Implementation of the MEGABYTE architecture that uses hierarchical transformers to model sequences at multiple scales. The global transformer handles coarse representations while local transformers refine finer details. Supports variable depth hierarchies, flash attention, and training/generation on arbitrary byte sequences.