add contiguous call to fix failing tests (#2831)

Summary:
Pull Request resolved: #2831

There is a sneaky upstream bug (pytorch/pytorch#151978) that caused some unit tests to fail and broke max posterior sampling. This contiguous call resolves it for now, until the upstream fix goes in.

Reviewed By: Balandat

Differential Revision: D73484645

fbshipit-source-id: 3daa91036a6db72c23046e9723a8a5726e71574a

Author: sdaulton

botorch/generation/sampling.py

@@ -145,8 +145,10 @@ def maximize_samples(self, X: Tensor, samples: Tensor, num_samples: int = 1):
         # to have shape batch_shape x num_samples
         if idcs.ndim > 1:
             idcs = idcs.permute(*range(1, idcs.ndim), 0)
-        # in order to use gather, we need to repeat the index tensor d times
-        idcs = idcs.unsqueeze(-1).expand(*idcs.shape, X.size(-1))
+        # in order to use gather, we need to repeat the index tensor d times.
+        # The contiguous call is needed due to a pytorch issue:
+        # https://github.com/pytorch/pytorch/issues/151978
+        idcs = idcs.unsqueeze(-1).expand(*idcs.shape, X.size(-1)).contiguous()
         # now if the model is batched batch_shape will not necessarily be the
         # batch_shape of X, so we expand X to the proper shape
         Xe = X.expand(*obj.shape[1:], X.size(-1))