fix(priors): avoid RecursionError when moving MultivariateNormalPrior to device

gpytorch/priors/prior.py

@@ -42,7 +42,11 @@ def __setattr__(self, name: str, value: Any) -> None:
         # If setting a BUFFERED_PREFIX attribute, update the base attribute instead.
         # Note: BUFFERED_PREFIX is just an indicator that this attribute belongs to a
         # TransformedDistribution, the value itself is not transformed.
-        if hasattr(self, name) and BUFFERED_PREFIX in name:
+        #
+        # Check BUFFERED_PREFIX before hasattr: torch.distributions lazy_property
+        # descriptors re-enter via setattr to cache (e.g. covariance_matrix), and
+        # hasattr() on those names would recurse infinitely.
+        if BUFFERED_PREFIX in name and hasattr(self, name):
             base_attr_name = name.replace(BUFFERED_PREFIX, "")
             # Convert to Tensor if needed
             tensor_value = torch.as_tensor(value)

gpytorch/priors/torch_priors.py

@@ -152,6 +152,12 @@ def cpu(self):
         _del_attributes(module, MVN_LAZY_PROPERTIES)
         return module
 
+    def to(self, *args, **kwargs):
+        """Applies module-level to() call and resets all lazy properties"""
+        module = TModule.to(self, *args, **kwargs)
+        _del_attributes(module, MVN_LAZY_PROPERTIES)
+        return module
+
     def expand(self, batch_shape):
         batch_shape = torch.Size(batch_shape)
         cov_shape = batch_shape + self.event_shape

test/priors/test_multivariate_normal_prior.py

@@ -18,6 +18,19 @@ def test_multivariate_normal_prior_to_gpu(self):
             self.assertEqual(prior.scale_tril.device.type, "cuda")
             self.assertEqual(prior.precision_matrix.device.type, "cuda")
 
+    def test_multivariate_normal_prior_device_move_no_recursion(self):
+        # Accessing torch.distributions lazy properties after a device move must not
+        # RecursionError via Prior.__setattr__ / lazy_property caching.
+        prior = MultivariateNormalPrior(torch.tensor([0.0, 1.0]), covariance_matrix=torch.eye(2))
+        _ = prior.covariance_matrix  # cache lazy properties first
+        prior = prior.to(torch.device("cpu"))
+        self.assertEqual(prior.loc.device.type, "cpu")
+        self.assertEqual(prior.covariance_matrix.device.type, "cpu")
+        self.assertEqual(prior.scale_tril.device.type, "cpu")
+        self.assertEqual(prior.precision_matrix.device.type, "cpu")
+        prior = prior.cpu()
+        self.assertEqual(prior.covariance_matrix.device.type, "cpu")
+
     def test_multivariate_normal_prior_validate_args(self):
         with self.assertRaises(ValueError):
             mean = torch.tensor([0.0, 1.0])