Fix error in AXLearn tests, and remove unnecessary ones

[Steboss]

This PR does the following:

• remove array_serialization_test.py that's causing the test to hang with the following error:

[pod/axlearn-14925960362-xb57t/axlearn]   File "/opt/jax/jax/experimental/array_serialization/serialization.py", line 193, in __del__
[pod/axlearn-14925960362-xb57t/axlearn]     logger.warning('Please add `.wait_until_finished()` in the main thread '
[pod/axlearn-14925960362-xb57t/axlearn] Message: 'Please add `.wait_until_finished()` in the main thread before your program finishes because there is a possibility of losing errors raised if the this class is deleted before writing is completed.'
[pod/axlearn-14925960362-xb57t/axlearn] Arguments: ()
[pod/axlearn-14925960362-xb57t/axlearn] sssssssssssssss.ssssssssssss.ssssssssssssssssssssssFsFs.s...sssssssFs.Fs [ 97%]

causing the EKS job to run out of time - so we can't get the tests

• remove tests that are redundant (namely, there are similar test already running) such as:

"/opt/axlearn/axlearn/common/deberta_test.py"
"/opt/axlearn/axlearn/common/distilbert_test.py"
"/opt/axlearn/axlearn/common/trainer_test.py"
"/opt/axlearn/axlearn/common/decoder_test.py"
"/opt/axlearn/axlearn/common/adapter_torch_test.py"
"/opt/axlearn/axlearn/common/attention_test.py"
"/opt/axlearn/axlearn/common/convolution_test.py"

• remove tests for models that we're not currently using:

"/opt/axlearn/axlearn/common/mixture_of_experts_test.py"
"/opt/axlearn/axlearn/common/t5_test.py"
"/opt/axlearn/axlearn/common/vision_transformer_test.py"
"/opt/axlearn/axlearn/common/input_reading_comprehension_test.py"
"/opt/axlearn/axlearn/common/input_t5_test.py"

• remove tests like summary_writer_test.py that is mostly using python library we're not employing here (e.g. wandb)
• add the installation of pytest-xdist and pytest-reportlog to avoid the following error:

ERROR: usage: pytest [options] [file_or_dir] [file_or_dir] [...]
pytest: error: unrecognized arguments: --report-log=/tmp/tmp.iL7rVQtXBq --dist=load --tx --tx 

Overall, this should allow us to reduce the testing time from 50 minutes to 30 minutes, covering the most important tests as well, that are dealing with general AXLearn infrastructure.

[olupton]

I see an error that seems like a test/infra bug, rather than a failing test: https://github.com/NVIDIA/JAX-Toolbox/actions/runs/14931992154/job/41952767038#step:6:41440

Some of the tests look like they are failing due to missing input data.

Also, the CI job is marked as successful despite the tests failing.

[Steboss]

In this PR I added a workflow dispatch, so we can test single parts of the CI.
In particular, this may be a better trick to be used during CI, so that we can have MODE=SOMETHING to trigger only specific parts of the workflow, rather than testing everything

[Steboss]

@olupton
I can see that the axlearn tests now are working fine. We have some tests that are still failing - I can have a look at those.
Working on the way we're monitoring and returning the k8s job status, as it's still giving green flag on the axlearn eks job.

[Steboss]

@olupton
It looks like we need the XLA_FLAGS="--xla_force_host_platform_device_count=8" for the for_8_devices tests, otherwise the XLA tests will fail as:

