DeepEP Ascend
An open-source large-scale cross-device communication component released by DeepSeek for the Huawei Ascend platform, used for Token Dispatch and Combine operations in MoE architectures.
Timeline (last 90 days)
The public API of DeepEP Ascend is fully aligned with the NVIDIA version of DeepEP, including EP Buffer, Dispatch, Combine, and the defer epilogue parameter, meaning code running on NVIDIA cards requires virtually no changes at the Python layer after being ported to Ascend.
DeepSeek ported the DeepEP communication library to the domestically produced Ascend NPU and open-sourced it on GitHub under the name DeepEP Ascend.
The Ascend version of DeepEP must use the expanded layout (Dispatch Expand) and currently operates as a single communication domain.
What is truly ready in DeepEP Ascend at present is the EP cross-card dispatch mechanism; APIs for other workload strategies such as PP, bucket, and ngram have been placeholder-defined but their implementations are not yet complete.
DeepEP Ascend achieves a sustained Dispatch bandwidth of 90%~95% of the physical payload limit on the Ascend 950, on the condition that the EP scale does not exceed 32; larger EP scales and Combine are still being optimized.
The open-source list includes TileLang, DeepGEMM Ascend, DeepEP Ascend, TileKernels, FlashMLA, and DeepSelect, all hosted on GitHub under the MIT license.
The DeepSeek community has open-sourced two underlying acceleration projects targeting the Ascend platform: DeepEP Ascend and DeepGEMM Ascend
DeepEP provides four types of buffers: EP Buffer (for expert-parallel Dispatch and Combine), PP Buffer (for pipeline-parallel Send/Receive), Ingram Buffer (for Table writes and per-layer Fetch), and Bucket Buffer (for bucketing operations)
In key training communication use cases, DeepEP can achieve communication bandwidth exceeding 90% of the theoretical maximum
All Facts (8)
The public API of DeepEP Ascend is fully aligned with the NVIDIA version of DeepEP, including EP Buffer, Dispatch, Combine, and the defer epilogue parameter, meaning code running on NVIDIA cards requires virtually no changes at the Python layer after being ported to Ascend.
50%UnverifiedThe Ascend version of DeepEP must use the expanded layout (Dispatch Expand) and currently operates as a single communication domain.
50%UnverifiedDeepEP Ascend achieves a sustained Dispatch bandwidth of 90%~95% of the physical payload limit on the Ascend 950, on the condition that the EP scale does not exceed 32; larger EP scales and Combine are still being optimized.
50%UnverifiedDeepEP provides four types of buffers: EP Buffer (for expert-parallel Dispatch and Combine), PP Buffer (for pipeline-parallel Send/Receive), Ingram Buffer (for Table writes and per-layer Fetch), and Bucket Buffer (for bucketing operations)
50%UnverifiedIn key training communication use cases, DeepEP can achieve communication bandwidth exceeding 90% of the theoretical maximum
50%UnverifiedDeepSeek ported the DeepEP communication library to the domestically produced Ascend NPU and open-sourced it on GitHub under the name DeepEP Ascend.
50%UnverifiedThe open-source list includes TileLang, DeepGEMM Ascend, DeepEP Ascend, TileKernels, FlashMLA, and DeepSelect, all hosted on GitHub under the MIT license.
50%UnverifiedThe DeepSeek community has open-sourced two underlying acceleration projects targeting the Ascend platform: DeepEP Ascend and DeepGEMM Ascend
50%