debug failing tests for Fabric with ddp_fork on PT 2.8 (#21093)

* let  `_get_default_process_group_backend_for_device` support more hardware platforms (#21057)

* support more hardware platforms and no longer hard code cuda when call _get_default_process_group_backend_for_device
* Apply suggestions from code review

---------

* try it
* chlog

---------

Signed-off-by: taozhiwei <taozhiweigis@163.com>
Co-authored-by: taozhiwei <taozhiweigis@163.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Nicki Skafte Detlefsen <skaftenicki@gmail.com>

(cherry picked from commit 3c81316)

Author: Borda

src/lightning/fabric/CHANGELOG.md

@@ -13,6 +13,10 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 - Include `exclude_frozen_parameters` to `DeepSpeedStrategy` ([#21060](https://github.com/Lightning-AI/pytorch-lightning/pull/21060))
 
 
+- Let `_get_default_process_group_backend_for_device` support more hardware platforms (
+    [#21057](https://github.com/Lightning-AI/pytorch-lightning/pull/21057), [#21093](https://github.com/Lightning-AI/pytorch-lightning/pull/21093))
+
+
 ### Fixed
 
 - Fixed with adding a missing device id for pytorch 2.8 ([#21105](https://github.com/Lightning-AI/pytorch-lightning/pull/21105))

src/lightning/fabric/strategies/ddp.py

@@ -160,7 +160,17 @@ def barrier(self, *args: Any, **kwargs: Any) -> None:
         if torch.distributed.get_backend() == "nccl":
             torch.distributed.barrier(device_ids=self._determine_ddp_device_ids())
         else:
-            torch.distributed.barrier()
+            # Handle PyTorch bug where barrier() fails on CPU with "PrivateUse1HooksInterface" error
+            try:
+                torch.distributed.barrier()
+            except RuntimeError as e:
+                if "PrivateUse1HooksInterface" in str(e):
+                    # Fallback: Use all_reduce as barrier - all processes must participate
+                    # This achieves the same synchronization effect as barrier()
+                    dummy_tensor = torch.tensor(0.0, device=self.root_device)
+                    torch.distributed.all_reduce(dummy_tensor)
+                else:
+                    raise
 
     @override
     def broadcast(self, obj: TBroadcast, src: int = 0) -> TBroadcast:

src/lightning/fabric/utilities/distributed.py

@@ -319,7 +319,11 @@ def _destroy_dist_connection() -> None:
 
 
 def _get_default_process_group_backend_for_device(device: torch.device) -> str:
-    return "nccl" if device.type == "cuda" else "gloo"
+    """Return corresponding distributed backend for a given device."""
+    device_backend_map = torch.distributed.Backend.default_device_backend_map
+    if device.type in device_backend_map:
+        return device_backend_map[device.type]
+    return "gloo"
 
 
 class _DatasetSamplerWrapper(Dataset):

tests/tests_fabric/utilities/test_distributed.py

@@ -17,6 +17,7 @@
 from lightning.fabric.utilities.distributed import (
     _destroy_dist_connection,
     _gather_all_tensors,
+    _get_default_process_group_backend_for_device,
     _InfiniteBarrier,
     _init_dist_connection,
     _is_dtensor,
@@ -243,6 +244,27 @@ def test_init_dist_connection_registers_destruction_handler(_, atexit_mock):
     atexit_mock.register.assert_not_called()
 
 
+def test_get_default_process_group_backend_for_device():
+    """Test that each device type maps to its correct default process group backend."""
+    # register a custom backend for test
+    torch.utils.rename_privateuse1_backend("pcu")
+
+    def mock_backend(store, group_rank, group_size, timeout):
+        pass
+
+    torch.distributed.Backend.register_backend(
+        "pccl",
+        lambda store, group_rank, group_size, timeout: mock_backend(store, group_rank, group_size, timeout),
+        devices=["pcu"],
+    )
+
+    # test that the default backend is correctly set for each device
+    devices = [torch.device("cpu"), torch.device("cuda:0"), torch.device("pcu:0")]
+    backends = ["gloo", "nccl", "pccl"]
+    for device, backend in zip(devices, backends):
+        assert _get_default_process_group_backend_for_device(device) == backend
+
+
 @RunIf(min_torch="2.4")
 def test_is_dtensor(monkeypatch):
     from torch.distributed._tensor import DTensor