[train][checkpoint] Add ray.train.get_all_reported_checkpoints method (#54555)

# Summary

This PR adds a `ray.train.get_all_reported_checkpoints` method that
allows users to get all the checkpoints they have reported from within
their training function.

This is different from
[Result](https://docs.ray.io/en/latest/train/user-guides/results.html)
in two ways:
* It is called from the training function on the training worker instead
of from the driver
* It can be called while training is still in progress

# Implementation Notes

The main idea is to use a worker-side counter and controller-side
counter as follows:
* Train worker: `ray.train.report` increments a
`num_reported_checkpoints` counter and puts the training result into its
queue
* Train controller: polls the training results from all worker,
registers the checkpoint, increments `num_reported_checkpoints`, and
then creates an asyncio task to notify asyncio Condition. This works
because asyncio Ray actors should always have an event loop.
* Train worker: `get_all_reported_results` uses an asyncio.Condition to
wait until the worker-side `num_reported_checkpoints` counter matches
its controller-side counterpart before returning the checkpoints. This
ensures that we wait for all pending reports to finish. It has access to
the controller actor through `init_train_context`.

`get_checkpoint` should be unaffected because it uses the local
checkpoint; we can consider changing it to use the "centrally committed"
checkpoint in the future.

# Testing

I ran the [ray train pytorch
example](https://docs.ray.io/en/latest/train/getting-started-pytorch.html)
and called `ray.train.get_all_reported_checkpoints` at the end of each
epoch. The results are as expected; here are a few examples

`
epoch 4: [TrainingResult(checkpoint=Checkpoint(filesystem=local,
path=/mnt/cluster_storage/my_run_name/checkpoint_2025-08-04_17-34-52.538994),
metrics={'loss': 0.24510294198989868, 'epoch': 0}),
TrainingResult(checkpoint=Checkpoint(filesystem=local,
path=/mnt/cluster_storage/my_run_name/checkpoint_2025-08-04_17-35-07.511694),
metrics={'loss': 0.23799467086791992, 'epoch': 1}),
TrainingResult(checkpoint=Checkpoint(filesystem=local,
path=/mnt/cluster_storage/my_run_name/checkpoint_2025-08-04_17-35-24.355974),
metrics={'loss': 0.39628422260284424, 'epoch': 2}),
TrainingResult(checkpoint=Checkpoint(filesystem=local,
path=/mnt/cluster_storage/my_run_name/checkpoint_2025-08-04_17-35-40.273211),
metrics={'loss': 0.15193207561969757, 'epoch': 3}),
TrainingResult(checkpoint=Checkpoint(filesystem=local,
path=/mnt/cluster_storage/my_run_name/checkpoint_2025-08-04_17-35-56.178119),
metrics={'loss': 0.17416314780712128, 'epoch': 4})]
`

`
epoch 9: [TrainingResult(checkpoint=Checkpoint(filesystem=local,
path=/mnt/cluster_storage/my_run_name/checkpoint_2025-08-04_17-34-52.538994),
metrics={'loss': 0.24510294198989868, 'epoch': 0}),
TrainingResult(checkpoint=Checkpoint(filesystem=local,
path=/mnt/cluster_storage/my_run_name/checkpoint_2025-08-04_17-35-07.511694),
metrics={'loss': 0.23799467086791992, 'epoch': 1}),
TrainingResult(checkpoint=Checkpoint(filesystem=local,
path=/mnt/cluster_storage/my_run_name/checkpoint_2025-08-04_17-35-24.355974),
metrics={'loss': 0.39628422260284424, 'epoch': 2}),
TrainingResult(checkpoint=Checkpoint(filesystem=local,
path=/mnt/cluster_storage/my_run_name/checkpoint_2025-08-04_17-35-40.273211),
metrics={'loss': 0.15193207561969757, 'epoch': 3}),
TrainingResult(checkpoint=Checkpoint(filesystem=local,
path=/mnt/cluster_storage/my_run_name/checkpoint_2025-08-04_17-35-56.178119),
metrics={'loss': 0.17416314780712128, 'epoch': 4}),
TrainingResult(checkpoint=Checkpoint(filesystem=local,
path=/mnt/cluster_storage/my_run_name/checkpoint_2025-08-04_17-36-12.547310),
metrics={'loss': 0.2924661934375763, 'epoch': 5}),
TrainingResult(checkpoint=Checkpoint(filesystem=local,
path=/mnt/cluster_storage/my_run_name/checkpoint_2025-08-04_17-36-28.538090),
metrics={'loss': 0.18640762567520142, 'epoch': 6}),
TrainingResult(checkpoint=Checkpoint(filesystem=local,
path=/mnt/cluster_storage/my_run_name/checkpoint_2025-08-04_17-36-44.583228),
metrics={'loss': 0.12567029893398285, 'epoch': 7}),
TrainingResult(checkpoint=Checkpoint(filesystem=local,
path=/mnt/cluster_storage/my_run_name/checkpoint_2025-08-04_17-37-00.540405),
metrics={'loss': 0.1620682030916214, 'epoch': 8}),
TrainingResult(checkpoint=Checkpoint(filesystem=local,
path=/mnt/cluster_storage/my_run_name/checkpoint_2025-08-04_17-37-17.129973),
metrics={'loss': 0.07022886723279953, 'epoch': 9})]
`

