Implement __torch_function__ for KeyedTensor (#3329)

Summary:
Pull Request resolved: #3329

X-link: pytorch/pytorch#161683

1.  There are a bunch of `torch.ops.aten` operations that can't handle `KeyedTensor`: The error was occurring because these ops expects a regular `Tensor` but was receiving a `KeyedTensor` object.

2.  Implement `__torch_function__` for `KeyedTensor`, so when these incompatible operations are called with a `KeyedTensor`, the `__torch_function__` method automatically delegates the op to the underlying values tensor from the `KeyedTensor` and returns a new `KeyedTensor` with updated values.

Reviewed By: malaybag

Differential Revision: D81047278

Author: jd7-tr

torchrec/sparse/jagged_tensor.py

@@ -12,7 +12,7 @@
 
 import operator
 
-from typing import Any, Callable, Dict, List, Optional, Tuple, Union
+from typing import Any, Callable, Dict, List, Optional, Tuple, Type, Union
 
 import torch
 from torch.autograd.profiler import record_function
@@ -3191,6 +3191,52 @@ class KeyedTensor(Pipelineable, metaclass=JaggedTensorMeta):
         # torch.Tensor([[2, 1, 2], [2, 1, 2], [2, 1, 2]])
     """
 
+    @classmethod
+    # pyre-ignore
+    def __torch_function__(cls: Type["KeyedTensor"], func, types, args=(), kwargs=None):
+        """
+        Enable KeyedTensor compatibility with PyTorch operations by delegating
+        operations to the underlying values tensor and reconstructing KeyedTensor
+        when appropriate.
+
+        This method allows KeyedTensor to work with various PyTorch operations
+        including compilation operations like torch.compile.
+        """
+        if kwargs is None:
+            kwargs = {}
+
+        # Handle operations that expect regular tensors but should return KeyedTensor
+        tensor_ops = [
+            torch.ops.aten._assert_tensor_metadata.default,
+            torch.ops.aten.to.dtype,
+            torch.ops.aten.to.device,
+            torch.ops.aten.detach.default,
+            torch.ops.aten.clone.default,
+        ]
+
+        if func in tensor_ops:
+            if len(args) > 0 and isinstance(args[0], cls):
+                keyed_tensor = args[0]
+                values_tensor = keyed_tensor.values()
+                new_args = (values_tensor,) + args[1:]
+                result = func(*new_args, **kwargs)
+
+                # For operations that return tensors, create new KeyedTensor with updated values
+                if isinstance(result, torch.Tensor):
+                    return cls(
+                        keys=keyed_tensor.keys(),
+                        length_per_key=keyed_tensor.length_per_key(),
+                        values=result,
+                        key_dim=keyed_tensor.key_dim(),
+                        offset_per_key=keyed_tensor._offset_per_key,
+                        index_per_key=keyed_tensor._index_per_key,
+                    )
+
+                return result
+
+        # For all other operations, return NotImplemented to allow normal handling
+        return NotImplementedError(f"{func} cannot be applied to KeyedTensor.")
+
     def __init__(
         self,
         keys: List[str],