[pod/axlearn-15024599931-2vfmc/axlearn] ___________________ HostArrayTest.test_fixed_process_shape67 ___________________
[pod/axlearn-15024599931-2vfmc/axlearn] [gw93] linux -- Python 3.12.3 /usr/bin/python3
[pod/axlearn-15024599931-2vfmc/axlearn] 
[pod/axlearn-15024599931-2vfmc/axlearn] self = <axlearn.common.host_array_test.HostArrayTest testMethod=test_fixed_process_shape67>
[pod/axlearn-15024599931-2vfmc/axlearn] platform = 'cpu', mesh_shape = (-1, 2), process_shape = [1]
[pod/axlearn-15024599931-2vfmc/axlearn] partition = PartitionSpec('data', 'model')
[pod/axlearn-15024599931-2vfmc/axlearn] 
[pod/axlearn-15024599931-2vfmc/axlearn]     @parameterized.product(
[pod/axlearn-15024599931-2vfmc/axlearn]         platform=("cpu", "tpu"),
[pod/axlearn-15024599931-2vfmc/axlearn]         mesh_shape=[
[pod/axlearn-15024599931-2vfmc/axlearn]             (-1, 1),  # Fully partitioned along one dim.
[pod/axlearn-15024599931-2vfmc/axlearn]             (2, -1),  # Partitioned along multiple dims.
[pod/axlearn-15024599931-2vfmc/axlearn]             (-1, 2),  # Test the other way.
[pod/axlearn-15024599931-2vfmc/axlearn]             (1, -1),
[pod/axlearn-15024599931-2vfmc/axlearn]         ],
[pod/axlearn-15024599931-2vfmc/axlearn]         process_shape=[
[pod/axlearn-15024599931-2vfmc/axlearn]             # Each process produces single dim.
[pod/axlearn-15024599931-2vfmc/axlearn]             [1],  # Not divisible by number of devices (replicated).
[pod/axlearn-15024599931-2vfmc/axlearn]             [8],  # Divisible by number of devices.
[pod/axlearn-15024599931-2vfmc/axlearn]             [16],  # Multiple elements per device.
[pod/axlearn-15024599931-2vfmc/axlearn]             # Each process produces multiple dims.
[pod/axlearn-15024599931-2vfmc/axlearn]             [1, 1],  # Not divisible by number of devices (replicated).
[pod/axlearn-15024599931-2vfmc/axlearn]             [2, 1],  # Can be partitioned over dim=0, replicated on dim=1.
[pod/axlearn-15024599931-2vfmc/axlearn]             [16, 1],  # Multiple elements per device.
[pod/axlearn-15024599931-2vfmc/axlearn]             [2, 4],  # Can be fully partitioned.
[pod/axlearn-15024599931-2vfmc/axlearn]             [8, 8],  # Can be fully partitioned.
[pod/axlearn-15024599931-2vfmc/axlearn]         ],
[pod/axlearn-15024599931-2vfmc/axlearn]         partition=(
[pod/axlearn-15024599931-2vfmc/axlearn]             DataPartitionType.FULL,
[pod/axlearn-15024599931-2vfmc/axlearn]             DataPartitionType.REPLICATED,
[pod/axlearn-15024599931-2vfmc/axlearn]             PartitionSpec("data"),
[pod/axlearn-15024599931-2vfmc/axlearn]             PartitionSpec("data", "model"),
[pod/axlearn-15024599931-2vfmc/axlearn]         ),
[pod/axlearn-15024599931-2vfmc/axlearn]     )
[pod/axlearn-15024599931-2vfmc/axlearn]     # NOTE: while annotated with `for_8_devices`, this runs on other configurations.
[pod/axlearn-15024599931-2vfmc/axlearn]     @pytest.mark.for_8_devices
[pod/axlearn-15024599931-2vfmc/axlearn]     def test_fixed_process_shape(
[pod/axlearn-15024599931-2vfmc/axlearn]         self,
[pod/axlearn-15024599931-2vfmc/axlearn]         platform: str,
[pod/axlearn-15024599931-2vfmc/axlearn]         mesh_shape: tuple[int, int],
[pod/axlearn-15024599931-2vfmc/axlearn]         process_shape: Sequence[int],
[pod/axlearn-15024599931-2vfmc/axlearn]         partition: Union[DataPartitionType, PartitionSpec],
[pod/axlearn-15024599931-2vfmc/axlearn]     ):
[pod/axlearn-15024599931-2vfmc/axlearn]         """Tests roundtrip host-to-global and global-to-host with fixed process shape."""
