Reference implementation of Megalodon 7B model

GPU-accelerated Levenberg-Marquardt curve fitting in CUDA

GPU-accelerated triangle mesh processing

Official code release for "Efficient Perspective-Correct 3D Gaussian Splatting Using Hybrid Transparency"

Qwen3-0.6B megakernel: 527 tok/s decode on RTX 3090 (3.8x faster than PyTorch)

[MICRO'23, MLSys'22] TorchSparse: Efficient Training and Inference Framework for Sparse Convolution on GPUs.