Fix DecorateInfo test-name matching for XPU overrides via register_test

[gplutop7]

Fixes:
#1973

This PR fixes a mismatch between overridden XPU test function names and DecorateInfo name matching.

What changed:

• Added a shared register_test() helper in XPU test utilities.
• Replaced direct Class.test_* = test* assignments with register_test(...) in affected XPU tests.

[Copilot]

Pull request overview

This PR fixes a name-mismatch problem where XPU test overrides assigned with Class.test_* = _test_* left the function's __name__ as _test_*, breaking DecorateInfo name matching. A new register_test() helper in test/xpu/xpu_test_utils.py wraps the override, rewrites __name__/__qualname__ to the public test_* name, and assigns it onto the target class. All affected XPU test modules are migrated to call this helper instead of doing the direct assignment.

Changes:

• Add register_test(cls, func, *, name=None) helper that validates the callable, derives the public test name (optionally via name=), and binds a properly named wrapper onto cls.
• Replace Class.test_* = _test_* patterns across many XPU test modules with register_test(...) calls, including a couple of cases that pass an explicit name= to remap to a CUDA-style name.
• Update imports in those test modules to also import register_test.

Reviewed changes

Copilot reviewed 16 out of 16 changed files in this pull request and generated 3 comments.

Show a summary per file

File	Description
test/xpu/xpu_test_utils.py	Adds the register_test helper and imports functools.
test/xpu/test_type_promotion_xpu.py	Migrates two test overrides to register_test.
test/xpu/test_spectral_ops_xpu.py	Migrates two overrides, one with explicit name=; leaves _compare_xpu_cpu helper as direct assignment.
test/xpu/test_sort_and_select_xpu.py	Migrates test_sort_large_slice override.
test/xpu/test_optim_xpu.py	Migrates three optimizer test overrides.
test/xpu/test_numba_integration_xpu.py	Migrates three numba integration test overrides.
test/xpu/test_multiprocessing_xpu.py	Migrates four overrides, two using explicit name= for CUDA-name remap.
test/xpu/test_modules_xpu.py	Migrates test_multiple_device_transfer; keeps _test_gradients_helper direct assignment.
test/xpu/test_fx_xpu.py	Migrates three FX profiler test overrides.
test/xpu/test_foreach_xpu.py	Migrates five foreach test overrides.
test/xpu/test_distributions_xpu.py	Migrates six distributions test overrides.
test/xpu/test_basic_torch_np_xpu.py	Migrates test_f16_on_cuda override.
test/xpu/quantization/core/test_quantized_tensor_xpu.py	Partially migrates overrides; two direct assignments remain.
test/xpu/quantization/core/test_quantized_op_xpu.py	Migrates test_max_pool2d_pt2e.
test/xpu/nn/test_pooling_xpu.py	Migrates many pooling test overrides.
test/xpu/export/test_export_training_ir_to_run_decomp_xpu.py	Migrates test_exception override.

[Copilot]

Pull request overview

Copilot reviewed 20 out of 20 changed files in this pull request and generated no new comments.

[Copilot]

Pull request overview

Copilot reviewed 22 out of 22 changed files in this pull request and generated 3 comments.

Comments suppressed due to low confidence (1)

test/xpu/xpu_test_utils.py:637

• removeprefix("_") only strips a single leading underscore. If a caller passes a function such as __test_foo (or any name with more than one leading underscore), the resulting test_name will still begin with _, the startswith("test") check will fail, and a ValueError will be raised. Either document the single-underscore expectation explicitly, or use lstrip("_") so any number of leading underscores are stripped before the prefix check.

    test_name = (name or func.__name__).removeprefix("_")
    if not test_name.startswith("test"):
        raise ValueError(
            f"Registered test name must start with 'test', got {test_name!r}"
        )

[Copilot]

Pull request overview

Copilot reviewed 22 out of 22 changed files in this pull request and generated no new comments.

