[Bug]: ModelListGP can't fantisize samples with observation_noise unless masked

[swierh]

What happened?

When using the ModelListGP.fantasize method, using observation_noise doesn't work, unless the evaluation_mask is also specified.

---

While looping over the outputs, the observation_noise_i is only set when there's a mask: https://github.com/pytorch/botorch/blob/main/botorch/models/model.py#L659

It should also be set when there's no mask.

Suggested fix, within the corresponding else statement, add:

if observation_noise is not None:
    observation_noise_i = observation_noise[..., i : i + 1]

Please provide a minimal, reproducible example of the unexpected behavior.

import torch
from botorch.fit import fit_gpytorch_mll
from botorch.models.gp_regression import SingleTaskGP
from botorch.models.model_list_gp_regression import ModelListGP
from botorch.sampling.list_sampler import ListSampler
from botorch.sampling.normal import SobolQMCNormalSampler
from gpytorch.mlls.sum_marginal_log_likelihood import SumMarginalLogLikelihood

model = ModelListGP(
    *[SingleTaskGP(train_X=torch.tensor([[0.0], [1.0]]), train_Y=torch.tensor([[0.0], [1.0]])) for _ in range(2)]
)
mll = SumMarginalLogLikelihood(model.likelihood, model)
mll = fit_gpytorch_mll(mll)
sampler = ListSampler(*[SobolQMCNormalSampler(sample_shape=torch.Size([1])) for _ in range(2)])

# Works:
model.fantasize(
    torch.tensor([[0.5]]),
    sampler=sampler,
    observation_noise=torch.tensor([[0.2, 0.2]]),
    evaluation_mask=torch.tensor([[1.0, 1.0]], dtype=torch.bool),
)

# Doesn't work:
model.fantasize(
    torch.tensor([[0.5]]),
    sampler=sampler,
    observation_noise=torch.tensor([[0.2, 0.2]]),
)

Please paste any relevant traceback/logs produced by the example provided.

File [...]/botorch/models/model.py:668, in ModelList.fantasize(self, X, sampler, observation_noise, evaluation_mask, **kwargs)
    660     else:
    661         sampler_i = (
    662             sampler.samplers[i] if isinstance(sampler, ListSampler) else sampler
    663         )
    665     fant_model = self.models[i].fantasize(
    666         X=X_i,
    667         sampler=sampler_i,
--> 668         observation_noise=observation_noise_i,
    669         **kwargs,
    670     )
    671     fant_models.append(fant_model)
    672 return self.__class__(*fant_models)

UnboundLocalError: cannot access local variable 'observation_noise_i' where it is not associated with a value

BoTorch Version

0.13.0

Python Version

3.11

Operating System

macOS 15.3

Code of Conduct

• I agree to follow BoTorch's Code of Conduct