This repository provides highly optimized Lidar AI solutions for autonomous driving, featuring GPU-accelerated deep learning networks for 3D object detection. It includes implementations of PointPillars, CenterPoint, and BEVFusion with custom CUDA kernels for sparse convolution, voxelization, and feature fusion. The project offers a standalone inference engine for 3D sparse convolutional networks independent of TensorRT, plus TensorRT-integrated solutions for camera-lidar BEV fusion. Pre/post-processing are optimized with CUDA kernels to maximize throughput on NVIDIA hardware.