DeepSeek's DeepJIT gives kernel authors one JIT runtime for both NVIDIA CUDA and Huawei Ascend NPUs
GitHub·medium signal
deepseek-ai/DeepJIT (created 2026-09-08, 261 stars) is a header-only C++20 runtime that compiles kernel source at runtime, caches binaries in memory and on shared distributed filesystems, and launches kernels through the same `Runtime<CUDA>` / `Runtime<Ascend>` interface. It picks up the current PyTorch CUDA or torch_npu stream by default. DeepSeek shipping Ascend as a first-class backend next to CUDA shows it is building its kernel stack to run on Huawei hardware as well as NVIDIA.