I also modified all the Ray Train v2 unit tests that call
`ray.train.report`:
* `test_persistence` also verifies that `get_all_reported_checkpoints`
works on resumption
* `test_data_parallel_trainer` verifies that
`get_all_reported_checkpoints` stalls until all workers report.

I also verified that `get_all_reported_checkpoints` produced similar
output when called from Tune + Train.

I tried to test that `get_all_reported_checkpoints` finished even with
graceful abort but was unable to create such a scenario since
`get_all_reported_checkpoints` returns very quickly and each `report`
forms a barrier.

---------

Signed-off-by: Timothy Seah <tseah@anyscale.com>

Author: TimothySeah

ci/lint/pydoclint-baseline.txt

@@ -2006,14 +2006,6 @@ python/ray/train/v2/_internal/callbacks/accelerators.py
 python/ray/train/v2/_internal/execution/checkpoint/checkpoint_manager.py
     DOC103: Method `CheckpointManager.register_checkpoint`: Docstring arguments are different from function arguments. (Or could be other formatting issues: https://jsh9.github.io/pydoclint/violation_codes.html#notes-on-doc103 ). Arguments in the function signature but not in the docstring: [checkpoint_result: _TrainingResult]. Arguments in the docstring but not in the function signature: [checkpoint: ].
 --------------------
-python/ray/train/v2/_internal/execution/context.py
-    DOC101: Method `TrainContext._save_checkpoint`: Docstring contains fewer arguments than in function signature.
-    DOC103: Method `TrainContext._save_checkpoint`: Docstring arguments are different from function arguments. (Or could be other formatting issues: https://jsh9.github.io/pydoclint/violation_codes.html#notes-on-doc103 ). Arguments in the function signature but not in the docstring: [checkpoint: Optional[Checkpoint], checkpoint_dir_name: str, metrics: Dict[str, Any]].
---------------------
-python/ray/train/v2/_internal/execution/controller/controller.py
-    DOC101: Method `TrainController._start_worker_group`: Docstring contains fewer arguments than in function signature.
-    DOC103: Method `TrainController._start_worker_group`: Docstring arguments are different from function arguments. (Or could be other formatting issues: https://jsh9.github.io/pydoclint/violation_codes.html#notes-on-doc103 ). Arguments in the function signature but not in the docstring: [num_workers: int, resources_per_worker: dict].
---------------------
 python/ray/train/v2/_internal/execution/storage.py
     DOC101: Method `_ExcludingLocalFilesystem.__init__`: Docstring contains fewer arguments than in function signature.
     DOC103: Method `_ExcludingLocalFilesystem.__init__`: Docstring arguments are different from function arguments. (Or could be other formatting issues: https://jsh9.github.io/pydoclint/violation_codes.html#notes-on-doc103 ). Arguments in the function signature but not in the docstring: [**kwargs: ].

python/ray/train/__init__.py

@@ -34,8 +34,10 @@
         RunConfig,
         ScalingConfig,
     )
+    from ray.train.v2.api.reported_checkpoint import ReportedCheckpoint  # noqa: F811
     from ray.train.v2.api.result import Result  # noqa: F811
     from ray.train.v2.api.train_fn_utils import (  # noqa: F811
+        get_all_reported_checkpoints,
         get_checkpoint,
         get_context,
         get_dataset_shard,
@@ -76,9 +78,14 @@
 SyncConfig.__module__ = "ray.train"
 TrainingIterator.__module__ = "ray.train"
 
