Fix test helper discovery in pytest and sanitizer runner

[lucifer1004]

Summary

• Mark the test_filter decorator factory with pytest's existing __test__ = False opt-out convention.
• Make the sanitizer runner respect the same marker instead of excluding names containing test_filter.

Why

Running an entire test module through pytest can collect the imported test_filter(condition) helper and fail with fixture 'condition' not found. Direct script execution avoids this, while the sanitizer runner currently has a name-specific exclusion. The shared marker handles aliases and explicit opt-outs without accidentally excluding real tests such as test_filter_behavior.

The marker applies only to the decorator factory. Decorated test functions remain discoverable, and existing condition-based execution/filtering behavior is unchanged.

Validation

Used a temporary CPU-only verification harness (not included in this PR):

• Actual pytest collection/execution and default sanitizer discovery checks: 2 passed with the fix; 2 failed as expected against unmodified upstream 78b6900.
• Confirmed the sanitizer-generated Python payloads execute successfully and temporary mocks restore original functions.
• Confirmed pytest's collection hook checks getattr(obj, "__test__", True) directly.
• git diff --check: passed.

No GPU kernels, actual Compute Sanitizer execution, or native rebuild were required. This small PR is independent of the SM120/API migration work.

[ds-review-bot]

🔵 suggestion: Consider adding a one-line comment explaining why test_filter.__test__ = False is set (pytest / test_sanitizer.py opt-out marker for a test-prefixed helper). Future readers may otherwise be tempted to remove what looks like a stray attribute assignment.

_{🤖 v5}

[ds-review-bot]

🔵 suggestion: With default discovery, the new tests/test_test_discovery.py module (test_pytest_discovery, test_sanitizer_discovery) will now itself be picked up and run under compute-sanitizer for every tool. These are pure-Python CPU tests that spawn pytest subprocesses and mock subprocess.run, so running them under the sanitizer is wasted time (and test_sanitizer_discovery re-executes this runner via runpy inside the sanitized process). Setting __test__ = False on them is not appropriate since pytest should collect them, so adding 'test_test_discovery.py' to exclude_files is the right lever. Not blocking.

_{🤖 v5}

[lucifer1004]

The whole test has been removed.

[ds-review-bot]

🔵 suggestion: The 60s timeout for the pytest subprocess may be tight: the sample module imports torch (via deep_gemm/testing/utils.py) in a fresh interpreter twice (collect-only + run). On cold caches or slow CI runners torch import alone can take 10–30s, so this could be flaky. Consider raising the timeout (e.g. 300s); a hang is still caught, just later.

_{🤖 v5}

[lucifer1004]

The whole test has been removed.

[ds-review-bot]

🔵 suggestion: patch.object(os, 'listdir', ...), patch.object(importlib, 'import_module', ...) and patch.object(subprocess, 'run', ...) are global patches held while runpy.run_path executes the runner. This works today because runpy/argparse don't call these, but if the runner later grows other imports or filesystem calls this test could break in confusing ways. A short comment noting the assumption (and that import_module returns the same sample module for any name) would help maintainers.

_{🤖 v5}

[lucifer1004]

The whole test has been removed.

[ds-review-bot]

🤖 ds-review-bot Code Review

v6

⚠️ 未完成评审（upstream_error：模型上游服务不可用）

v5

This MR fixes accidental collection of the test_filter decorator factory by both pytest and the sanitizer runner. The change is small (+93/−1 across 3 files), correct, and well-targeted: test_filter.__test__ = False in deep_gemm/testing/utils.py opts the factory out of pytest collection without affecting decorated tests (functools.wraps copies __dict__ from the wrapped test function, not from test_filter), and tests/test_sanitizer.py now checks getattr(obj, '__test__', True) instead of the fragile 'test_filter' not in name substring test, which correctly handles aliases and explicit opt-outs while no longer dropping legitimate tests like test_filter_behavior. The two new CPU-only regression tests in tests/test_test_discovery.py exercise real pytest collection/execution in a subprocess and the sanitizer runner's default discovery via runpy with mocks; the sample module covers enabled/filtered/alias/test_filter-named/explicit-opt-out cases, which directly demonstrates both regressions. All existing @test_filter(...) call sites (test_einsum, test_hyperconnection, test_mega_mhc, test_mega_gate, test_attention) use it only as a decorator, so runtime behavior is unchanged. The assert_called_once_with on import_module implicitly verifies exclude_files filtering too. git diff --check is clean. Note: the sandbox shell was restricted to git commands, so the pytest run reported in the MR description could not be re-executed here; the review is based on code inspection. Recommendation: approve, with a few minor optional suggestions below.

v4

本 MR 修复了测试助手被 pytest / sanitizer runner 误收集的问题。deep_gemm/testing/utils.py 为装饰器工厂 test_filter 增加 __test__ = False，使导入它的测试模块在 pytest 收集时不再把该工厂当作测试（此前会因缺少 condition fixture 而收集失败）；tests/test_sanitizer.py 将默认发现逻辑从「按名称排除 test_filter」改为 getattr(obj, '__test__', True)，既兼容 test_filter 的别名/重导出，又避免误伤 test_filter_behavior 这类真实测试，并与 pytest 的 opt-out 语义保持一致。新增的 tests/test_test_discovery.py 用两个 CPU-only 回归测试分别覆盖 pytest 的实际收集/执行（含 __test__ = False 显式 opt-out 与别名）以及 sanitizer runner 的默认发现与命令生成。

标记只作用于装饰器工厂，被装饰的测试函数仍可被发现，条件过滤行为不变；实现简洁、与现有模式一致。未发现正确性、安全性、性能或行为回归问题，建议合并。

Files reviewed: 3
Issues found: 🔵 4 suggestion
Inline comments posted: 4

⚠️ Parse warning: [v6] upstream_error：模型上游服务不可用