Main bottleneck is the $predict() and $predict_newdata() steps, so minimizing how often they are called will be helpful.

• One approach is to create the datasets for e.g. PFI individually, rbind them together while keeping track in which chunk of data which feature is permuted, and then only calling $predict once on the combined data.
• Analogously, this can be done across multiple iterations (iters_perm) of the permutations for PFI and SAGE.
• Since resamplings are independent and have different trained models, parallelizing across those would be an option as well.
• Chunking the data beforehand