Fix issue in ckpt saver.

.github/workflows/main.yml

@@ -3,6 +3,7 @@ name: CI
 
 on:
  pull_request:
+ workflow_dispatch:
  push:
  branches: [master]
 

dlrover/python/elastic_agent/torch/ckpt_saver.py

@@ -414,11 +414,7 @@ def __init__(
  )
  self._master_client = None
 
- # remove the history temp path if exists
- self.storage.safe_rmtree(
- os.path.join(self.checkpoint_dir, self._STAGE_DIR)
- )
- logger.info("AsyncSaver initialized.")
+ logger.info(f"AsyncSaver({self.__class__.__name__}) initialized.")
 
  def __del__(self):
  self.close()
@@ -781,10 +777,17 @@ def _sync_node_checkpoint(
  if elapsed_time > timeout:
  logger.info(
  "It is timeout to sync checkpoint "
- "bacause some nodes may fail."
+ "because some nodes may fail."
  )
  return False
 
+ def _remove_sub_dir_of_target_path(self, path):
+ if os.path.exists(path):
+ for entry in os.listdir(path):
+ full_path = os.path.join(path, entry)
+ if os.path.isdir(full_path):
+ self.storage.safe_rmtree(full_path)
+
  @classmethod
  def reset(cls):
  """Reset the shared memory of all shards."""
@@ -999,6 +1002,28 @@ class TempDirCheckpointSaver(AsyncCheckpointSaver):
  by users.
  """
 
+ def __init__(
+ self,
+ checkpoint_dir,
+ storage_meta: ClassMeta,
+ local_shard_num=1,
+ global_shard_num=1,
+ save_timeout=CheckpointConstant.SAVE_TIMEOUT,
+ ) -> None:
+ super().__init__(
+ checkpoint_dir,
+ storage_meta,
+ local_shard_num,
+ global_shard_num,
+ save_timeout,
+ )
+
+ if self._node_rank == 0:
+ # remove the history temp path if exists
+ self._remove_sub_dir_of_target_path(
+ os.path.join(self.checkpoint_dir, self._STAGE_DIR)
+ )
+
  def save_step_checkpoint(self, step):
  """
  Save the checkpoint of a step into the storage.

dlrover/python/tests/test_ckpt_saver.py

@@ -157,6 +157,22 @@ def test_create_checkpoint_saver(self):
  id(AsyncCheckpointSaver._saver_instance._master_client),
  )
 
+ # test
+ test_path = "/tmp/test_ckpt"
+ AsyncCheckpointSaver._saver_instance.checkpoint_dir = test_path
+ os.makedirs(test_path, exist_ok=True)
+ os.makedirs(os.path.join(test_path, "td1"), exist_ok=True)
+ with open(
+ os.path.join(test_path, "tf1"), "w", encoding="utf-8"
+ ) as file:
+ file.write("test")
+ AsyncCheckpointSaver._saver_instance._remove_sub_dir_of_target_path(
+ test_path
+ )
+ self.assertTrue(os.path.exists(test_path))
+ self.assertTrue(os.path.exists(os.path.join(test_path, "tf1")))
+ self.assertFalse(os.path.exists(os.path.join(test_path, "td1")))
+
  def test_close_saver(self):
  saver = DdpCheckpointSaver("test_ckpt", self.storage.get_class_meta())
  try:

dlrover/trainer/torch/flash_checkpoint/engine.py

@@ -174,7 +174,7 @@ def __init__(
  self.storage = storage
  self._save_timeout = save_timeout
  self._local_rank = env_utils.get_local_rank()
- self._cached_step = 0
+ self._cached_step = -1
  self._restart_count = env_utils.get_torch_restart_count()
 
  # init saver