Added support for all_gather object (#3047)

* Added support for all_gather object

* Apply suggestions from code review

Co-authored-by: Sadra Barikbin <[email protected]>

* Added new test in _test_distrib_all_gather_group

* Handling pytorch old versions

---------

Co-authored-by: Sadra Barikbin <[email protected]>

ignite/distributed/comp_models/base.py

@@ -181,7 +181,7 @@ def _apply_op(
         return tensor
 
     def _collective_op(
-        self, tensor: Union[torch.Tensor, float, str], fn: Callable, *args: Any, **kwargs: Any
+        self, tensor: Union[torch.Tensor, Number, str], fn: Callable, *args: Any, **kwargs: Any
     ) -> Union[torch.Tensor, float, List[float], List[str]]:
         tensor_to_number = tensor_to_str = False
         device = self.device()
@@ -216,10 +216,10 @@ def all_reduce(
         return cast(Union[torch.Tensor, float], self._collective_op(tensor, self._do_all_reduce, op, group=group))
 
     def all_gather(
-        self, tensor: Union[torch.Tensor, float, str], group: Optional[Any] = None
+        self, tensor: Union[torch.Tensor, float, str, Any], group: Optional[Any] = None
     ) -> Union[torch.Tensor, float, List[float], List[str]]:
         if not isinstance(tensor, (torch.Tensor, Number, str)):
-            raise TypeError(f"Unhandled input type {type(tensor)}")
+            return self._do_all_gather_object(tensor, group=group)
 
         return self._collective_op(tensor, self._do_all_gather, group=group)
 
@@ -282,6 +282,10 @@ def _do_all_reduce(self, tensor: torch.Tensor, op: str = "SUM", group: Optional[
     def _do_all_gather(self, tensor: torch.Tensor, group: Optional[Any] = None) -> torch.Tensor:
         pass
 
+    @abstractmethod
+    def _do_all_gather_object(self, tensor: Any, group: Optional[Any] = None) -> List[Any]:
+        pass
+
     @abstractmethod
     def _do_broadcast(self, tensor: torch.Tensor, src: int) -> torch.Tensor:
         pass
@@ -373,6 +377,9 @@ def _do_all_reduce(self, tensor: torch.Tensor, op: str = "SUM", group: Optional[
     def _do_all_gather(self, tensor: torch.Tensor, group: Optional[Any] = None) -> torch.Tensor:
         return tensor
 
+    def _do_all_gather_object(self, tensor: Any, group: Optional[Any] = None) -> Any:
+        return tensor
+
     def _do_new_group(self, ranks: List[int], **kwargs: Any) -> Any:
         return ranks
 

ignite/distributed/comp_models/horovod.py

@@ -192,6 +192,12 @@ def _do_all_gather(self, tensor: torch.Tensor, group: Optional[Any] = None) -> t
                 tensor = tensor.unsqueeze(0)
             return hvd.allgather(tensor)
 
+        def _do_all_gather_object(self, tensor: Any, group: Optional[Any] = None) -> List[Any]:
+            if group is not None:
+                raise NotImplementedError("all_gather with group for horovod is not implemented")
+
+            return hvd.allgather_object(tensor)
+
         def _do_new_group(self, ranks: List[int], **kwargs: Any) -> Any:
             return hvd.ProcessSet(ranks)
 

ignite/distributed/comp_models/native.py

@@ -423,6 +423,7 @@ def _do_all_reduce(self, tensor: torch.Tensor, op: str = "SUM", group: Optional[
             if group is not None and not isinstance(group, dist.ProcessGroup):
                 raise ValueError("Argument group should be list of int or ProcessGroup")
             reduce_op = self._reduce_op_map[op]
+            # We do if/else here for compatibility with older pytorch versions
             if group is not None:
                 dist.all_reduce(tensor, reduce_op, group=group)
             else:
@@ -432,7 +433,8 @@ def _do_all_reduce(self, tensor: torch.Tensor, op: str = "SUM", group: Optional[
         def _do_all_gather(self, tensor: torch.Tensor, group: Optional[Any] = None) -> torch.Tensor:
             if group == dist.GroupMember.NON_GROUP_MEMBER:
                 return tensor
-            elif group is None:
+
+            if group is None:
                 group_size = self.get_world_size()
             elif isinstance(group, dist.ProcessGroup):
                 group_size = group.size()
@@ -441,12 +443,38 @@ def _do_all_gather(self, tensor: torch.Tensor, group: Optional[Any] = None) -> t
             if tensor.ndimension() == 0:
                 tensor = tensor.unsqueeze(0)
             output = [torch.zeros_like(tensor) for _ in range(group_size)]
+            # We do if/else here for compatibility with older pytorch versions
             if group is not None:
                 dist.all_gather(output, tensor, group=group)
             else:
                 dist.all_gather(output, tensor)
             return torch.cat(output, dim=0)
 
