Sharded state dicts save correctly when save_weights_only=True (#19524

)

Co-authored-by: Dimitri <dvoytan@sparkcognition.com>
Co-authored-by: awaelchli <aedu.waelchli@gmail.com>
Co-authored-by: Jirka Borovec <6035284+Borda@users.noreply.github.com>

src/lightning/pytorch/strategies/fsdp.py

@@ -568,7 +568,8 @@ def save_checkpoint(
 
             converted_state = {"model": checkpoint.pop("state_dict")}
             converted_state.update({
-                f"optimizer_{idx}": optim_state for idx, optim_state in enumerate(checkpoint.pop("optimizer_states"))
+                f"optimizer_{idx}": optim_state
+                for idx, optim_state in enumerate(checkpoint.pop("optimizer_states", []))
             })
 
             _distributed_checkpoint_save(converted_state, path)

tests/tests_pytorch/strategies/test_fsdp.py

@@ -185,7 +185,8 @@ def _run_multiple_stages(trainer, model, model_path: Optional[str] = None):
     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__)
+    if not model_path.is_dir():  # TODO (@awaelchli): Add support for asserting equality of sharded checkpoints
+        _assert_save_equality(trainer, model_path, cls=model.__class__)
 
     with torch.inference_mode():
         # Test entry point
@@ -279,11 +280,13 @@ def training_step(self, batch, batch_idx):
 
 @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):
+@pytest.mark.parametrize("state_dict_type", ["sharded", "full"])
+def test_fsdp_strategy_checkpoint(state_dict_type, precision, tmpdir):
     """Test to ensure that checkpoint is saved correctly when using a single GPU, and all stages can be run."""
     model = TestFSDPModel()
+    strategy = FSDPStrategy(state_dict_type=state_dict_type)
     trainer = Trainer(
-        default_root_dir=tmpdir, accelerator="gpu", devices=2, strategy="fsdp", precision=precision, max_epochs=1
+        default_root_dir=tmpdir, accelerator="gpu", devices=2, strategy=strategy, precision=precision, max_epochs=1
     )
     _run_multiple_stages(trainer, model, os.path.join(tmpdir, "last.ckpt"))