diff --git a/src/lightning/fabric/strategies/fsdp.py b/src/lightning/fabric/strategies/fsdp.py
index 95e44b0dd553c..2edcd47b13ce9 100644
--- a/src/lightning/fabric/strategies/fsdp.py
+++ b/src/lightning/fabric/strategies/fsdp.py
@@ -895,8 +895,11 @@ def _get_full_state_dict_context(
     from torch.distributed.fsdp import FullyShardedDataParallel as FSDP
     from torch.distributed.fsdp.api import FullOptimStateDictConfig
 
-    state_dict_config = FullStateDictConfig(offload_to_cpu=True, rank0_only=rank0_only)
-    optim_state_dict_config = FullOptimStateDictConfig(offload_to_cpu=True, rank0_only=rank0_only)
+    # Offloading to CPU when FSDP is on CPU triggers a use-after-free in PyTorch's FlatParamHandle.to_cpu().
+    param = next(module.parameters(), None)
+    offload_to_cpu = param is None or param.device.type != "cpu"
+    state_dict_config = FullStateDictConfig(offload_to_cpu=offload_to_cpu, rank0_only=rank0_only)
+    optim_state_dict_config = FullOptimStateDictConfig(offload_to_cpu=offload_to_cpu, rank0_only=rank0_only)
     state_dict_type_context = FSDP.state_dict_type(
         module=module,
         state_dict_type=StateDictType.FULL_STATE_DICT,
diff --git a/tests/tests_fabric/strategies/test_fsdp.py b/tests/tests_fabric/strategies/test_fsdp.py
index 990fdf4845344..37c844a882c6b 100644
--- a/tests/tests_fabric/strategies/test_fsdp.py
+++ b/tests/tests_fabric/strategies/test_fsdp.py
@@ -433,17 +433,24 @@ def test_set_timeout(init_process_group_mock):
 
 @mock.patch("torch.distributed.fsdp.fully_sharded_data_parallel.FullyShardedDataParallel.set_state_dict_type")
 def test_get_full_state_dict_context_offload(set_type_mock, monkeypatch):
-    """Test that the state dict context manager handles CPU offloading."""
-
-    with _get_full_state_dict_context(module=Mock(spec=FullyShardedDataParallel), world_size=1):
+    """Test that the state dict context manager only offloads to CPU when the shards live on an accelerator."""
+    # Shards on accelerator: offload to CPU.
+    accelerator_param = Mock()
+    accelerator_param.device = torch.device("cuda", 0)
+    module = Mock(spec=FullyShardedDataParallel)
+    module.parameters = Mock(return_value=iter([accelerator_param]))
+    with _get_full_state_dict_context(module=module, world_size=4):
         assert set_type_mock.call_args_list[0][0][2].offload_to_cpu  # model config
         assert set_type_mock.call_args_list[0][0][3].offload_to_cpu  # optim config
 
     set_type_mock.reset_mock()
 
-    with _get_full_state_dict_context(module=Mock(spec=FullyShardedDataParallel), world_size=4):
-        assert set_type_mock.call_args_list[0][0][2].offload_to_cpu  # model config
-        assert set_type_mock.call_args_list[0][0][3].offload_to_cpu  # optim config
+    # Shards on CPU: do not offload (prevents PyTorch use-after-free).
+    module = Mock(spec=FullyShardedDataParallel)
+    module.parameters = Mock(return_value=iter([torch.nn.Parameter(torch.zeros(1))]))
+    with _get_full_state_dict_context(module=module, world_size=4):
+        assert not set_type_mock.call_args_list[0][0][2].offload_to_cpu  # model config
+        assert not set_type_mock.call_args_list[0][0][3].offload_to_cpu  # optim config
 
 
 def test_device_mesh_type_annotation():