Skip to main content

Module checkpointer

Module checkpointer 

Source
Expand description

TensorCheckpointer — periodic checkpointing of tensor computation state with rollback.

Provides a configurable checkpoint manager that snapshots tensor values at regular intervals (or on demand), maintains a bounded rolling window of checkpoints, and supports rollback to any previously saved state.

Structs§

Checkpoint
A single snapshot of tensor computation state.
CheckpointConfig
Configuration for the TensorCheckpointer.
CheckpointerStats
Aggregate statistics about the checkpointer state.
TensorCheckpointer
Manages a bounded, rolling window of tensor computation checkpoints with rollback support.