[pod/axlearn-15024599931-2vfmc/axlearn]     
[pod/axlearn-15024599931-2vfmc/axlearn] >       mesh_shape = infer_mesh_shape(mesh_shape)
[pod/axlearn-15024599931-2vfmc/axlearn] 
[pod/axlearn-15024599931-2vfmc/axlearn] axlearn/common/host_array_test.py:124: 
[pod/axlearn-15024599931-2vfmc/axlearn] _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 
[pod/axlearn-15024599931-2vfmc/axlearn] 
[pod/axlearn-15024599931-2vfmc/axlearn] mesh_shape = (-1, 2)
[pod/axlearn-15024599931-2vfmc/axlearn] 
[pod/axlearn-15024599931-2vfmc/axlearn]     def infer_mesh_shape(mesh_shape: MeshShape, *, num_devices: Optional[int] = None) -> MeshShape:
[pod/axlearn-15024599931-2vfmc/axlearn]         """Infer the value for -1 from len(jax.devices()) and other dims if there is -1 in mesh shape.
[pod/axlearn-15024599931-2vfmc/axlearn]     
[pod/axlearn-15024599931-2vfmc/axlearn]         Args:
[pod/axlearn-15024599931-2vfmc/axlearn]             mesh_shape: The original MeshShape, which might have -1 in one axis.
[pod/axlearn-15024599931-2vfmc/axlearn]             num_devices: The devices that will be used to construct the mesh.
[pod/axlearn-15024599931-2vfmc/axlearn]                 If None, defaults to len(jax.devices()).
[pod/axlearn-15024599931-2vfmc/axlearn]     
[pod/axlearn-15024599931-2vfmc/axlearn]         Returns
[pod/axlearn-15024599931-2vfmc/axlearn]             A new MeshShape with inferred value for -1.
[pod/axlearn-15024599931-2vfmc/axlearn]         """
[pod/axlearn-15024599931-2vfmc/axlearn]         if -1 not in mesh_shape:
[pod/axlearn-15024599931-2vfmc/axlearn]             return mesh_shape
[pod/axlearn-15024599931-2vfmc/axlearn]     
[pod/axlearn-15024599931-2vfmc/axlearn]         if mesh_shape.count(-1) > 1:
[pod/axlearn-15024599931-2vfmc/axlearn]             raise ValueError(f"Only one axis can be -1 in {mesh_shape=}.")
[pod/axlearn-15024599931-2vfmc/axlearn]     
[pod/axlearn-15024599931-2vfmc/axlearn]         # Handle the case with one -1.
[pod/axlearn-15024599931-2vfmc/axlearn]         prod = math.prod(mesh_shape, start=-1)
[pod/axlearn-15024599931-2vfmc/axlearn]         if num_devices is None:
[pod/axlearn-15024599931-2vfmc/axlearn]             num_devices = len(jax.devices())
[pod/axlearn-15024599931-2vfmc/axlearn]         if num_devices % prod != 0:
[pod/axlearn-15024599931-2vfmc/axlearn] >           raise ValueError(
[pod/axlearn-15024599931-2vfmc/axlearn]                 f"Unable to infer -1 in mesh shape {mesh_shape} as num_devices {num_devices} "
[pod/axlearn-15024599931-2vfmc/axlearn]                 f"is not a multiple of the product {prod} of mesh axes."
[pod/axlearn-15024599931-2vfmc/axlearn]             )
[pod/axlearn-15024599931-2vfmc/axlearn] E           ValueError: Unable to infer -1 in mesh shape (-1, 2) as num_devices 1 is not a multiple of the product 2 of mesh axes.
[pod/axlearn-15024599931-2vfmc/axlearn] 
[pod/axlearn-15024599931-2vfmc/axlearn] axlearn/common/utils.py:1834: ValueError

I ran a test, that results in:

• with flag: 1 failed, 161 passed
• without flag: 203 failed, 100 passed

[Steboss]

This may be a new JAX version error

AttributeError: module 'jax.experimental.array_serialization.serialization' has no attribute '_spec_has_metadata'

[olupton]

I think we can merge this to speed up the pipeline, but I left some more comments on error handling and robustness.

[olupton]

I think we can merge this as an improvement to the current situation.

The job is still marked green despite test errors/failures, although they do seem to be captured in the badge. This should be fixed in a follow-up.