Max Balandat
Max Balandat
Hmm I think we can probably inject some logic into the caching code that would check the object for which device/dtype it's on and then if needed move the caches...
Could the same not be achieved by transposing the edge_features input?
I am not aware of any plans for this, but glancing at the code it seems like it should be relatively straightfoward to implement something like this in gpytorch. We're...
Could you please post the full stack trace to make it easier to debug?
@irinaespejo is the downstream code based on BoTorch by any chance?
> I believe that there's utilities for returning only one task during the posterior call in Botorch's MultitaskGP models. Yeah there is some functionality for a model here: https://github.com/pytorch/botorch/blob/main/botorch/models/multitask.py#L105-L106 I...
Interesting. Large negative eval indeed. Is the CUDA version the same for both torch versions? Maybe @mruberry or @ngimel have any thoughts what could be causing this discrepancy?
Do you know whether the eval computations themselves are ok? I.e. what I would suggest is to compute `f_covar` both in 1.9.0 and in 1.9.1 and save the tensors to...
How do the state dicts of the model look like when fitted with either torch version? Seems like some of the parameters must be degenerate somehow to result in this.
Interesting. There should be no explicit inverses anywhere on the code though (there better not be!). Does the same happen when you use cholesky + cholesky_solce instead of the inverse?