When switching to TF2.13 and CUDA 11.8, there seems to be an issue with BinnedKNN (or related ops). Should be investigated.