DeepSeek open-sources DeepEP-Ascend communication library for Huawei Ascend NPUs
DeepSeek published DeepEP-Ascend, a high-performance communication library for machine-learning training and inference on Huawei Ascend NPUs.
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 use these libraries to optimize MoE training and inference on Huawei Ascend NPUs, with APIs aligned to NVIDIA DeepEP and DeepGEMM.
- For investors
- The release indicates DeepSeek's commitment to Huawei's Ascend ecosystem, increasing the viability of Ascend-based AI infrastructure for large-scale deployments.
Translated from Chinese. Quotes and facts link to the original sources.
The library implements expert-parallel communication for MoE dispatch and combine. Its README documents the hardware, firmware and test conditions used for reported performance.
Its public buffer APIs align with the NVIDIA DeepEP implementation, while Ascend kernels use the communication and compilation components documented by the project.
The repository adds an implementation that infrastructure teams can examine when evaluating Ascend-based MoE training and inference.
Inspect the code and reproduce a workload on compatible Ascend hardware before comparing communication overhead with the existing stack.
Watch performance on intended deployment hardware, firmware compatibility and progress on the communication primitives marked as work in progress.