+# TODO: consider implementing these in v1 and raising ImportError instead.
 if is_v2_enabled():
     __all__.append("UserCallback")
     UserCallback.__module__ = "ray.train"
+    __all__.append("get_all_reported_checkpoints")
+    get_all_reported_checkpoints.__module__ = "ray.train"
+    __all__.append("ReportedCheckpoint")
+    ReportedCheckpoint.__module__ = "ray.train"
 
 
 # DO NOT ADD ANYTHING AFTER THIS LINE.

python/ray/train/v2/_internal/execution/checkpoint/checkpoint_manager.py

@@ -1,3 +1,4 @@
+import asyncio
 import logging
 from typing import Any, Dict, List, Optional
 
@@ -16,6 +17,7 @@
 from ray.train.v2._internal.execution.context import StorageContext
 from ray.train.v2._internal.execution.storage import _delete_fs_path, _exists_at_fs_path
 from ray.train.v2._internal.execution.worker_group import Worker
+from ray.train.v2.api.reported_checkpoint import ReportedCheckpoint
 
 try:
     from pydantic import BaseModel
@@ -81,6 +83,12 @@ def __init__(
     ):
         self._storage_context = storage_context
         self._checkpoint_config = checkpoint_config
+
+        # This tracks the number of report calls that have been processed
+        # for the current worker group.
+        self._num_report_calls = 0
+
+        self._condition = asyncio.Condition()
         super().__init__(checkpoint_config)
         # If the snapshot is found, the checkpoint manager will restore its state.
         self._maybe_load_state_from_storage()
@@ -139,6 +147,14 @@ def register_checkpoint(self, checkpoint_result: _TrainingResult):
             logger.debug("Deleting checkpoint: ", checkpoint)
             _delete_fs_path(fs=checkpoint.filesystem, fs_path=checkpoint.path)
 
+        self._num_report_calls += 1
+
+        async def async_notify():
+            async with self._condition:
+                self._condition.notify_all()
+
+        asyncio.create_task(async_notify())
+
     # --------------------------
     # CheckpointManager state
     # --------------------------
@@ -267,6 +283,7 @@ def after_report(
         self, metrics: List[Dict[str, Any]], checkpoint: Optional[Checkpoint]
     ):
         if not checkpoint:
+            self._num_report_calls += 1
             return
 
         rank_0_metrics = metrics[0]
@@ -279,9 +296,31 @@ def after_report(
     # --------------------------
 
     def before_init_train_context(self, workers: List[Worker]) -> Dict[str, List[Any]]:
+        self._num_report_calls = 0
         latest_checkpoint = (
             self.latest_checkpoint_result.checkpoint
             if self.latest_checkpoint_result
             else None
         )
-        return {"checkpoint": [latest_checkpoint] * len(workers)}
+        train_context_args = {
+            "checkpoint": [latest_checkpoint] * len(workers),
+        }
+        return train_context_args
+
+    async def get_all_reported_checkpoints(
+        self, expected_num_report_calls: int
+    ) -> List[ReportedCheckpoint]:
+        """Once expected_num_checkpoints are reported, return the ReportedCheckpoints."""
+        async with self._condition:
+            await self._condition.wait_for(
+                lambda: self._num_report_calls == expected_num_report_calls
+            )
+            # TODO: might be nice for CheckpointManager to manage ReportedCheckpoint
+            # instead of _TrainingResult but that is a large refactor.
+            return [
+                ReportedCheckpoint(
+                    checkpoint=tr.checkpoint,
+                    metrics=tr.metrics,
+                )
+                for tr in self._checkpoint_results
+            ]

python/ray/train/v2/_internal/execution/context.py

@@ -7,8 +7,8 @@
 from typing import TYPE_CHECKING, Any, Dict, List, Optional
 
 import ray
+from ray.actor import ActorHandle
 from ray.data import DataIterator, Dataset
-from ray.train import BackendConfig, Checkpoint, DataConfig
 from ray.train._internal import session
 from ray.train._internal.session import _TrainingResult
 from ray.train.v2._internal.execution.checkpoint.sync_actor import SynchronizationActor
@@ -17,12 +17,14 @@
 from ray.train.v2.api.config import RunConfig, ScalingConfig
 
 if TYPE_CHECKING:
+    from ray.train import BackendConfig, Checkpoint, DataConfig
     from ray.train.v2._internal.data_integration.interfaces import (
         DatasetShardMetadata,
         DatasetShardProvider,
     )
     from ray.train.v2._internal.execution.callback import TrainContextCallback
     from ray.train.v2._internal.execution.worker_group.thread_runner import ThreadRunner
+    from ray.train.v2.api.reported_checkpoint import ReportedCheckpoint
 
 
 logger = logging.getLogger(__file__)
@@ -45,13 +47,13 @@ class TrainRunContext:
     scaling_config: ScalingConfig
 
