Avoid using __torch_dispatch__ prior to v2.3.0 (#22)

There are bugs with `__torch_dispatch__` in earlier versions.

pyproject.toml

@@ -19,6 +19,7 @@ dependencies = [
     "pydantic>=2.0,<3.0",
     "cached-path",
     "requests",
+    "packaging",
 ]
 
 [project.optional-dependencies]

src/olmo_core/distributed/tensors/sharded_flat_tensor.py

@@ -8,6 +8,7 @@
 import torch
 import torch.distributed as dist
 import torch.nn.functional as F
+from packaging import version
 
 try:
     from torch.utils import _cxx_pytree as pytree
@@ -105,30 +106,36 @@ def __new__(cls, data: torch.Tensor, requires_grad: bool = False) -> ShardedFlat
 
         return tensor
 
-    def __torch_dispatch__(self, func, types, args=(), kwargs=None):
-        del types
-        kwargs = kwargs or {}
+    if version.parse(torch.__version__) >= version.parse("2.3.0"):
+        # There are some bugs with __torch_dispatch__ in earlier versions.
 
-        def unwrap(x):
-            if isinstance(x, ShardedFlatTensor):
-                return x._global_tensor if x._global_tensor is not None else x._local_tensor
-            else:
-                return x
+        def __torch_dispatch__(self, func, types, args=(), kwargs=None):
+            del types
+            kwargs = kwargs or {}
+
+            def unwrap(x):
+                if isinstance(x, ShardedFlatTensor):
+                    return x._global_tensor if x._global_tensor is not None else x._local_tensor
+                else:
+                    return x
 
-        def wrap(x):
-            if isinstance(x, torch.Tensor):
-                if x.shape == self.shape:
-                    return self.wrap(x, requires_grad=x.requires_grad)
-            return x
+            def wrap(x):
+                if isinstance(x, torch.Tensor):
+                    if x.shape == self.shape:
+                        return self.wrap(x, requires_grad=x.requires_grad)
+                return x
 
-        out = func(*pytree.tree_map(unwrap, args), **pytree.tree_map(unwrap, kwargs))
+            out = func(*pytree.tree_map(unwrap, args), **pytree.tree_map(unwrap, kwargs))
 
-        if func in {torch.ops.aten.empty_like.default, torch.ops.aten.zeros_like.default, torch.ops.aten.ones_like.default}:  # type: ignore
-            out = pytree.tree_map(wrap, out)
+            if func in {torch.ops.aten.empty_like.default, torch.ops.aten.zeros_like.default, torch.ops.aten.ones_like.default}:  # type: ignore
+                out = pytree.tree_map(wrap, out)
 
-        return out
+            return out
 
     def __repr__(self) -> str:
+        if not self.metadata_set:
+            return super().__repr__()
+
         if self._global_tensor is not None:
             return f"ShardedFlatTensor(local_tensor={self._local_tensor}, global_tensor={self._global_tensor})"
         else:

src/test/distributed/checkpoint_test.py

@@ -205,6 +205,22 @@ def test_tensor_shard_spec_for_dtensor_2D_rowwise():
     ]
 
 
+def run_get_local_tensor_data_with_dtensor():
+    mesh = init_device_mesh("cuda", (dist.get_world_size(),))
+    dtensor = distribute_tensor(torch.randn(16, device=get_default_device()), mesh, [Shard(dim=0)])
+
+    # Make sure modifying the data returned from `_get_local_tensor_data` will modify the data
+    # in the actual tensor.
+    _get_local_tensor_data(dtensor).fill_(torch.nan)
+    assert _get_local_tensor_data(dtensor).isnan().all()
+    assert dtensor.full_tensor().isnan().all()
+
+
+@requires_multi_gpu
+def test_get_local_tensor_data_with_dtensor():
+    run_distributed_test(run_get_local_tensor_data_with_dtensor, backend="nccl")
+
+
 def save_and_load_checkpoint_with_regular_and_sharded_tensors(dir):
     checkpointer = Checkpointer()
 
@@ -251,22 +267,6 @@ def save_and_load_checkpoint_with_regular_and_sharded_tensors(dir):
         assert full_state_dict["y"].shape == (2, 3)
 
 
-def run_get_local_tensor_data_with_dtensor():
-    mesh = init_device_mesh("cuda", (dist.get_world_size(),))
-    dtensor = distribute_tensor(torch.randn(16, device=get_default_device()), mesh, [Shard(dim=0)])
-
-    # Make sure modifying the data returned from `_get_local_tensor_data` will modify the data
-    # in the actual tensor.
-    _get_local_tensor_data(dtensor).fill_(torch.nan)
-    assert _get_local_tensor_data(dtensor).isnan().all()
-    assert dtensor.full_tensor().isnan().all()
-
-
-@requires_multi_gpu
-def test_get_local_tensor_data_with_dtensor():
-    run_distributed_test(run_get_local_tensor_data_with_dtensor, backend="nccl")
-
-
 @pytest.mark.parametrize("backend", BACKENDS)
 def test_save_and_load_checkpoint_with_regular_and_sharded_tensors(backend, tmp_path):
     run_distributed_test(