+        def _do_all_gather_object(self, tensor: Any, group: Optional[Any] = None) -> List[Any]:
+            if Version(torch.__version__) < Version("1.7.0"):
+                raise RuntimeError(
+                    "Current torch version does not implement dist.all_gather_object. "
+                    "Required version should be >=1.7.0"
+                )
+
+            if group == dist.GroupMember.NON_GROUP_MEMBER:
+                return tensor
+
+            if group is None:
+                group_size = self.get_world_size()
+            elif isinstance(group, dist.ProcessGroup):
+                group_size = group.size()
+            else:
+                raise ValueError("Argument group should be list of int or ProcessGroup")
+            output = [None for _ in range(group_size)]
+            # We do if/else here for compatibility with older pytorch versions
+            if group is not None:
+                dist.all_gather_object(output, tensor, group=group)
+            else:
+                dist.all_gather_object(output, tensor)
+
+            return output
+
         def _do_new_group(self, ranks: List[int], **kwargs: Any) -> Any:
             return dist.new_group(ranks=ranks, **kwargs)
 

ignite/distributed/comp_models/xla.py

@@ -155,6 +155,9 @@ def _do_all_gather(self, tensor: torch.Tensor, group: Optional[Any] = None) -> t
             xm.all_reduce("sum", [output], groups=group)
             return output.reshape(-1, *output.shape[2:])
 
+        def _do_all_gather_object(self, tensor: Any, group: Optional[Any] = None) -> List[Any]:
+            raise NotImplementedError("all_gather on object is not implemented for xla")
+
         def _do_new_group(self, ranks: List[int], **kwargs: Any) -> Any:
             return [ranks]
 

tests/ignite/distributed/utils/__init__.py

@@ -156,21 +156,22 @@ def _test_distrib_all_reduce_group(device):
 
 def _test_distrib_all_gather(device):
     rank = idist.get_rank()
+    ws = idist.get_world_size()
 
     res = torch.tensor(idist.all_gather(10), device=device)
-    true_res = torch.tensor([10] * idist.get_world_size(), device=device)
+    true_res = torch.tensor([10] * ws, device=device)
     assert (res == true_res).all()
 
     t = torch.tensor(rank, device=device)
     res = idist.all_gather(t)
-    true_res = torch.tensor([i for i in range(idist.get_world_size())], device=device)
+    true_res = torch.tensor([i for i in range(ws)], device=device)
     assert (res == true_res).all()
 
     x = "test-test"
     if rank == 0:
         x = "abc"
     res = idist.all_gather(x)
-    true_res = ["abc"] + ["test-test"] * (idist.get_world_size() - 1)
+    true_res = ["abc"] + ["test-test"] * (ws - 1)
     assert res == true_res
 
     base_x = "tests/ignite/distributed/utils/test_native.py" * 2000
@@ -179,27 +180,46 @@ def _test_distrib_all_gather(device):
         x = "abc"
 
     res = idist.all_gather(x)
-    true_res = ["abc"] + [base_x] * (idist.get_world_size() - 1)
+    true_res = ["abc"] + [base_x] * (ws - 1)
     assert res == true_res
 
     t = torch.arange(100, device=device).reshape(4, 25) * (rank + 1)
     in_dtype = t.dtype
     res = idist.all_gather(t)
-    assert res.shape == (idist.get_world_size() * 4, 25)
+    assert res.shape == (ws * 4, 25)
     assert res.dtype == in_dtype
-    true_res = torch.zeros(idist.get_world_size() * 4, 25, device=device)
-    for i in range(idist.get_world_size()):
+    true_res = torch.zeros(ws * 4, 25, device=device)
+    for i in range(ws):
         true_res[i * 4 : (i + 1) * 4, ...] = torch.arange(100, device=device).reshape(4, 25) * (i + 1)
     assert (res == true_res).all()
 
