testing

Recipe YAMLs for Megatron-LM functional test buckets are defined using structured workload definitions, expanded products combinations, and runtime placeholders for assets and artifacts to ensure deterministic parsing and execution.

You can reproduce Megatron-LM CI test failures locally by matching CI behavior using torch.distributed.run and CI scripts, alongside downloading and updating golden outputs to diagnose regressions or mismatches.

The best way to manage flaky Megatron-LM distributed tests is applying stability controls like marker filters, CI-parity guidance, and environment-specific scopes to handle flaky and disabled test behavior reliably.

Pytest works with distributed testing in Megatron-LM unit tests by running via distributed pytest layouts, while functional tests run end-to-end using CI shell scripts and artifacts for consistent execution.

Yes, you can use cartesian workload expansion for H100 and GB200 tests by defining structured recipe YAMLs that expand products combinations and apply runtime placeholders for distributed artifacts and logging.

Golden values mismatch locally because of missing golden-value download workflows or environment drift; you must download and update golden outputs and match CI behavior exactly to diagnose regressions accurately.