[chuanqi129]

@copilot New UT failures detected in op_ut tests (4 total). Please check these new failures and analyze whether they are caused by the PR changes.

Important: Do NOT update the plan in the PR description directly. Use reply comments to update the status.

Job log: https://github.com/intel/torch-xpu-ops/actions/runs/25928824349
Full new failure report: https://github.com/intel/torch-xpu-ops/actions/runs/25928824349/artifacts/7032504143

New failures:

Category	Class name	Test name	Status	Message
op_ut	third_party.torch-xpu-ops.test.xpu.dynamo.test_misc_xpu.MiscTests	test_const_dict_variable_python_type	failed	AttributeError: '_thread._local' object has no attribute 'current_tx'
op_ut	third_party.torch-xpu-ops.test.xpu.profiler.test_profiler_xpu.TestProfiler	test_disable_external_correlation	failed	AssertionError: False is not true
op_ut	third_party.torch-xpu-ops.test.xpu.profiler.test_profiler_xpu.TestProfiler	test_dynamic_toggle	failed	AssertionError: False is not true
op_ut	third_party.torch-xpu-ops.test.xpu.test_custom_ops_xpu.TestCustomOpAPI	test_no_grad_skips_autograd	failed	AssertionError: Scalars are not equal!

Instructions:

1. Download the full new failure report artifact for complete error messages
2. Analyze whether the failures are caused by PR changes
3. If caused by PR changes, fix those failures on this PR branch directly
4. Add failing test cases with pytest commands in the PR description for the reproducer job to use
5. Use reply comments to report analysis status - do NOT modify the PR description plan directly

[Copilot]

@copilot New UT failures detected in op_ut tests (4 total). Please check these new failures and analyze whether they are caused by the PR changes.

**Important: Do NOT update the plan in the PR de...

After analyzing all 4 failures, they are not caused by this PR's changes:

1. test_const_dict_variable_python_type (test_misc_xpu.py) — This file doesn't exist in the repo locally; it's from a third-party component. The AttributeError: '_thread._local' object has no attribute 'current_tx' is a dynamo/torch.compile threading issue unrelated to test name registration.

2. test_disable_external_correlation and test_dynamic_toggle (test_profiler_xpu.py) — This file was not modified by this PR (zero diff) and does not import xpu_test_utils or register_test.

3. test_no_grad_skips_autograd (test_custom_ops_xpu.py) — This file was also not modified by this PR and does not import xpu_test_utils or register_test.

This PR only changes: test/xpu/xpu_test_utils.py and the 21 test files that use register_test (none of which include the 3 files with failures above). These failures appear to be pre-existing flaky tests or failures from unrelated environmental issues in the CI run.

[BBBela]

The PR converts many test registrations that don't appear to have confirmed
DecorateInfo breakage.
Given the stated plan to port all XPU tests to PyTorch directly, it might be worth
scoping this PR to only the confirmed broken cases to keep the PR focused.
It looks like most of those changes are preventive rather than fixing existing problems - is that correct?

[BBBela]

The current signature hides its contract — you have to read the body to know that the function name must start with test_ and that the target method name is derived from it. A three-argument form like that:

def register_test(original_test_class, original_test_func, overriden_func):

would make the relationship explicit at the call site and removes all the name-derivation logic entirely.

What do you think?

[gplutop7]

Thanks, I had the same thought initially.

A fully explicit 3-argument form makes the contract clearer, but it adds cost to the common case. Most overrides follow test* -> test_*, so requiring the target name everywhere would duplicate that repeatedly and make renames more fragile.

I also considered introducing the helper only for new/broken cases and leaving existing direct assignments as-is. But that would create two inconsistent patterns in the codebase: some tests using the helper, others the old direct assignment. That invites confusion for future contributors — which pattern should they copy? — and risks repeating the same fix in different ways.

So I went with a hybrid approach:

def register_test(cls, override_func, *, target_name=None):

This keeps the common case concise while requiring target_name for non-standard mappings. And I updated all existing registrations to use it consistently, so the pattern is now uniform across the test suite.

