Fix: Allow trainer to accept CUDAAccelerator instance as accelerator with FSDP strategy (#20964)

* Add test for FSDP with CUDAAccelerator instance
* update error message
* update the test

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
(cherry picked from commit 460c60c)

Author: 2 people

src/lightning/pytorch/CHANGELOG.md

@@ -15,7 +15,7 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 
 ### Fixed
 
--
+- Fixed Allowing trainer to accept CUDAAccelerator instance as accelerator with FSDP strategy ([#20964](https://github.com/Lightning-AI/pytorch-lightning/pull/20964))
 
 
 ## [2.5.2] - 2025-06-20

src/lightning/pytorch/trainer/connectors/accelerator_connector.py

@@ -453,10 +453,11 @@ def _check_strategy_and_fallback(self) -> None:
 
         if (
             strategy_flag in FSDPStrategy.get_registered_strategies() or type(self._strategy_flag) is FSDPStrategy
-        ) and self._accelerator_flag not in ("cuda", "gpu"):
+        ) and not (self._accelerator_flag in ("cuda", "gpu") or isinstance(self._accelerator_flag, CUDAAccelerator)):
             raise ValueError(
-                f"The strategy `{FSDPStrategy.strategy_name}` requires a GPU accelerator, but got:"
-                f" {self._accelerator_flag}"
+                f"The strategy `{FSDPStrategy.strategy_name}` requires a GPU accelerator, but received "
+                f"`accelerator={self._accelerator_flag!r}`. Please set `accelerator='cuda'`, `accelerator='gpu'`,"
+                " or pass a `CUDAAccelerator()` instance to use FSDP."
             )
         if strategy_flag in _DDP_FORK_ALIASES and "fork" not in torch.multiprocessing.get_all_start_methods():
             raise ValueError(

tests/tests_pytorch/trainer/connectors/test_accelerator_connector.py

@@ -582,6 +582,11 @@ class AcceleratorSubclass(CPUAccelerator):
     Trainer(accelerator=AcceleratorSubclass(), strategy=FSDPStrategySubclass())
 
 
+@RunIf(min_cuda_gpus=1)
+def test_check_fsdp_strategy_and_fallback_with_cudaaccelerator():
+    Trainer(strategy="fsdp", accelerator=CUDAAccelerator())
+
+
 @mock.patch.dict(os.environ, {}, clear=True)
 def test_unsupported_tpu_choice(xla_available, tpu_available):
     # if user didn't set strategy, _Connector will choose the SingleDeviceXLAStrategy or XLAStrategy