DeepEP-Ascend
DeepEP-Ascend is a communication library for machine learning training and inference on Huawei Ascend NPUs. It provides expert-parallel operations for MoE dispatch and combine, alongside additional buffer interfaces, some of which are experimental.
What it does
Developers can use it to coordinate communication between Ascend NPU ranks in MoE training and inference workloads.
- Availability
- Installation instructions are provided for Linux on supported Ascend hosts; validated stack uses Ascend 950DT and CANN 9.2.0
- Pricing model
- Not established
- Deployment
- Linux on an Ascend host with supported NPU drivers and firmware
- Privacy
- Not established
- Licence
- Not established
Capability
Modality
Access
Deployment
Pricing
Target user
Evidence
5 claims- The library provides expert-parallel all-to-all operations for MoE dispatch and combine. [1]
It provides expert-parallel (EP) all-to-all operations for MoE dispatch and combine, including FP8 dispatch and deferred epilogues.
GitHub new repositories. Quoted text is outside CC BY 4.0 and keeps its original rights. - EP dispatch supports FP8, and combine supports BF16. [1]
It provides expert-parallel (EP) all-to-all operations for MoE dispatch and combine, including FP8 dispatch and deferred epilogues.
GitHub new repositories. Quoted text is outside CC BY 4.0 and keeps its original rights. - Bucket batched all-gather is available; Ascend reduce-scatter and all-reduce kernels are still being built. [1]
- Bucket collectives: batched all-gather is available; Ascend reduce-scatter and all-reduce kernels are still being built.
GitHub new repositories. Quoted text is outside CC BY 4.0 and keeps its original rights. - Installation is documented using Git clone and pip on Linux with an Ascend host. [1]
- The validated stack includes Ascend 950DT and CANN 9.2.0. [1]
The validated stack is Ascend 950DT, CANN 9.2.0, Python 3.12, PyTorch 2.13.0+cpu and torch_npu 2.13.0rc1, with the PoC HDK configuration described below.
GitHub new repositories. Quoted text is outside CC BY 4.0 and keeps its original rights.
Sources
1
1
GitHub new repositoriesOwner-controlled
https://github.com/deepseek-ai/DeepEP-Ascend
Retrieved 2026-10-29 00:08 UTC
Verified
Versions
2| Version | Effective | Change | Changed fields | State |
|---|---|---|---|---|
| 2 | 2026-10-03 01:44 UTC | material change | availability, deployment | Current |
| 1 | 2026-10-02 02:38 UTC | publication | None | Current |
In briefs
2- Backburner leads 25 AI developments in week 40, 2026 Weekly brief 2026-W40
- Google introduces Gemini 4 Argon reasoning model Daily brief 2026-10-02
Original catalog content is licensed under CC BY 4.0, attributed to wowsuchai. Source quotations and linked content keep their original rights.