use Accelerator to prepare distributed training components :param accelerator: Accelerate instance :param components: any num

Validate a latent-world init checkpoint. LaWAM training accepts either a complete LaWAM checkpoint or a VLM-only initialization checkpoint pr

Keep full VLM context values for flow, but split direct-flow gradients by sequence boundary: only tokens after the last act placeholder keep

Args: batch: video: [T, V, H, W, C]

Convert batched videos to uint8 [B,T,H,W,C] once before GPU augmentation.

Convert one clip to uint8 [T,H,W,C] for GPU augmentation.

Apply two-view GPU augmentation for heterogeneous-shape batches.

Build local-step indexing arrays for fast local->absolute index mapping.

Clone ROBOT_TYPE_CONFIG_MAP[robot_type].modality_config() and override delta_indices. We only keep video/state by default to reduce IO for LA

Build deterministic train/val split using trajectory order and tail ratio.

Build AdamW decay/no-decay param groups for LAM training.

calculate global batch size

Use the config to check if the keys are valid and detect silent data corruption.

Duplicate helper (retained for backward compatibility). See primary _check_unnorm_key above.

Parse config booleans explicitly to avoid `bool("false") == True` surprises.

create data iterators

Drop legacy VLM lm_head keys when the current VLA has removed lm_head.

EMA update of codebooks using current assignments.

Fallback path for samples that PyAV fails to decode at a target timestamp.

training end processing