lightning-transformers VS Pointnet2_PyTorch

Lightning Transformers is for users who want to train, evaluate and predict using HuggingFace models and datasets with PyTorch Lightning. Full customizability of the code using the LightningModule and Trainer, with Hydra config composition for quick and easy experimentation. No boilerplate code is required; easily swap out models, optimizers, schedulers, and more without touching the code. Check out the blog post: Training Transformers at Scale with PyTorch Lightning for more information or the documentation.

Apologies if this isn't the right place to ask. But I'm currently studying point cloud-based networks like pointcloud++, and all the related 3d object detection networks like pointpillars, voxelnet, etc. While I (think) understand the algorithms like feature propagation in pointnet++. I'm having trouble understanding how would one implement them. Or Where could I learn about writing operations in cuda and making sure they are compatible with backprop?