Smart Batch Composition¶

Flux uses smart batching to optimize training efficiency.

Strategies¶

1. Length Bucketing¶

Groups similar-length sequences to minimize padding.

2. Staleness Balancing¶

Stratified sampling ensures balanced staleness distribution.

3. Curriculum Learning¶

Progressive difficulty ordering (easy → hard).

Configuration¶

batch_composer:
  length_bucket_boundaries: [256, 512, 1024, 2048]
  staleness_balance_weight: 0.3
  curriculum_enabled: true
  curriculum_decay: 0.995