add-reference-tests

Generate pytest scaffolds for flashinfer_trace by targeting a specific op_type or all definitions under flashinfer_trace/definitions, placing new tests in flashinfer_trace/tests/references/ while skipping existing ones.

Reference testing validates flashinfer_trace definitions against FlashInfer or SGLang ground-truth kernels by verifying constants, input shapes, and numerical correctness across multiple sources using configurable tolerances and dtypes.

Yes, you must complete the clone-repos step before generating reference tests, because the validation process requires accessing ground-truth kernels from FlashInfer or SGLang to compare against your local definitions.

Yes, you can test all relevant definitions at once by running the command without a specific op_type filter, which discovers all definitions under flashinfer_trace/definitions and generates corresponding pytest files for each.

No, generating new reference tests skips existing files in flashinfer_trace/tests/references/ and only creates scaffolds for definitions that lack a corresponding test file, preserving your previously written or modified tests.

The best way to validate numerical correctness is using pytest-based reference tests that compare your definitions against FlashInfer or SGLang ground truth with configurable tolerances, shapes, and dtypes to ensure kernel implementation alignment.