Does that resolve the scope concern? Do you see any risk in standardizing on this approach repo-wide?

[BBBela]

Thank you for response and clarification! 😃

I understand your point of view and it makes sense. I may even agree on the repo-wide scope to be consistent. However, regarding the function signature I am still not convinced, because the main argument for current approach is that it will make each registration several lines shorter, while making the contract less clear. So we trade the number of code lines at the cost of simplicity and clarity - at least that is my point of view.

But to be clear - those are my personal, subjective design preferences, so I am not going to block anything on this PR and I am not going to request any changes, I am just not approving something I do not fully agree with. So feel free to request others for review and proceed with this PR 😉

[gplutop7]

Thanks a lot for the thoughtful feedback and for the constructive discussion.

You were absolutely right about the clarity tradeoff, and I decided to follow your suggestion. I updated the helper to the fully explicit 3-argument form so the contract is now clear at each call site.

I also migrated all registrations in this PR to match that explicit style (including non-standard mappings), so the pattern stays consistent repo-wide while keeping the interface unambiguous

[chuanqi129]

@copilot New UT failures detected in op_ut tests (1 total). Please check these new failures and analyze whether they are caused by the PR changes.

Important: Do NOT update the plan in the PR description directly. Use reply comments to update the status.

Job log: https://github.com/intel/torch-xpu-ops/actions/runs/26033399176
Full new failure report: https://github.com/intel/torch-xpu-ops/actions/runs/26033399176/artifacts/7066563163

New failures:

Category	Class name	Test name	Status	Message
op_ut	third_party.torch-xpu-ops.test.xpu.functorch.test_vmap_xpu.TestVmapOperatorsOpInfoXPU	test_op_has_batch_rule_count_nonzero_xpu_float32	failed	Unexpected success

Instructions:

1. Download the full new failure report artifact for complete error messages
2. Analyze whether the failures are caused by PR changes
3. If caused by PR changes, fix those failures on this PR branch directly
4. Add failing test cases with pytest commands in the PR description for the reproducer job to use
5. Use reply comments to report analysis status - do NOT modify the PR description plan directly

[Copilot]

@copilot New UT failures detected in op_ut tests (1 total). Please check these new failures and analyze whether they are caused by the PR changes.

**Important: Do NOT update the plan in the PR de...

The failure in test_op_has_batch_rule_count_nonzero_xpu_float32 is not caused by this PR.

Evidence:

• test/xpu/functorch/test_vmap_xpu.py has zero diff in this PR (git diff HEAD~11 HEAD -- test/xpu/functorch/test_vmap_xpu.py is empty)
• The file does not import xpu_test_utils or register_test at all

The failure is "Unexpected success" — meaning count_nonzero was marked xfail("count_nonzero") as expected-to-fail, but it now passes. This indicates an upstream PyTorch change added a batching rule for count_nonzero, making the previously expected failure now succeed. The fix (removing xfail("count_nonzero") from test_vmap_xpu.py) should be done in a separate PR targeting that test file.

The PR converts many test registrations that don't appear to have confirmed
DecorateInfo breakage.
Given the stated plan to port all XPU tests to PyTorch directly, it might be worth...

That's correct — the majority of the conversions are preventive. The __name__ mismatch exists for every Class.test_* = _test_* assignment in these files, but only manifests as a visible bug when a DecorateInfo skip/xfail targeting that test name is added later.

The trade-off considered: leaving confirmed-broken cases on the helper and the rest on the old pattern creates two inconsistent conventions in the codebase, making it unclear which pattern future contributors should follow. A uniform register_test(...) pattern makes the contract explicit and catches regressions early (as the reviewer comment on the helper itself noted). Given that the overhead per call-site is minimal and the scope of the conversion is limited to the XPU test utilities, standardizing upfront seemed lower-risk than deferring.