DeepSeek open-sources DeepJIT, a C++20 JIT runtime for CUDA and Huawei Ascend NPUs
DeepSeek released DeepJIT, a header-only C++20 JIT runtime for NVIDIA CUDA GPUs and Huawei Ascend NPUs, on GitHub. It provides kernel compilation, caching, loading, and launching with PyTorch integration. The source release was published on 2026-09-08.
China context
- Original name
- 深度求索
- Outside China
- Open weights · github.com
- Claims
- Company-reported; not yet independently evaluated
- For builders
- Developers outside China can integrate DeepJIT into their C++/Python extensions to support both CUDA and Ascend backends with a single JIT interface, potentially simplifying cross-platform kernel development.
- For investors
- DeepSeek's release of a tool supporting Huawei Ascend indicates continued investment in the Chinese AI hardware ecosystem, which may affect the competitive landscape for AI accelerator software.
DeepJIT is a lightweight, header-only C++20 JIT runtime that gives C++/Python extension authors a shared interface for compiling kernel source at runtime, caching binaries, loading them onto devices, and launching them with backend-specific options. It supports both NVIDIA CUDA GPUs and Huawei Ascend NPUs, with features including kernel caching, distributed filesystem support, lazy initialization, PyTorch integration, and compilation controls. The project is in development, with cache warmup and a Python compilation API listed as work-in-progress.
DeepJIT abstracts JIT compilation across CUDA and Ascend backends, using a unified runtime configuration, source and include hashing, in-memory and on-disk caches, and lazy initialization. It requires CUDA headers 12.4+, NVCC 12.9+, and PyTorch with CUDA support for the CUDA backend; for Ascend, it needs CANN with bisheng, ld.lld, and Ascend advapi headers. The cache supports distributed filesystems with atomic directory rename and fsync, enabling shared caches across users and nodes.
Developers building cross-vendor kernel libraries can now use one JIT interface for both NVIDIA and Huawei hardware, reducing the cost of maintaining separate compilation paths. This matters for teams targeting Chinese AI accelerators alongside CUDA, as it lowers the integration effort for PyTorch extensions.
DeepJIT could reduce development time for teams supporting both CUDA and Ascend, potentially accelerating deployment of AI models on Huawei hardware. However, the project is new and its production readiness is unproven.
The project's roadmap includes cache warmup from history and a Python compilation API, both marked as under development. Verification of actual adoption can be tracked through GitHub stars, forks, and issues on the DeepJIT repository.