     # The configuration for the training backend (e.g., PyTorch, XGBoost).
-    backend_config: BackendConfig
+    backend_config: "BackendConfig"
 
     # The datasets used in the current training run.
     datasets: Dict[str, Dataset]
 
     # The configuration for dataset ingestion and sharding.
-    dataset_config: DataConfig
+    dataset_config: "DataConfig"
 
     def get_run_config(self) -> RunConfig:
         """Returns the run config of the current training run."""
@@ -96,8 +98,11 @@ class TrainContext:
     distributed_context: DistributedContext
     execution_context: ExecutionContext
     storage_context: StorageContext
+    controller_actor: ActorHandle
+
     dataset_shard_provider: "DatasetShardProvider"
-    checkpoint: Optional[Checkpoint] = None
+    checkpoint: Optional["Checkpoint"] = None
+    num_report_calls: int = 0
 
     @_copy_doc(session.get_experiment_name)
     def get_experiment_name(self) -> str:
@@ -137,6 +142,13 @@ def get_synchronization_actor(self):
     def get_checkpoint(self):
         return self.checkpoint
 
+    def get_all_reported_checkpoints(self) -> List["ReportedCheckpoint"]:
+        return ray.get(
+            self.controller_actor.get_all_reported_checkpoints.remote(
+                self.num_report_calls
+            )
+        )
+
     def get_dataset_shard(self, dataset_info: "DatasetShardMetadata") -> DataIterator:
         """Returns the :class:`ray.data.DataIterator` shard for this worker.
 
@@ -189,10 +201,15 @@ def _save_checkpoint(
         self,
         checkpoint_dir_name: str,
         metrics: Dict[str, Any],
-        checkpoint: Optional[Checkpoint] = None,
+        checkpoint: Optional["Checkpoint"] = None,
     ) -> _TrainingResult:
         """Save the checkpoint to remote storage.
 
+        Args:
+            checkpoint_dir_name: The checkpoint dir to persist to.
+            metrics: The metrics to report.
+            checkpoint: The checkpoint to report.
+
         Returns:
             The training result object containing the persisted checkpoint.
         """
@@ -212,7 +229,7 @@ def _save_checkpoint(
     def report(
         self,
         metrics: Dict[str, Any],
-        checkpoint: Optional[Checkpoint] = None,
+        checkpoint: Optional["Checkpoint"] = None,
         checkpoint_dir_name: Optional[str] = None,
     ) -> None:
         """
@@ -265,6 +282,7 @@ def report(
             # TODO (hpguo): Add a metrics to track the blocking time waiting for the
             # training result to be consumed by the controller.
             self.get_result_queue().put(training_result)
+            self.num_report_calls += 1
 
 
 # The global variable holding the current TrainContext

python/ray/train/v2/_internal/execution/controller/controller.py

@@ -3,7 +3,7 @@
 import os
 import uuid
 from dataclasses import dataclass
-from typing import Callable, List, Optional, Union
+from typing import TYPE_CHECKING, Callable, List, Optional, Union
 
 import pandas as pd
 
@@ -67,6 +67,10 @@
 )
 from ray.train.v2.api.result import Result
 
+if TYPE_CHECKING:
+    from ray.train.v2.api.reported_checkpoint import ReportedCheckpoint
+
+
 logger = logging.getLogger(__name__)
 
 
@@ -275,6 +279,10 @@ def _start_worker_group(
     ) -> Optional[ControllerError]:
         """Start the worker group and launch the train function.
 
