DeepJIT

A lightweight, header-only C++20 runtime for just-in-time compilation of accelerator kernels that targets both NVIDIA CUDA and Huawei Ascend (Bisheng and ACL), with shared on-disk and distributed kernel caches and PyTorch and torch_npu integration. Released the week of V4.1-Flash by the DeepGEMM authors, it is the compilation layer under DeepSeek's kernel libraries and a marker of the lab's dual-stack CUDA-plus-Ascend engineering (257 stars in two days).