[xieminghui187]

Problem

The definition loader keeps module-level assignments only when their RHS is a literal. Valid Triton globals such as LOG2_E = tl.constexpr(1.4426950408889634) and NEG_INF = tl.constexpr(float("-inf")) are silently discarded. Kernels referencing them then fail with undefined-name errors before an effective accuracy comparison.

Changes

• Resolve Triton constexpr constructors from imports, including aliases.
• Preserve literal-valued constexpr assignments and annotated assignments, including literal float(...) sentinels.
• Preserve source order without special-casing contestant constant names.
• Continue filtering arbitrary call assignments and executable top-level statements.
• Leave output structure/type checks, tolerances, timing, and device handling unchanged.

This remains a narrow definition-loading policy, not a security sandbox: imports, decorators, and class bodies already execute.

Validation

python -m unittest discover -s tests -p test_ks_auto_bench_loader.py -v
git diff --check

All 6 CPU-only regression tests pass. PyTorch/Triton dependencies are stubbed, while the actual evaluator loader is exercised. Coverage includes constant visibility in a kernel, aliases, annotated assignments, arbitrary constant names, source ordering, rejected arbitrary calls, unrelated constexpr methods, and unchanged strict list/Tensor rejection.

The regression case also reproduces NameError: name 'LOG2_E' is not defined with the original upstream filter. Uploaded files were fetched back and verified byte-for-byte against the locally tested files.

No accelerator end-to-end benchmark has been run with this generalized patch. This PR does not change submissions or determine competition scores.

[CLAassistant]

All committers have signed the CLA.