Thanks for the awesome job! Currently nr3d_lib supports float64, float32 and float16, I think bfloat16 support can be introduced without too much effort and for my experience it can stabilize NeRF training with half precision. Hope you can consider this option later, thanks!