+        Args:
+            num_workers: The number of workers to start.
+            resources_per_worker: The resources per worker to start.
+
         Returns:
             None if the worker group was successfully started,
             ControllerError if the worker group failed to start.
@@ -537,7 +545,6 @@ def get_result(self) -> Result:
             raise ValueError(
                 f"Cannot get result when controller is in state {controller_state}"
             )
-
         return self._build_result()
 
     def get_training_failed_error(self) -> Optional[TrainingFailedError]:
@@ -553,3 +560,10 @@ def get_training_failed_error(self) -> Optional[TrainingFailedError]:
             return controller_state.training_failed_error
 
         return None
+
+    async def get_all_reported_checkpoints(
+        self, expected_num_report_calls: int
+    ) -> List["ReportedCheckpoint"]:
+        return await self._checkpoint_manager.get_all_reported_checkpoints(
+            expected_num_report_calls
+        )

python/ray/train/v2/_internal/execution/train_fn_utils.py

@@ -1,14 +1,17 @@
 import threading
-from typing import Any, Dict, Optional
+from typing import TYPE_CHECKING, Any, Dict, List, Optional
 
 from ray.data import DataIterator
-from ray.train import Checkpoint
 from ray.train.v2._internal.execution import collective_impl
 from ray.train.v2._internal.execution.context import (
     get_train_context as get_internal_train_context,
 )
 from ray.train.v2.api.context import TrainContext as ExternalTrainContext
 
+if TYPE_CHECKING:
+    from ray.train import Checkpoint
+    from ray.train.v2.api.reported_checkpoint import ReportedCheckpoint
+
 
 class TrainFnUtils:
     """Utility class providing an abstraction layer between user-facing APIs
@@ -21,7 +24,7 @@ class TrainFnUtils:
     def report(
         self,
         metrics: Dict[str, Any],
-        checkpoint: Optional[Checkpoint] = None,
+        checkpoint: Optional["Checkpoint"] = None,
         checkpoint_dir_name: Optional[str] = None,
     ) -> None:
         """Upload checkpoint to remote storage and put a training result on the result queue.
@@ -46,6 +49,15 @@ def get_checkpoint(self):
         """
         return get_internal_train_context().get_checkpoint()
 
+    def get_all_reported_checkpoints(self) -> List["ReportedCheckpoint"]:
+        """Get all the checkpoints reported by the workers.
+
+        Returns:
+            A list of ReportedCheckpoint objects that represent the checkpoints and
+            corresponding metrics reported by the workers.
+        """
+        return get_internal_train_context().get_all_reported_checkpoints()
+
     def get_dataset_shard(self, dataset_name: str) -> DataIterator:
         """Get the dataset shard for this worker.
 

python/ray/train/v2/_internal/execution/worker_group/worker.py

@@ -194,6 +194,7 @@ def init_train_context(
         synchronization_actor: SynchronizationActor,
         storage_context: StorageContext,
         worker_callbacks: List[Union[WorkerCallback, TrainContextCallback]],
+        controller_actor: ActorHandle,
         dataset_shard_provider: Optional["DatasetShardProvider"] = None,
         checkpoint: Optional[Checkpoint] = None,
     ):
@@ -213,6 +214,7 @@ def init_train_context(
                 train_context_callbacks=context_callbacks_to_propagate,
             ),
             storage_context=storage_context,
+            controller_actor=controller_actor,
             checkpoint=checkpoint,
             dataset_shard_provider=dataset_shard_provider,
         )

python/ray/train/v2/_internal/execution/worker_group/worker_group.py

@@ -437,6 +437,7 @@ def _init_train_context_on_workers(
                 synchronization_actor=sync_actor,
                 storage_context=self._storage_context,
                 worker_callbacks=self._worker_callbacks_to_propagate,
+                controller_actor=ray.get_runtime_context().current_actor,
                 **{
                     arg: arg_values[i] for arg, arg_values in train_context_args.items()
                 },

python/ray/train/v2/api/reported_checkpoint.py

@@ -0,0 +1,21 @@
+from dataclasses import dataclass
+from typing import TYPE_CHECKING, Any, Dict
+
+from ray.util.annotations import PublicAPI
+
+if TYPE_CHECKING:
+    from ray.train import Checkpoint
+
+
+@dataclass
+@PublicAPI(stability="alpha")
+class ReportedCheckpoint:
+    """A user-reported checkpoint and its associated metrics.
+
+    Attributes:
+        checkpoint: The checkpoint reported by the user.
+        metrics: The metrics associated with that checkpoint.
+    """
+
+    checkpoint: "Checkpoint"
+    metrics: Dict[str, Any]