maxtext.input_pipeline.multihost_dataloading module#
SPMD Multihost Dataloading Utilities.
Adapted from Sholto’s: sholtodouglas/multihost_dataloading
- class maxtext.input_pipeline.multihost_dataloading.MultiHostDataLoadIterator(dataloader, global_mesh, generate_padding_batch=False, expansion_loading_factor_for_grain=-1)[source]#
Bases:
objectfold get_next_batch_sharded into a iterator class. expansion_factor_for_grain is only used for grain pipeline when having a subset of hosts loading real data.
- Parameters:
dataloader (DatasetV2 | Iterable)
global_mesh (Mesh)
generate_padding_batch (bool)
expansion_loading_factor_for_grain (int)