Train a LoRA adapter on the caller's model. Obligations of the control.signal
seam: cancellation is observed only at boundaries (entry, before every optimizer
step, and once more after the last step once every checkpoint write has
completed), so a step is never interrupted and numerics are unchanged; a
cancellation observed at a boundary rejects with the signal's reason, detaches
the caller's model, releases the LoRA leaves after any checkpoint write that
borrowed them has settled, keeps checkpoints saved earlier on disk, and writes
no final adapter. Starting the final save is the commit point: a signal that
arrives after it does not stop the complete save, and the run reports success.
Every exit, including failure, leaves the model detached from training.
Train a LoRA adapter on the caller's model. Obligations of the
control.signalseam: cancellation is observed only at boundaries (entry, before every optimizer step, and once more after the last step once every checkpoint write has completed), so a step is never interrupted and numerics are unchanged; a cancellation observed at a boundary rejects with the signal's reason, detaches the caller's model, releases the LoRA leaves after any checkpoint write that borrowed them has settled, keeps checkpoints saved earlier on disk, and writes no final adapter. Starting the final save is the commit point: a signal that arrives after it does not stop the complete save, and the run reports success. Every exit, including failure, leaves the model detached from training.