[dynamo] do not raise an unimplemented error with boolean masking set…

…item (pytorch#134902)

Cudagraph breaks on boolean masking setitem, however the code runs fine. There is no need to raise an unimplemented error here, since it already warns that its an incompatible op.

Fixes pytorch#134241

Pull Request resolved: pytorch#134902
Approved by: https://github.com/jansel, https://github.com/henrylhtsang

test/dynamo/test_export.py

@@ -4565,10 +4565,7 @@ def forward(self, x, b, y):
     return pytree.tree_unflatten([x], self._out_spec)""",  # NOQA: B950
         )
 
-        with self.assertRaisesRegex(
-            torch._dynamo.exc.Unsupported, "boolean masking setitem backwards"
-        ):
-            gm, _ = torch._dynamo.export(fn)(x, b, y)
+        gm, _ = torch._dynamo.export(fn)(x, b, y)
 
     def test_dynamo_list_index(self):
         def fn(x, in_list):

torch/_dynamo/variables/tensor.py

@@ -874,15 +874,6 @@ def has_bool_key(v):
             else:
                 return False
 
-        if (
-            has_bool_key(key)
-            and isinstance(value, TensorVariable)
-            and value.requires_grad
-            and torch.is_grad_enabled()
-        ):
-            unimplemented(
-                "boolean masking setitem backwards, see https://github.com/pytorch/pytorch/issues/114123"
-            )
         from ..symbolic_convert import InstructionTranslator
 
         tx = InstructionTranslator.current_tx()