lrm/lrm_sana/docs/checklistv2.md

SANA LRM Checklist V2

1. Scope and Dependency Rules

• Core parity scope is trainer architecture and runtime contracts across flux, lrm_15, lrm_xl, and lrm_sana.
• information_related_to_sana is treated as model-specific external adapter code.
• information_related_to_flux is treated as model-specific external adapter code.
• External model adapter modules are excluded from core trainer parity checks.

2. Codebase Parity Against Existing Variants

• SANA has expected variant-specific replacements for config/task/model/dataset/criterion modules.
• SANA keeps shared trainer scaffolding structure aligned with other variants.
• Launcher flow preserves profile-based behavior and runtime overrides.
• Optimizer/lr-scheduler defaults are set for runnable quick smoke behavior.
• Quick profile precision default updated to BF16 for stable H200 smoke runs.

3. Runtime Verification Evidence (centralized in log.log)

• Comparison inventory and focused diffs recorded in log output.
• Hydra composition succeeds (Hydra compose bg exit code: 0).
• Deterministic quick smoke run reaches final evaluation end marker (EVAL END (final@gstep=1)).
• Deterministic quick smoke run saves checkpoint (checkpoint-gstep1).
• All verification command outputs are consolidated under lrm_sana/log.log.

4. Remaining Validation Items

• Initialization smoke checks for all four SANA model profiles.
• Optional distributed_sharded path validation on nodes with CUDA toolchain available for DeepSpeed ops.
• Extended multi-step sanity run (beyond max_steps=1).

5. Canonical Artifacts

• Main aggregated run/comparison log: lrm_sana/log.log
• Previous checklist: lrm_sana/docs/checklist.md
• This checklist revision: lrm_sana/docs/checklistv2.md