-    if idist.get_world_size() > 1:
-        with pytest.raises(TypeError, match=r"Unhandled input type"):
-            idist.all_reduce([0, 1, 2])
+    if ws > 1 and idist.backend() != "xla-tpu":
+        t = {
+            "a": [rank + 1, rank + 2, torch.tensor(rank + 3, device=device)],
+            "b": torch.tensor([[rank + 1, rank + 2, rank + 3]], device=device),
+            "c": {"abcd": rank, "cdfg": torch.tensor(rank, dtype=torch.uint8, device=device)},
+        }
+        res = idist.all_gather(t)
+        assert isinstance(res, list) and len(res) == ws
+        for i, obj in enumerate(res):
+            assert isinstance(obj, dict)
+            assert list(obj.keys()) == ["a", "b", "c"], obj
+            expected_device = (
+                device if torch.device(device).type == "cpu" else torch.device(f"{torch.device(device).type}:{i}")
+            )
+            expected = {
+                "a": [i + 1, i + 2, torch.tensor(i + 3, device=expected_device)],
+                "b": torch.tensor([[i + 1, i + 2, i + 3]], device=expected_device),
+                "c": {"abcd": i, "cdfg": torch.tensor(i, dtype=torch.uint8, device=expected_device)},
+            }
+            assert obj["a"] == expected["a"]
+            assert (obj["b"] == expected["b"]).all()
+            assert obj["c"] == expected["c"]
 
 
 def _test_distrib_all_gather_group(device):
     if idist.get_world_size() > 1:
-        ranks = [0, 1]
+        ranks = list(range(idist.get_world_size() - 1, 0, -1))  # [0, 1, 2, 3] -> [3, 2, 1]
         rank = idist.get_rank()
         bnd = idist.backend()
 
@@ -226,6 +246,40 @@ def _test_distrib_all_gather_group(device):
             else:
                 assert res == t
 
+        t = {
+            "a": [rank + 1, rank + 2, torch.tensor(rank + 3, device=device)],
+            "b": torch.tensor([[rank + 1, rank + 2, rank + 3]], device=device),
+            "c": {"abcd": rank, "cdfg": torch.tensor(rank, dtype=torch.uint8, device=device)},
+        }
+        if bnd in ("xla-tpu"):
+            with pytest.raises(NotImplementedError, match=r"all_gather on object is not implemented for xla"):
+                res = idist.all_gather(t, group=ranks)
+        elif bnd in ("horovod"):
+            with pytest.raises(NotImplementedError, match=r"all_gather with group for horovod is not implemented"):
+                res = idist.all_gather(t, group=ranks)
+        else:
+            res = idist.all_gather(t, group=ranks)
+            if rank in ranks:
+                assert isinstance(res, list) and len(res) == len(ranks)
+                for i, obj in zip(ranks, res):
+                    assert isinstance(obj, dict)
+                    assert list(obj.keys()) == ["a", "b", "c"], obj
+                    expected_device = (
+                        device
+                        if torch.device(device).type == "cpu"
+                        else torch.device(f"{torch.device(device).type}:{i}")
+                    )
+                    expected = {
+                        "a": [i + 1, i + 2, torch.tensor(i + 3, device=expected_device)],
+                        "b": torch.tensor([[i + 1, i + 2, i + 3]], device=expected_device),
+                        "c": {"abcd": i, "cdfg": torch.tensor(i, dtype=torch.uint8, device=expected_device)},
+                    }
+                    assert obj["a"] == expected["a"], (obj, expected)
+                    assert (obj["b"] == expected["b"]).all(), (obj, expected)
+                    assert obj["c"] == expected["c"], (obj, expected)
+            else:
+                assert res == t
+
         if bnd in ("nccl", "gloo", "mpi"):
             with pytest.raises(ValueError, match=r"Argument group should be list of int or ProcessGroup"):
                 res = idist.all_gather(t, group="abc")

tests/ignite/distributed/utils/test_native.py

@@ -3,6 +3,7 @@
 import pytest
 import torch
 import torch.distributed as dist
+from packaging.version import Version
 
 import ignite.distributed as idist
 from ignite.distributed.utils import has_native_dist_support
@@ -236,6 +237,7 @@ def test_idist_all_reduce_gloo(distributed_context_single_node_gloo):
 @pytest.mark.distributed
 @pytest.mark.skipif(not has_native_dist_support, reason="Skip if no native dist support")
 @pytest.mark.skipif(torch.cuda.device_count() < 1, reason="Skip if no GPU")
+@pytest.mark.skipif(Version(torch.__version__) < Version("1.7.0"), reason="dist.all_gather_object is not implemented")
 def test_idist_all_gather_nccl(distributed_context_single_node_nccl):
     device = idist.device()
     _test_distrib_all_gather(device)
@@ -244,6 +246,7 @@ def test_idist_all_gather_nccl(distributed_context_single_node_nccl):
 
 @pytest.mark.distributed
 @pytest.mark.skipif(not has_native_dist_support, reason="Skip if no native dist support")
+@pytest.mark.skipif(Version(torch.__version__) < Version("1.7.0"), reason="dist.all_gather_object is not implemented")
 def test_idist_all_gather_gloo(distributed_context_single_node_gloo):
     device = idist.device()
     _test_distrib_all_gather(device)