Add an exception if model inputs require gradients.

botorch/models/gpytorch.py

@@ -68,12 +68,15 @@ class (e.g. an `ExactGP`) and this `GPyTorchModel`. See e.g. `SingleTaskGP`.
     def _validate_tensor_args(
         X: Tensor, Y: Tensor, Yvar: Optional[Tensor] = None, strict: bool = True
     ) -> None:
-        r"""Checks that `Y` and `Yvar` have an explicit output dimension if strict.
-        Checks that the dtypes of the inputs match, and warns if using float.
+        r"""Check the input tensors to verify that they are compatible with
+        BoTorch conventions. This checks that
 
-        This also checks that `Yvar` has the same trailing dimensions as `Y`. Note
-        we only infer that an explicit output dimension exists when `X` and `Y` have
-        the same `batch_shape`.
+        - `Y` and `Yvar` have an explicit output dimension if strict.
+        - The dtypes of the inputs match and warns if using float.
+        - `Yvar` has the same trailing dimensions as `Y`.
+          Note: We only infer that an explicit output dimension exists when
+          `X` and `Y` have the same `batch_shape`.
+        - The input tensors do not require gradients.
 
         Args:
             X: A `batch_shape x n x d`-dim Tensor, where `d` is the dimension of
@@ -131,6 +134,17 @@ def _validate_tensor_args(
                 InputDataWarning,
                 stacklevel=3,  # Warn at model constructor call.
             )
+        if (
+            X.requires_grad
+            or Y.requires_grad
+            or (Yvar is not None and Yvar.requires_grad)
+        ):
+            raise InputDataError(
+                "The BoTorch model inputs should not require gradients. This leads to "
+                f"errors during model fitting. Got {X.requires_grad=}, "
+                f"{Y.requires_grad=}"
+                + ("." if Yvar is None else f", and {Yvar.requires_grad=}.")
+            )
 
     @property
     def batch_shape(self) -> torch.Size:

test/models/test_gpytorch.py

@@ -255,6 +255,20 @@ def test_validate_tensor_args(self) -> None:
                     ):
                         GPyTorchModel._validate_tensor_args(X, Y, Yvar, strict=strict)
 
+    def test_validate_tensor_args_with_grad(self) -> None:
+        with self.assertRaisesRegex(
+            InputDataError, "inputs should not require gradients"
+        ):
+            GPyTorchModel._validate_tensor_args(
+                X=torch.randn(1, 1, requires_grad=True), Y=torch.randn(1, 1)
+            )
+        with self.assertRaisesRegex(InputDataError, "Yvar.requires_grad=False"):
+            GPyTorchModel._validate_tensor_args(
+                X=torch.randn(1, 1, requires_grad=True),
+                Y=torch.randn(1, 1),
+                Yvar=torch.randn(1, 1),
+            )
+
     def test_condition_on_observations_tensor_validation(self) -> None:
         model = SimpleGPyTorchModel(torch.rand(5, 1), torch.randn(5, 1))
         model.posterior(torch.rand(2, 1))  # evaluate the model to form caches.