dvlt.cu
A dependency-free, pure CUDA and C++ reimplementation of NVIDIA's DVLT transformer that reconstructs a 3D point cloud and camera poses from a handful of images in a single forward pass, packaged as one small binary.
dvlt.cu is a from-scratch CUDA and C++ port of NVIDIA's DVLT (Deja View Looping Transformer) model, written without Python, PyTorch or any inference framework. Given a folder of images or a video of a static scene, it reconstructs the scene in 3D in one forward pass, with no per-scene training, and writes out a merged point cloud along with the recovered camera pose for each view.
Internally it memory-maps bf16 model weights, performs a single bulk upload to the GPU, uses a one-shot memory arena, and fuses operations such as attention, normalization and rotary embeddings into custom kernels built on cuBLASLt and cuTLASS, with float32 kept only for precision-sensitive output layers. It depends on little beyond cuBLASLt, header-only cuTLASS and a small image-loading header, and is released under the Apache-2.0 license.