Regressor memory increase with each estimator iteration

[noahho]

Describe the workflow you want to enable

The regressor saves the regression output which is a vector of probabilities for each output https://github.com/PriorLabs/TabPFN/blob/main/src/tabpfn/regressor.py#L914 This can grow quite large: 10000 samples * 5000 borders * 2 byte * 8 estimators = 6.4GB which are all stored on GPU memory

Describe your proposed solution

Saving a compressed version of the borders (dropping out low probability borders) would be much more memory efficient. Since the probabilities and borers are recombined anyways with different borders, we could do the same and remove a lot of the low probability borders.

Describe alternatives you've considered, if relevant

No response

Additional context

No response

Impact

None

[andy7002]

For the regression task, if the predict interface is called multiple times, the GPU memory will keep increasing. How to fix it ?