fix(checkpointing): Using global step count for checkpoint retention.

BlueCrescent · Copilot · web-flow · commit afc5dbfd1399 · 2026-05-08T16:17:32.000+02:00
Co-authored-by: Copilot Autofix powered by AI &lt;175728472+Copilot@users.noreply.github.com&gt;
diff --git a/src/modalities/checkpointing/checkpoint_saving_strategies.py b/src/modalities/checkpointing/checkpoint_saving_strategies.py
@@ -172,7 +172,7 @@ def get_checkpoint_instruction(
             if len(self._saved_recent_checkpoints) > self._num_recent_checkpoints_to_keep
             else ([], self._saved_recent_checkpoints)
         )
-        # Do not delete checkpoints that are divisible by k.
-        checkpoints_to_delete = [cp for cp in checkpoints_to_delete if cp.num_seen_steps_current_run % self._k != 0]
+        # Do not delete checkpoints that are divisible by k in total training steps.
+        checkpoints_to_delete = [cp for cp in checkpoints_to_delete if cp.num_seen_steps_total % self._k != 0]
 
         return CheckpointingInstruction(save_current=True, checkpoints_to_delete=checkpoints_to_delete)

Original file line number	Diff line number	Diff line change
`@@ -172,7 +172,7 @@ def get_checkpoint_instruction(`
`172`	`172`	`if len(self._saved_recent_checkpoints) > self._num_recent_checkpoints_to_keep`
`173`	`173`	`else ([], self._saved_recent_checkpoints)`
`174`	`174`	`)`
`175`		`- # Do not delete checkpoints that are divisible by k.`
`176`		`- checkpoints_to_delete = [cp for cp in checkpoints_to_delete if cp.num_seen_steps_current_run % self._k != 0]`
	`175`	`+ # Do not delete checkpoints that are divisible by k in total training steps.`
	`176`	`+ checkpoints_to_delete = [cp for cp in checkpoints_to_delete if cp.num_seen_steps_total % self._k != 0]`
`177`	`177`
`178`	`178`	`return CheckpointingInstruction(save_current=True, checkpoints_to_delete=checkpoints_to_delete)`