Support async checkpointing through CheckpointManager (pytorch#5697)

* Support async checkpointing through CheckpointManager

* Allow threads to exit when CheckpointManager is freed

* Use rank from tracked process group

* Add TODO

test/spmd/test_xla_distributed_checkpoint.py

@@ -4,6 +4,7 @@
 import tempfile
 import unittest
 import test_xla_sharding_base
+import threading
 
 import torch
 import torch.distributed as dist
@@ -412,6 +413,70 @@ def test_manager_max_to_keep(self, tmpdir):
     self.assertTrue(chkpt_mgr.save(20, state_dict))
     self.assertEqual(set(chkpt_mgr.all_steps()), {20, 10})
 
+  @run_with_tmpdir
+  def test_manager_async(self, tmpdir):
+    chkpt_mgr = CheckpointManager(tmpdir, save_interval=10)
+    state_dict = self._get_sharded_model().state_dict()
+
+    # Patch the manager's save method to block until this thread signals.
+    cond = threading.Condition()
+    old_save = chkpt_mgr.save
+
+    def patched_save(*args, **kwargs):
+      cond.wait()
+      old_save(*args, **kwargs)
+
+    with unittest.mock.patch.object(chkpt_mgr, 'save', patched_save):
+      chkpt_mgr.save_async(10, state_dict)
+
+    # No new steps should be tracked immediately after calling save_async
+    self.assertEqual(chkpt_mgr.all_steps(), [])
+
+    # Trigger the actual checkpoint in the background thread and wait for
+    # completion.
+    with cond:
+      cond.notify()
+    chkpt_mgr.join()
+
+    # The manager should track all steps which were asynchronously saved.
+    self.assertEqual(set(chkpt_mgr.all_steps()), {10})
+
+  @run_with_tmpdir
+  def test_manager_async_step_tracking(self, tmpdir):
+    chkpt_mgr = CheckpointManager(tmpdir, save_interval=10)
+    state_dict = self._get_sharded_model().state_dict()
+
+    self.assertEqual(chkpt_mgr.all_steps(), [])
+
+    # Steps not divisible by 10 should not be saved
+    for step in range(1, 10):
+      self.assertFalse(chkpt_mgr.save_async(step, state_dict))
+      self.assertEqual(chkpt_mgr.all_steps(), [])
+
+    # Steps divisible by 10 should be saved
+    saved = set()
+    for step in range(0, 100, 10):
+      self.assertTrue(chkpt_mgr.save_async(step, state_dict))
+      saved.add(step)
+
+    # Join to allow pending async checkpoints to complete
+    chkpt_mgr.join()
+
+    # The manager should track all steps which were asynchronously saved.
+    self.assertEqual(set(chkpt_mgr.all_steps()), saved)
+
+    # Load a checkpoint into a new state_dict
+    new_state_dict = self._get_sharded_model().state_dict()
+    self.assertFalse(
+        any(
+            torch.allclose(v, new_state_dict[k])
+            for k, v in state_dict.items()))
+    chkpt_mgr.restore(0, new_state_dict)
+    self.assertTrue(
+        all(
+            torch.allclose(v, new_state_dict[k])
+            for k, v in state_dict.items()))
+
 
 if __name__ == '__main__':
   test = unittest.main()

torch_xla/experimental/distributed_checkpoint/manager.py

@@ -2,10 +2,13 @@
 import logging
 import os
 import pickle
+import queue
+import threading
 import torch.distributed as dist
 import torch.distributed.checkpoint as dist_cp
 import torch_xla.runtime as xr
 import torch_xla.experimental.distributed_checkpoint as xc
+import traceback
 
 from dataclasses import dataclass
 from datetime import datetime
@@ -14,6 +17,7 @@
 from os.path import basename
 from typing import Deque, List, Optional, Union
 from torch.distributed.checkpoint.metadata import STATE_DICT_TYPE
+from ._helpers import _sharded_cpu_state_dict
 
