diff --git a/gpytorch/models/approximate_gp.py b/gpytorch/models/approximate_gp.py
index 85e2674f4..36165576e 100644
--- a/gpytorch/models/approximate_gp.py
+++ b/gpytorch/models/approximate_gp.py
@@ -105,4 +105,5 @@ def get_fantasy_model(self, inputs, targets, **kwargs):
     def __call__(self, inputs, prior=False, **kwargs):
         if inputs.dim() == 1:
             inputs = inputs.unsqueeze(-1)
+        inputs = self.apply_input_transforms(X=inputs, is_training_input=self.training)
         return self.variational_strategy(inputs, prior=prior, **kwargs)
diff --git a/gpytorch/models/exact_gp.py b/gpytorch/models/exact_gp.py
index a82ef7d19..72a259555 100644
--- a/gpytorch/models/exact_gp.py
+++ b/gpytorch/models/exact_gp.py
@@ -210,7 +210,11 @@ def get_fantasy_model(self, inputs, targets, **kwargs):
         except KeyError:
             fantasy_kwargs = {}
 
-        full_output = super(ExactGP, self).__call__(*full_inputs, **kwargs)
+        # Prediction strategy should have transformed train inputs.
+        prediction_strategy_inputs = [
+            self.apply_input_transforms(X=t_input, is_training_input=True) for t_input in full_inputs
+        ]
+        full_output = super(ExactGP, self).__call__(*prediction_strategy_inputs, **kwargs)
 
         # Copy model without copying training data or prediction strategy (since we'll overwrite those)
         old_pred_strat = self.prediction_strategy
@@ -229,7 +233,7 @@ def get_fantasy_model(self, inputs, targets, **kwargs):
 
         new_model.likelihood = old_likelihood.get_fantasy_likelihood(**fantasy_kwargs)
         new_model.prediction_strategy = old_pred_strat.get_fantasy_strategy(
-            inputs, targets, full_inputs, full_targets, full_output, **fantasy_kwargs
+            inputs, targets, prediction_strategy_inputs, full_targets, full_output, **fantasy_kwargs
         )
 
         # if the fantasies are at the same points, we need to expand the inputs for the new model
@@ -242,8 +246,17 @@ def get_fantasy_model(self, inputs, targets, **kwargs):
         return new_model
 
     def __call__(self, *args, **kwargs):
-        train_inputs = list(self.train_inputs) if self.train_inputs is not None else []
-        inputs = [i.unsqueeze(-1) if i.ndimension() == 1 else i for i in args]
+        train_inputs = (
+            [self.apply_input_transforms(X=t_input, is_training_input=True) for t_input in self.train_inputs]
+            if self.train_inputs is not None
+            else []
+        )
+        inputs = [
+            self.apply_input_transforms(
+                X=i.unsqueeze(-1) if i.ndimension() == 1 else i, is_training_input=self.training
+            )
+            for i in args
+        ]
 
         # Training mode: optimizing
         if self.training:
diff --git a/gpytorch/models/gp.py b/gpytorch/models/gp.py
index 922e981b0..e74d38637 100644
--- a/gpytorch/models/gp.py
+++ b/gpytorch/models/gp.py
@@ -1,7 +1,14 @@
 #!/usr/bin/env python3
 
+from torch import Tensor
+
 from ..module import Module
 
 
 class GP(Module):
-    pass
+    def apply_input_transforms(self, X: Tensor, is_training_input: bool) -> Tensor:
+        input_transform = getattr(self, "input_transform", None)
+        if input_transform is not None:
+            return input_transform(X=X, is_training_input=is_training_input)
+        else:
+            return X