Update outdated Sobol dimension error message (#924)

Summary:
Pull Request resolved: #924

The current behavior looks for hard-coded `1111` in the error message. This is outdated as the current `MAXDIM` is `21201`.

I simply removed the hard coded dimension from the error message. An alternative would be to get rid of the `make_and_optimize_acqf` helper and the error handling around it. I'm happy to do that as well. Left it in here, in case someone tries to use SAASBO in 25000 dimensions (I'm not even sure if SAASBO uses this part of the code...)

Reviewed By: Balandat

Differential Revision: D35763023

fbshipit-source-id: 0f0b4b78f2b66e5bf1bc4dab1dd873d06eb71167

ax/models/torch/botorch.py

@@ -359,10 +359,6 @@ def gen(
 
         botorch_rounding_func = get_rounding_func(rounding_func)
 
-        # The following logic is to work around the limitation of PyTorch's Sobol
-        # sampler to <1111 dimensions.
-        # TODO: Remove once https://github.com/pytorch/pytorch/issues/41489 is resolved.
-
         from botorch.exceptions.errors import UnsupportedError
 
         def make_and_optimize_acqf(override_qmc: bool = False) -> Tuple[Tensor, Tensor]:
@@ -396,7 +392,7 @@ def make_and_optimize_acqf(override_qmc: bool = False) -> Tuple[Tensor, Tensor]:
         try:
             candidates, expected_acquisition_value = make_and_optimize_acqf()
         except UnsupportedError as e:
-            if "SobolQMCSampler only supports dimensions q * o <= 1111" in str(e):
+            if "SobolQMCSampler only supports dimensions" in str(e):
                 # dimension too large for Sobol, let's use IID
                 candidates, expected_acquisition_value = make_and_optimize_acqf(
                     override_qmc=True