[Bug] `ard_num_dims` no longer passed from `LinearKernel` to `Kernel` base class

[AdrianSosic]

🐛 Bug

With the newest release, the ard_num_dims attribute (and potentially others!?) is no longer set in the Kernel base class. I think the problem is that it entered via kwargs in the previous version but now comes as an explicit parameter, which was perhaps forgotten to be passed to the super().__init__ call?

To reproduce

Code snippet to reproduce

from gpytorch.kernels import LinearKernel

print(LinearKernel(ard_num_dims=1).ard_num_dims)

Stack trace/error message

1  # gpytorch==1.13
None  # gpytorch==1.14

System information

Please complete the following information:

• GPyTorch Version: 1.14
• PyTorch Version: 2.5.1
• Computer OS: macOS

[Balandat]

This is for the LinearKernel only, right? If so, can you adjust title and description of the issue?

Also, does this actually cause any issues? The variance parameter is still correctly initialized with the right number of dimensions as specified in ard_num_dims. Is this mostly a convenience / API inconsistency or are there situations in which this leads to incorrect results or errors?

cc @sdaulton who made the recent change affecting this behavior in #2593

[sdaulton]

I'm also curious if this is causing issues @AdrianSosic . ard_num_dims in the base Kernel refers to the number of lengthscales, which isn't relevant for the LinearKernel.

[gpleiss]

It is just an issue on LinearKernel. Looking at the diff between v1.13 and v1.14

+    :param ard_num_dims: Set this if you want a separate variance priors for each weight. (Default: `None`)
     :param variance_prior: Prior over the variance parameter. (Default `None`.)
     :param variance_constraint: Constraint to place on variance parameter. (Default: `Positive`.)
-    :param active_dims: List of data dimensions to operate on. `len(active_dims)` should equal `num_dimensions`.
+    :param active_dims: List of data dimensions to operate on.
     """

     def __init__(
         self,
-        num_dimensions: Optional[int] = None,
-        offset_prior: Optional[Prior] = None,
+        ard_num_dims: Optional[int] = None,
         variance_prior: Optional[Prior] = None,
         variance_constraint: Optional[Interval] = None,
         **kwargs,
     ):
-        super(LinearKernel, self).__init__(**kwargs)
+        super().__init__(**kwargs)

So we simply forgot pass in ard_num_dims into the super constructor. I can put up a PR..

[gpleiss]

@sdaulton since the variance parameter uses ard_num_dims as per the new PR, it seems reasonable to add it as an attribute on the kernel.

[AdrianSosic]

Also, does this actually cause any issues? The variance parameter is still correctly initialized with the right number of dimensions as specified in ard_num_dims. Is this mostly a convenience / API inconsistency or are there situations in which this leads to incorrect results or errors?

I'm also curious if this is causing issues @AdrianSosic

No, I don't think it causes any problems, but:

1. It's still a backward incompatible change, strictly speaking
2. I think it is very unintuitive API-wise to pass an argument to a constructor that is actually used by then object, the object has an actual attribute with the same name, but the passed argument is not reflected in the attribute value.

In fact, I noticed this since one of our CI tests started to fail that checks the API of our custom wrapper around the gpytorch kernels – which sort of underpins both points above.