#4003: return TorchTensor from ttnn.to_torch

ttnn/ttnn/operations/core.py

@@ -7,6 +7,7 @@
 from typing import Union, Tuple, Optional, Any
 
 from loguru import logger
+import torch
 
 import tt_lib as ttl
 
@@ -368,6 +369,16 @@ def _to_torch_validate_input_tensors(operation_name, input_tensor, *args, **kwar
     )
 
 
+class TorchTensor(torch.Tensor):
+    @classmethod
+    def __torch_function__(cls, func, types, args=(), kwargs=None):
+        # this tells torch to treat TorchTensor just like torch.Tensor's.
+        # Otherwise, torch will complain that it doesn't know how to handle it.
+        types = tuple(torch.Tensor if t == TorchTensor else t for t in types)
+        func = ttl.tensor.decorate_external_operation(func, function_name=f"(torch) {func.__name__}")
+        return super().__torch_function__(func, types, args, kwargs)
+
+
 @ttnn.register_operation(name="ttnn.to_torch", validate_input_tensors=_to_torch_validate_input_tensors)
 def to_torch(tensor: ttnn.Tensor, *, torch_rank: Optional[int] = None) -> "torch.Tensor":
     """
@@ -413,21 +424,7 @@ def impl(ttl_tensor):
     tensor = ttnn.Tensor(ttl_tensor.reshape(tensor.shape.with_tile_padding().value))
     tensor = ttl.tensor.decorate_external_operation(impl, function_name="ttnn.to_torch")(tensor.value)
 
-    import torch
-
-    class TorchTensor(torch.Tensor):
-        @classmethod
-        def __torch_function__(
-            cls: Any,
-            func,
-            types: Any,
-            args: Any = (),
-            kwargs: Any = None,
-        ) -> Any:
-            func = ttl.tensor.decorate_external_operation(func, function_name=f"(torch) {func.__name__}")
-            return super().__torch_function__(func, types, args, kwargs)
-
-    return tensor
+    return TorchTensor(tensor)
 
 
 def _to_device_validate_input_tensors(operation_name, tensor, *args, **kwargs):