[Train] Fix off-by-one AIR Trainer checkpoint ID indexing on restore (#…

…31423)

This PR is a follow-up to #31231 to save checkpoints to the correctly indexed directory upon restore. The "latest checkpoint ID" that's used to generate the next checkpoint directory (`checkpoint_0000<latest_checkpoint_id>`) is off by one when restoring an AIR trainer.

Signed-off-by: Justin Yu <justinvyu@berkeley.edu>

python/ray/train/_internal/checkpoint.py

@@ -214,7 +214,12 @@ def _load_checkpoint(
     ) -> Optional[Union[Dict, Checkpoint]]:
         loaded_checkpoint = super()._load_checkpoint(checkpoint_to_load)
         assert not loaded_checkpoint or isinstance(loaded_checkpoint, Checkpoint)
-        self._latest_checkpoint_id = getattr(loaded_checkpoint, TUNE_CHECKPOINT_ID, 0)
+        # `latest_checkpoint_id` will be the id assigned to the next checkpoint,
+        # which should be one more than the loaded checkpoint's id
+        # If no checkpoint is loaded, initialize this to 0
+        self._latest_checkpoint_id = (
+            getattr(loaded_checkpoint, TUNE_CHECKPOINT_ID, -1) + 1
+        )
         return loaded_checkpoint
 
     def add_tune_checkpoint_id(self, checkpoint: Checkpoint):

python/ray/tune/tests/test_tuner_restore.py

@@ -10,8 +10,16 @@
 
 import ray
 from ray import tune
-from ray.air import Checkpoint, CheckpointConfig, FailureConfig, RunConfig, session
+from ray.air import (
+    Checkpoint,
+    CheckpointConfig,
+    FailureConfig,
+    RunConfig,
+    ScalingConfig,
+    session,
+)
 from ray.air._internal.remote_storage import delete_at_uri, download_from_uri
+from ray.train.data_parallel_trainer import DataParallelTrainer
 from ray.tune import Callback, Trainable
 from ray.tune.execution.trial_runner import _find_newest_experiment_checkpoint
 from ray.tune.experiment import Trial
@@ -803,7 +811,8 @@ def on_trial_result(self, runner, trial, result):
     )
 
 
-def test_checkpoints_saved_after_resume(tmp_path):
+@pytest.mark.parametrize("use_air_trainer", [True, False])
+def test_checkpoints_saved_after_resume(tmp_path, use_air_trainer):
     """Checkpoints saved after experiment restore should pick up at the correct
     iteration and should not overwrite the checkpoints from the original run.
     Old checkpoints should still be deleted if the total number of checkpoints
@@ -831,19 +840,30 @@ def get_checkpoints(experiment_dir):
     fail_marker = tmp_path / "fail_marker"
     fail_marker.write_text("", encoding="utf-8")
 
+    trainable = (
+        DataParallelTrainer(
+            _train_fn_sometimes_failing, scaling_config=ScalingConfig(num_workers=1)
+        )
+        if use_air_trainer
+        else _train_fn_sometimes_failing
+    )
+    param_space = {
+        "failing_hanging": (fail_marker, None),
+        "num_epochs": 2,
+    }
+    if use_air_trainer:
+        param_space = {"train_loop_config": param_space}
+
     num_to_keep = 4
     tuner = Tuner(
-        _train_fn_sometimes_failing,
+        trainable,
         tune_config=TuneConfig(num_samples=1),
         run_config=RunConfig(
             name="exp_name",
             local_dir=str(tmp_path),
             checkpoint_config=CheckpointConfig(num_to_keep=num_to_keep),
         ),
-        param_space={
-            "failing_hanging": (fail_marker, None),
-            "num_epochs": 2,
-        },
+        param_space=param_space,
     )
     results = tuner.fit()
     training_iteration = results[0].metrics["training_iteration"]