packed_memmap#

Read-only fixed-sequence token storage shared by every distributed stage.

Classes

PackedTokenMemmapDataset

Map N x (S+1) uint32 tokens to shifted input/target tensors.

class PackedTokenMemmapDataset#

Bases: Dataset

Map N x (S+1) uint32 tokens to shifted input/target tensors.

__init__(path, *, limit=None, sequence_length=None)#
Parameters:
  • path (str | Path)

  • limit (int | None)

  • sequence_length (int | None)