- CheckpointConfig.start_step: suppress saves before a given step
- TrainingMethod.apply_configured_lrs(): re-apply YAML LRs after a DCP
resume restores the checkpoint's optimizer/scheduler LRs
- tracker: if online W&B init fails at boot, retry offline, then run
trackerless instead of crashing a multi-node job
- dcp_to_diffusers: --weights-only (skip optimizer state that OOMs a
single-GPU conversion) and robust base-pipeline copytree
- GB200 sbatch: lustre HOME, NCCL/MNNVL env pins, boot watchdog,
SP/HSDP topology env overrides; plus the 32-GPU test configs and the
64-prompt held-out H3 validation set
(dataclass/parse plumbing for the new checkpoint fields lands with the
next commit, which touches the same config files)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