 # TODO(jonbolin): Import path will change
 from torch.distributed.checkpoint._fsspec_filesystem import FsspecReader, FsspecWriter
@@ -94,7 +98,8 @@ def __init__(self,
                path: str,
                save_interval: int,
                max_to_keep: Optional[int] = 0,
-               async_queue_size: Optional[int] = 1):
+               async_queue_size: Optional[int] = 1,
+               process_group: dist.ProcessGroup = None):
     """
     Create a checkpoint manager that reads and writes checkpoints into
     the provided directory.
@@ -113,6 +118,9 @@ def __init__(self,
             network issues which slow down the active checkpoint.
             Default: 1, which only allows a single async checkpoint to be
             pending at a time.
+      process_group: The process group to use when coordinating the checkpoint.
+            Default: None, in which case a subgroup of the default process
+            group will be created.
     """
     assert dist.is_initialized(), "A process group is required."
     assert save_interval > 0, "save_interval must be positive"
@@ -124,6 +132,16 @@ def __init__(self,
     self.max_to_keep = max_to_keep
 
     self._tracked_chkpts = self._load_tracked_chkpts()
+    self._async_queue = queue.Queue(maxsize=async_queue_size)
+    self._alive = threading.Event()
+    self._alive.set()
+    self._chkpt_thread = threading.Thread(
+        target=self._async_worker, daemon=True)
+    self._chkpt_thread.start()
+
+    # Create a new group if none is provided
+    # TODO(jonbolin): Verify subgroup on GPU backend
+    self.pg = process_group or dist.new_group()
 
   def _load_tracked_chkpts(self) -> Deque[_CheckpointMetadata]:
     """
@@ -143,6 +161,25 @@ def _load_tracked_chkpts(self) -> Deque[_CheckpointMetadata]:
       logging.warning(f'Ignoring invalid checkpoints: {invalid_paths}')
     return deque(sorted(all_chkpts, key=lambda m: m.ts))
 
+  def __del__(self):
+    self._alive.clear()
+    # Send a sentinel value to tell the worker to exit, and wait for pending
+    # checkpoints to complete.
+    self._async_queue.put(None)
+    self._chkpt_thread.join()
+
+  def _async_worker(self):
+    while self._alive.is_set():
+      try:
+        item = self._async_queue.get()
+        if item:
+          step, state_dict = item
+          self.save(step, state_dict, force=True)
+      except:
+        traceback.print_exc()
+      finally:
+        self._async_queue.task_done()
+
   def _get_path(self, step: int) -> str:
     return os.path.join(self.base_path, str(step))
 
@@ -157,7 +194,7 @@ def _release_oldest_checkpoints(self):
     Delete oldest checkpoints until the number of tracked checkpoints is below
     self.max_to_keep. This operation is only execution on the rank 0 process.
     """
-    if dist.get_rank() == 0 and self.max_to_keep > 0:
+    if dist.get_rank(self.pg) == 0 and self.max_to_keep > 0:
       while len(self._tracked_chkpts) > self.max_to_keep:
         oldest_chkpt = self._tracked_chkpts.popleft()
         self._delete_chkpt_at_step(oldest_chkpt.step)
@@ -193,6 +230,7 @@ def save(self,
           state_dict=state_dict,
           storage_writer=FsspecWriter(path),
           planner=xc.SPMDSavePlanner(),
+          process_group=self.pg,
       )
       metadata = _CheckpointMetadata(step=step, ts=datetime.now())
       with fsspec.open(os.path.join(path, _MANAGER_METADATA_FILE), 'wb') as f:
@@ -225,7 +263,12 @@ def save_async(self,
     Returns:
       True if a checkpoint was taken and False otherwise.
     """
-    raise NotImplementedError
+    if self.should_save(step) or force:
+      # Move the state_dict to CPU
+      cpu_state_dict = _sharded_cpu_state_dict(state_dict)
+      self._async_queue.put((step, cpu_state_dict))
+      return True
+    return False
 
   def restore(self, step: int, state_dict: STATE_DICT_TYPE) -> None:
     """
@@ -245,10 +288,15 @@ def restore(self, step: int, state_dict: STATE_DICT_TYPE) -> None:
         state_dict=state_dict,
         storage_reader=FsspecReader(path),
         planner=xc.SPMDLoadPlanner(),
+        process_group=self.pg,
     )
 
   def all_steps(self) -> List[int]:
     """
     List all steps tracked by the CheckpointManager.
     """
     return sorted(x.step for x in self._tracked_chkpts)
+
+  def join(self):
+    """ Wait for all pending async checkpoints to complete. """
+    self._async_queue.join()