(VariationalHierarchicalHyperparameters, DirichletKernelMulticlassClassification) decorator combination test causes unrelated tests to fail

[rg936672]

What's the problem?

This pair of decorators is normally excluded from the combination tests due to "can't aggregate multitask predictions". However, removing this disable does not only introduce two failing tests (as would be expected, one for each ordering of the two decorators), it also causes several unrelated combinations DirichletMulticlassClassification to fail. This indicates that the tests are not properly isolated from one another..

How can we reproduce the issue?

Comment out the lines

    (
        VariationalHierarchicalHyperparameters,
        DirichletMulticlassClassification,
    ),  # can't aggregate multitask predictions

in test_decorator_combinations.py (in the definition of EXCLUDED_COMBINATIONS). Run the tests. Observe that unrelated tests (which passed before) now fail.

Python version

3.12

Package version

2.1.0

Operating system

No response

Other packages

No response

Relevant log output

For example:

FAILED [ 14%]
tests\units\test_decorator_combinations.py:186 (test_combinations[Upper: DirichletMulticlassClassification - Lower: SetWarp])
test_decorator_combinations.py:256: in test_combinations
    controller.fit(1)
..\..\vanguard\base\gpcontroller.py:148: in fit
    loss = self._sgd_round(n_iters=n_sgd_iters, gradient_every=gradient_every)
..\..\vanguard\warps\decorator.py:187: in _sgd_round
    loss = super()._sgd_round(n_iters=n_iters, gradient_every=gradient_every)
..\..\vanguard\base\basecontroller.py:381: in _sgd_round
    raise err
..\..\vanguard\base\basecontroller.py:364: in _sgd_round
    loss = self._single_optimisation_step(train_x, train_y, retain_graph=iter_num < n_iters - 1)
..\..\vanguard\base\basecontroller.py:405: in _single_optimisation_step
    loss = self._loss(x, y)
..\..\vanguard\classification\dirichlet.py:234: in _loss
    return super()._loss(train_x, train_y).sum()
..\..\vanguard\warps\decorator.py:207: in _loss
    nmll = super()._loss(train_x, warped_train_y)
..\..\vanguard\base\basecontroller.py:426: in _loss
    return -self._mll(output, train_y.squeeze(dim=-1))
..\..\vanguard\base\basecontroller.py:602: in __call__
    result = super().__call__(*args, **kwargs)
..\..\..\..\.virtualenvs\Vanguard-ShqWj01e\Lib\site-packages\gpytorch\module.py:31: in __call__
    outputs = self.forward(*inputs, **kwargs)
..\..\..\..\.virtualenvs\Vanguard-ShqWj01e\Lib\site-packages\gpytorch\mlls\exact_marginal_log_likelihood.py:66: in forward
    output = self.likelihood(function_dist, *params, **kwargs)
..\..\..\..\.virtualenvs\Vanguard-ShqWj01e\Lib\site-packages\gpytorch\likelihoods\gaussian_likelihood.py:469: in __call__
    return super().__call__(input, *args, **kwargs)
..\..\..\..\.virtualenvs\Vanguard-ShqWj01e\Lib\site-packages\gpytorch\likelihoods\likelihood.py:76: in __call__
    return self.marginal(input, *args, **kwargs)
..\..\..\..\.virtualenvs\Vanguard-ShqWj01e\Lib\site-packages\gpytorch\likelihoods\gaussian_likelihood.py:461: in marginal
    return super().marginal(function_dist, *args, **kwargs)
..\..\..\..\.virtualenvs\Vanguard-ShqWj01e\Lib\site-packages\gpytorch\likelihoods\gaussian_likelihood.py:359: in marginal
    return super().marginal(function_dist, *args, **kwargs)
..\..\..\..\.virtualenvs\Vanguard-ShqWj01e\Lib\site-packages\gpytorch\likelihoods\gaussian_likelihood.py:117: in marginal
    full_covar = covar + noise_covar
..\..\..\..\.virtualenvs\Vanguard-ShqWj01e\Lib\site-packages\linear_operator\operators\_linear_operator.py:2774: in __add__
    return AddedDiagLinearOperator(self, other)
..\..\..\..\.virtualenvs\Vanguard-ShqWj01e\Lib\site-packages\gpytorch\lazy\lazy_tensor.py:46: in __init__
    return __orig_init__(self, *args, **new_kwargs)
..\..\..\..\.virtualenvs\Vanguard-ShqWj01e\Lib\site-packages\linear_operator\operators\added_diag_linear_operator.py:43: in __init__
    super(AddedDiagLinearOperator, self).__init__(*linear_ops, preconditioner_override=preconditioner_override)
..\..\..\..\.virtualenvs\Vanguard-ShqWj01e\Lib\site-packages\gpytorch\lazy\lazy_tensor.py:46: in __init__
    return __orig_init__(self, *args, **new_kwargs)
..\..\..\..\.virtualenvs\Vanguard-ShqWj01e\Lib\site-packages\linear_operator\operators\sum_linear_operator.py:23: in __init__
    batch_shape = torch.broadcast_shapes(*[lt.batch_shape for lt in linear_ops])
..\..\..\..\.virtualenvs\Vanguard-ShqWj01e\Lib\site-packages\torch\functional.py:133: in broadcast_shapes
    raise RuntimeError("Shape mismatch: objects cannot be broadcast to a single shape")
E   RuntimeError: Shape mismatch: objects cannot be broadcast to a single shape