Fix trainer.save_checkpoint after trainer.test with FSDP (#18992)

src/lightning/pytorch/CHANGELOG.md

@@ -52,6 +52,9 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 - Fixed an issue where Metric instances from `torchmetrics` wouldn't get moved to the device when using FSDP ([#18954](https://github.com/Lightning-AI/lightning/issues/18954))
 
 
+- Fixed an issue preventing the user to `Trainer.save_checkpoint()` an FSDP model when `Trainer.test/validate/predict()` ran after `Trainer.fit()` ([#18992](https://github.com/Lightning-AI/lightning/issues/18992))
+
+
 ## [2.1.0] - 2023-10-11
 
 ### Added

src/lightning/pytorch/strategies/fsdp.py

@@ -341,6 +341,11 @@ def setup(self, trainer: "pl.Trainer") -> None:
 
     @override
     def setup_optimizers(self, trainer: "pl.Trainer") -> None:
+        # If we're setting up for evaluation after fitting, we need to discard the optimizers
+        # since we're rewrapping the model, otherwise optimizer param references are no longer valid
+        # and subsequent checkpoint saving can fail
+        self._reset_optimizers_and_schedulers()
+
         if self.kwargs.get("use_orig_params"):
             return super().setup_optimizers(trainer)
 

src/lightning/pytorch/strategies/strategy.py

@@ -575,6 +575,11 @@ def on_exception(self, exception: BaseException) -> None:
         """Called when the trainer execution is interrupted by an exception."""
         pass
 
+    def _reset_optimizers_and_schedulers(self) -> None:
+        self._optimizers = []
+        self._lightning_optimizers = []
+        self.lr_scheduler_configs = []
+
     def __getstate__(self) -> Dict:
         # `LightningOptimizer` overrides `self.__class__` so they cannot be pickled
         state = dict(vars(self))  # copy

tests/tests_pytorch/strategies/test_fsdp.py

@@ -173,9 +173,13 @@ def _assert_layer_fsdp_instance(self) -> None:
 
 def _run_multiple_stages(trainer, model, model_path: Optional[str] = None):
     trainer.fit(model)
+    trainer.test(model)
+
     model_path = trainer.strategy.broadcast(model_path)
-    model_path = model_path if model_path else trainer.checkpoint_callback.last_model_path
+    model_path = Path(model_path if model_path else trainer.checkpoint_callback.last_model_path)
 
+    # Save another checkpoint after testing, without optimizer states
+    trainer.save_checkpoint(model_path.with_name("after-test"))
     trainer.save_checkpoint(model_path, weights_only=True)
 
     _assert_save_equality(trainer, model_path, cls=model.__class__)
@@ -270,13 +274,13 @@ def training_step(self, batch, batch_idx):
     trainer.fit(model)
 
 
-@RunIf(min_cuda_gpus=1, skip_windows=True, standalone=True)
+@RunIf(min_cuda_gpus=2, skip_windows=True, standalone=True)
 @pytest.mark.parametrize("precision", ["16-mixed", pytest.param("bf16-mixed", marks=RunIf(bf16_cuda=True))])
 def test_fsdp_strategy_checkpoint(tmpdir, precision):
     """Test to ensure that checkpoint is saved correctly when using a single GPU, and all stages can be run."""
     model = TestFSDPModel()
     trainer = Trainer(
-        default_root_dir=tmpdir, accelerator="gpu", devices=1, strategy="fsdp", precision=precision, max_epochs=1
+        default_root_dir=tmpdir, accelerator="gpu", devices=2, strategy="fsdp", precision=precision, max_epochs=1
     )
     _run_multiple_stages(trainer, model, os.path.join(tmpdir, "last.ckpt"))