Let TorchCollective works on the torch.distributed WORLD process group by default (#16995)

Co-authored-by: Adrian Wälchli <aedu.waelchli@gmail.com>
Co-authored-by: Carlos Mocholí <carlossmocholi@gmail.com>
Co-authored-by: Jirka Borovec <6035284+Borda@users.noreply.github.com>

Author: 4 people

src/lightning/fabric/plugins/collectives/torch_collective.py

@@ -30,6 +30,12 @@ def __init__(self) -> None:
             raise RuntimeError("Torch distributed is not available.")
         super().__init__()
 
+    @property
+    def group(self) -> CollectibleGroup:
+        if self._group is None:
+            self._group = dist.GroupMember.WORLD
+        return super().group
+
     @property
     def rank(self) -> int:
         # local rank
@@ -138,17 +144,20 @@ def setup(self, main_address: Optional[str] = None, main_port: Optional[str] = N
         return self
 
     def teardown(self) -> Self:
-        non_group_member = self.group == dist.GroupMember.NON_GROUP_MEMBER
+        group_member = self.group != dist.GroupMember.NON_GROUP_MEMBER
         super().teardown()  # will destroy its own group
         # try to destroy the default group. this should only be done by a group member to avoid race conditions,
         # and only if the class is managing it
-        if not non_group_member and TorchCollective.manages_default_group:
-            default_group = dist.GroupMember.WORLD
-            if default_group is not None:  # not destroyed already
-                group_map = dist.distributed_c10d._pg_map
-                if len(group_map) == 1 and default_group in group_map:  # only the default group is left
-                    self.destroy_group(default_group)
-                    TorchCollective.manages_default_group = False
+        if (
+            group_member
+            and TorchCollective.manages_default_group
+            and (default_group := dist.GroupMember.WORLD) is not None  # not destroyed already
+            and len(dist.distributed_c10d._pg_map) == 1  # only the default group is left
+        ):
+            self.destroy_group(default_group)
+            TorchCollective.manages_default_group = False
+        elif TorchCollective.manages_default_group and dist.GroupMember.WORLD is None:
+            TorchCollective.manages_default_group = False
         return self
 
     @classmethod
@@ -171,7 +180,8 @@ def new_group(cls, **kwargs: Any) -> CollectibleGroup:
     def destroy_group(cls, group: CollectibleGroup) -> None:
         # can be called by all processes in the default group, group will be `object()` if they are not part of the
         # current group
-        dist.destroy_process_group(group)  # type: ignore[arg-type]
+        if group in dist.distributed_c10d._pg_map:
+            dist.destroy_process_group(group)  # type: ignore[arg-type]
 
     @classmethod
     def _convert_to_native_op(cls, op: Union[str, ReduceOp, RedOpType]) -> Union[ReduceOp, RedOpType]:

tests/tests_fabric/plugins/collectives/test_torch_collective.py

@@ -146,7 +146,7 @@ def test_convert_ops():
 def test_repeated_create_and_destroy():
     collective = TorchCollective()
     with mock.patch("torch.distributed.init_process_group"):
-        collective.setup(main_address="foo", main_port=123)
+        collective.setup(main_address="foo", main_port="123")
 
     assert not os.environ
 
@@ -157,7 +157,9 @@ def test_repeated_create_and_destroy():
     with pytest.raises(RuntimeError, match="TorchCollective` already owns a group"):
         collective.create_group()
 
-    with mock.patch("torch.distributed.destroy_process_group") as destroy_mock:
+    with mock.patch.dict("torch.distributed.distributed_c10d._pg_map", {collective.group: ("", None)}), mock.patch(
+        "torch.distributed.destroy_process_group"
+    ) as destroy_mock:
         collective.teardown()
     # this would be called twice if `init_process_group` wasn't patched. once for the group and once for the default
     # group
@@ -269,3 +271,38 @@ def _test_two_groups(strategy, left_collective, right_collective):
 @pytest.mark.skip(reason="TODO(carmocca): causing hangs in CI")
 def test_two_groups():
     collective_launch(_test_two_groups, [torch.device("cpu")] * 3, num_groups=2)
+
+
+def _test_default_process_group(strategy, *collectives):
+    for collective in collectives:
+        assert collective.group == torch.distributed.group.WORLD
+    world_size = strategy.world_size
+    for c in collectives:
+        tensor = torch.tensor(world_size)
+        r = c.all_reduce(tensor)
+        assert world_size**2 == r
+
+
+@skip_distributed_unavailable
+@RunIf(skip_windows=True)
+@mock.patch.dict(os.environ, os.environ.copy(), clear=True)  # sets CUDA_MODULE_LOADING in torch==1.13
+def test_default_process_group():
+    collective_launch(_test_default_process_group, [torch.device("cpu")] * 3, num_groups=2)
+
+
+@skip_distributed_unavailable
+@mock.patch.dict(os.environ, {}, clear=True)
+def test_collective_manages_default_group():
+    collective = TorchCollective()
+    with mock.patch("torch.distributed.init_process_group"):
+        collective.setup(main_address="foo", main_port="123")
+
+    assert TorchCollective.manages_default_group
+
+    with mock.patch.object(collective, "_group") as mock_group, mock.patch.dict(
+        "torch.distributed.distributed_c10d._pg_map", {mock_group: ("", None)}
+    ), mock.patch("torch.distributed.destroy_process_group") as destroy_mock:
+        collective.teardown()
+    destroy_mock.assert_called_once_with(mock_group)
+
+    assert not TorchCollective.manages_default_group