Remove current_x validation from mixed optimizer (#2678)

Summary:
## Motivation

Removes the validation that `current_x` is within bounds from the `continuous_step` of the mixed optimizer. In some cases (e.g., with `Log` transform in Ax), the `post_processing_func` can bring the `current_x` slightly outside of the bounds, which leads to the optimization erroring out. The optimizer is capable of handling these outside of bounds starting points, so the validation doesn't provide much benefit.

Pull Request resolved: #2678

Test Plan: Removed the test checking for the validation.

Reviewed By: esantorella

Differential Revision: D68276788

Pulled By: saitcakmak

fbshipit-source-id: caf82a867eb448b9fa19a083f0c5a29acdfd8f6f

botorch/optim/optimize_mixed.py

@@ -524,10 +524,7 @@ def continuous_step(
         A tuple of two tensors: a (1 x d)-dim tensor of optimized points
             and a (1)-dim tensor of acquisition values.
     """
-    bounds = opt_inputs.bounds
     options = opt_inputs.options or {}
-    if (current_x < bounds[0]).any() or (current_x > bounds[1]).any():
-        raise ValueError("continuous_step requires current_x to be within bounds.")
     if len(discrete_dims) == len(current_x):  # nothing continuous to optimize
         with torch.no_grad():
             return current_x, opt_inputs.acq_function(current_x.unsqueeze(0))

test/optim/test_optimize_mixed.py

@@ -420,23 +420,6 @@ def test_continuous_step(self):
         self.assertTrue(X is X_out)  # testing pointer equality for due to short cut
         self.assertAllClose(ei_val, ei(X[None]))
 
-        # Input outside of bounds raises error.
-        invalid_X = X.clone()
-        invalid_X[2] = 2
-        with self.assertRaisesRegex(
-            ValueError,
-            "continuous_step requires current_x to be",
-        ):
-            X_new, ei_val = continuous_step(
-                opt_inputs=_make_opt_inputs(
-                    acq_function=ei,
-                    bounds=bounds,
-                    options={"maxiter_continuous": 32},
-                ),
-                discrete_dims=binary_dims,
-                current_x=invalid_X,
-            )
-
     def test_optimize_acqf_mixed_binary_only(self) -> None:
         train_X, train_Y, binary_dims, cont_dims = self._get_data()
         dim = len(binary_dims) + len(cont_dims)