Skip to content

[Bug Skip]: test_activation_checkpointing_xpu.test_ac_rematerialize_with_sdpa_dropout_zero failed with error DumpError: can't serialize <class 'torch.nn.attention._SDPBackend'#4946

Description

@shangerxin

馃悰 Describe the bug with skip template

third_party.torch-xpu-ops.test.xpu.dynamo.test_activation_checkpointing_xpu.RematerializeACNodesPassTests::test_ac_rematerialize_with_sdpa_dropout_zero failed because the test runner could not serialize a torch.nn.attention._SDPBackend object while reporting a subtest result through pytest-xdist.

The actual failure is not from the test assertion itself, but from test infrastructure during result reporting. While pytest_subtests and xdist were sending the subtest report from the worker process, Python serialization raised:

DumpError: can't serialize <class 'torch.nn.attention._SDPBackend'>

This indicates the subtest failure metadata or attached exception context included an _SDPBackend value that cannot be pickled/serialized for transport between processes. As a result, the test run aborts with an infrastructure serialization error instead of a normal test failure report.

Short version for an issue template:

The test test_ac_rematerialize_with_sdpa_dropout_zero fails in test reporting, not necessarily in model logic. pytest-xdist cannot serialize torch.nn.attention._SDPBackend when emitting a subtest report, causing DumpError: can't serialize <class 'torch.nn.attention._SDPBackend'>. This appears to be a test infrastructure / serialization issue involving subtest result payloads.

Versions

XPU PR

#4894

Versions

Version context from report: pytorch main 2.15.0a0
pytorch main 2.15.0a0

Metadata

Metadata

Assignees

No one assigned

    Labels

    skippedUsed for temp UT failure to parallel fix

    Type

    No type

    Projects

    No projects

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions