Add independent flag to track checkpoint resumption.

Signed-off-by: sudipto baral <sudiptobaral.me@gmail.com>

Author: sudiptob2

src/lightning/pytorch/loops/fit_loop.py

@@ -204,11 +204,6 @@ def skip(self) -> bool:
         # so we cannot use it solely
         return self.done or self.trainer.limit_train_batches == 0
 
-    @property
-    def _is_resuming(self) -> bool:
-        """Whether we're resuming training from a checkpoint."""
-        return self._loaded_from_state_dict
-
     def run(self) -> None:
         self.setup_data()
         if self.skip:

src/lightning/pytorch/loops/loop.py

@@ -23,13 +23,19 @@ class _Loop:
     def __init__(self, trainer: "pl.Trainer") -> None:
         self._restarting = False
         self._loaded_from_state_dict = False
+        self._resuming_from_checkpoint = False
         self.trainer = trainer
 
     @property
     def restarting(self) -> bool:
         """Whether the state of this loop was reloaded and it needs to restart."""
         return self._restarting
 
+    @property
+    def is_resuming(self) -> bool:
+        """Whether we're resuming training from a checkpoint."""
+        return self._resuming_from_checkpoint
+
     @restarting.setter
     def restarting(self, restarting: bool) -> None:
         """Connects this loop's restarting value and its children."""
@@ -87,6 +93,7 @@ def load_state_dict(
                 v.load_state_dict(state_dict.copy(), prefix + k + ".")
         self.restarting = True
         self._loaded_from_state_dict = True
+        self._resuming_from_checkpoint = True
 
     def _load_from_state_dict(self, state_dict: dict, prefix: str) -> None:
         for k, v in self.__dict__.items():
@@ -102,4 +109,5 @@ def _load_from_state_dict(self, state_dict: dict, prefix: str) -> None:
     def on_iteration_done(self) -> None:
         self._restarting = False
         self._loaded_from_state_dict = False
+        self._resuming_from_checkpoint = False
         self.reset_restart_stage()

src/lightning/pytorch/loops/training_epoch_loop.py

@@ -239,9 +239,9 @@ def on_run_start(self, data_fetcher: _DataFetcher) -> None:
         # `iter()` was called once in `FitLoop.setup_data()` already
         # Only call `iter()` if all following cases:
         # 1. Not restarting
-        # 2. Not resuming from checkpoint (not _is_resuming)
+        # 2. Not resuming from checkpoint (not is_resuming)
         # 3. Past first epoch (current_epoch > 0)
-        if (self.trainer.current_epoch > 0 and not self.trainer.fit_loop._is_resuming) and not self.restarting:
+        if (self.trainer.current_epoch > 0 and not self.trainer.fit_loop.is_resuming) and not self.restarting:
             iter(data_fetcher)  # creates the iterator inside the fetcher
 
         # add the previous `fetched` value to properly track `is_last_batch` with no prefetching