests/callbacks/test_quantization.py:56: AssertionError

It seems quant_score is higher than it should be. When I run the test locally (Pytorch 1.9) quant_score = 0.5105

@Borda since you worked on this do you have any idea what might be the issue? If not should we check with Pytorch folks?

[Borda]

It seems quant_score is higher than it should be. When I run the test locally quant_score = 0.5105
@Borda since you worked on this do you have any idea what might be the issue? If not should we check with Pytorch folks?

as nothing changed with PL callback I would open parallel talk to Pytorch folks and Ill check it on PL side

[daniellepintz]

Sounds good, Pytorch issue created: pytorch/pytorch#64564

[daniellepintz]

So as of latest commit on master we now have 30 tests failing in the Pytorch 1.10 env: https://github.com/PyTorchLightning/pytorch-lightning/runs/3571593663

Here is what I notice about this job:

1. it has been failing for months at least

2. it seems like the failures are mostly due to changes on the PT side (since we are pulling new PT1.10 nightly) rather than reflecting actual issues from changes on our side. Also, the failures are constantly changing/increasing every day, due to the changing nature of the core dependency, PT1.10. It seems unsustainable to be constantly fixing issues from this test environment.

3. this CI job is the last thing standing in our way of a green CI. There are a few other jobs which are flaky and sometimes fail, but this test consistently fails. Some recent commits on PL master only have one CI failure (this job).

Given this, what do people think about deleting this test? If this test is important, then we should make it required which will force us to keep it green, but I feel it is really harmful to have it constantly making our CI red, and don't think it is sustainable to keep fixing it all the time.

cc @PyTorchLightning/core-contributors

[carmocca]

Testing nightly is useful to track these errors. However, you don't want to mark it as required to avoid upstream blocking us.

Since GitHub doesn't allow having "informative" jobs, if we want to keep the CI green but not mark a nightly job as required, the only option I can think of is only enabling it when there's a release candidate available which should be more stable.

[Borda]

Testing nightly is useful to track these errors. However, you don't want to mark it as required to avoid upstream blocking us.

we do not have this nightly versions as required 🐰