utils.callbacks.dynamic_checkpoint

utils.callbacks.dynamic_checkpoint

Classes

Name	Description
DynamicCheckpointCallback	Callback to save checkpoints on-demand during training via:

utils.callbacks.dynamic_checkpoint.DynamicCheckpointCallback(cfg)

Callback to save checkpoints on-demand during training via: 1. File-based trigger (works everywhere, rank 0 checks file)

Thread-safe for multi-GPU distributed training.

touch /path/to/output_dir/axolotl_checkpoint.save

Name	Description
on_step_end	Check for checkpoint triggers at the end of each step.

utils.callbacks.dynamic_checkpoint.DynamicCheckpointCallback.on_step_end(
    args,
    state,
    control,
    **_kwargs,
)

Check for checkpoint triggers at the end of each step. ONLY rank 0 checks the file, then all ranks synchronize.