[GH-ISSUE #332] ​​Proposal: Create DeepEP-Ascend Extension for Seamless NPU Integration While Exploring Future Upstream Path​​

bsakaguchi commented

2026-10-26 09:57:05 +00:00

Owner

Originally created by @Yael-X on GitHub (Jul 28, 2025).
Original GitHub issue: https://github.com/deepseek-ai/DeepEP/issues/332

Official Ascend Backend for DeepEP

1. Background & Motivation

We (Huawei Ascend Ecosystem Team) have successfully implemented an Ascend-native EP communication backend that fully complies with DeepEP’s low-latency mode API. This adaptation enables end-to-end Expert Parallelism (EP) for DeepSeek-V3 inference on Ascend clusters via the SGLang framework. And we are simultaneously adapting DeepEP’s normal mode for the prefill phase, which optimizes long-sequence processing by maximizing computational throughput while maintaining strict API compatibility.

Urgent Industry Needs:

🔥 Ascend NPU users (e.g., ByteDance, Tencent) require unified EP solutions to avoid fragmented forks
⚡️ SGLang’s moe model performance now depends on DeepEP-like optimizations, yet current DeepEP only supports NVIDIA GPUs

2. Gradual Integration Proposal

To balance immediate needs and architectural integrity:

✅ Step 1: Create an official DeepEP-Ascend repo under deepseek-ai, hosting our Ascend-optimized backend (API-compatible with DeepEP).
✅ Step 2: Jointly design abstract multi-backend interfaces, enabling seamless upstream merging to DeepEP main.

3. Why Start with Dedicated Repo?

🔧 Minimal Disruption: Isolate Ascend-specific logic from NVIDIA-centric CI pipelines
🚀 Accelerated Adoption: Immediate support for validated SGLang/Ascend workloads
🔜 Upstream Alignment: Refine APIs via community feedback

4. Call to Action

We request DeepEP maintainers to:

Schedule a meeting to discuss collaboration, or email us directly at xuyue32@huawei.com.
Create DeepEP-Ascend repo in deepseek-ai (we will submit full PR immediately upon creation)
Designate technical liaison to coordinate Abstract backend interface design.

5. Huawei’s Commitments

Responsibility	Details
👥 Dedicated Maintainers	2+ senior engineers
⚙️ CI/CD Resources	Huawei Cloud Ascend testbeds for nightly builds

Originally created by @Yael-X on GitHub (Jul 28, 2025). Original GitHub issue: https://github.com/deepseek-ai/DeepEP/issues/332 ## Official Ascend Backend for DeepEP ### 1. Background & Motivation We (Huawei Ascend Ecosystem Team) have successfully implemented an **Ascend-native EP communication backend** that fully complies with DeepEP’s low-latency mode API. This adaptation enables end-to-end **Expert Parallelism (EP) for DeepSeek-V3** inference on Ascend clusters via the SGLang framework. And we are simultaneously adapting DeepEP’s normal mode for the prefill phase, which optimizes long-sequence processing by maximizing computational throughput while maintaining strict API compatibility. **Urgent Industry Needs**: > 🔥 Ascend NPU users (e.g., ByteDance, Tencent) require unified EP solutions to avoid fragmented forks > ⚡️ SGLang’s moe model performance now depends on DeepEP-like optimizations, yet current DeepEP only supports NVIDIA GPUs ### 2. Gradual Integration Proposal To balance immediate needs and architectural integrity: > ✅ **Step 1:** Create an official `DeepEP-Ascend` repo under `deepseek-ai`, **hosting our Ascend-optimized backend** (API-compatible with DeepEP). > ✅ **Step 2:** Jointly design **abstract multi-backend interfaces**, enabling seamless upstream merging to DeepEP main. ### 3. Why Start with Dedicated Repo? - 🔧 **Minimal Disruption**: Isolate Ascend-specific logic from NVIDIA-centric CI pipelines - 🚀 **Accelerated Adoption**: Immediate support for validated SGLang/Ascend workloads - 🔜 **Upstream Alignment**: Refine APIs via community feedback ### 4. Call to Action We request DeepEP maintainers to: 1. **Schedule a meeting** to discuss collaboration, or email us directly at `xuyue32@huawei.com`. 2. **Create `DeepEP-Ascend` repo** in `deepseek-ai` (we will submit full PR immediately upon creation) 3. **Designate technical liaison** to coordinate Abstract backend interface design. ### 5. Huawei’s Commitments | Responsibility | Details | |-------------------------------|------------------------------------------| | 👥 **Dedicated Maintainers** | 2+ senior engineers | | ⚙️ **CI/CD Resources** | Huawei Cloud Ascend testbeds for nightly builds |

bsakaguchi closed this issue

2026-10-26 09:57:06 +00:00

bsakaguchi commented

2026-10-26 09:57:08 +00:00

Author

Owner

@sphish commented on GitHub (Jul 29, 2025):

Why not place it under Ascend, just like https://github.com/Ascend/pytorch?

@sphish commented on GitHub (Jul 29, 2025): Why not place it under Ascend, just like https://github.com/Ascend/pytorch?

bsakaguchi commented

2026-10-26 09:57:08 +00:00

Author

Owner

@Yael-X commented on GitHub (Jul 30, 2025):

Why not place it under Ascend, just like https://github.com/Ascend/pytorch?

@sphish Thank you for the suggestion! As @LyricZhao recommended in #269, we propose hosting DeepEP-Ascend under deepseek-ai (not Ascend/), adopting the proven decoupled repository model of vllm-ascend. This approach:

Core Benefits

👀 Unified Community Visibility
→ Centralizes support for DeepEP's users
→ Avoids ecosystem fragmentation from org-specific forks
🛡️ DeepEP Integrity Preservation
→ Isolates Ascend-specific logic from NVIDIA-centric CI pipelines
→ Maintains single-source truth for EP semantics

Why deepseek-ai > Ascend/?

vllm-ascend (hosted by vllm-project) demonstrates higher adoption and cleaner abstraction than hardware-vendor-owned repos.
DeepEP interfaces deserve shared community stewardship — a goal best realized under deepseek-ai, where hardware innovations and core standards evolve together.

Let’s collaborate! ✨
We’d welcome a technical discussion via email (948800153@qq.com) or synchronous meeting to align on next steps.

@Yael-X commented on GitHub (Jul 30, 2025): > Why not place it under Ascend, just like https://github.com/Ascend/pytorch? @sphish Thank you for the suggestion! As @LyricZhao [recommended in #269](https://github.com/deepseek-ai/DeepEP/issues/269), we propose hosting **`DeepEP-Ascend` under `deepseek-ai`** (not `Ascend/`), adopting the proven decoupled repository model of [`vllm-ascend`](https://github.com/vllm-project/vllm-ascend). This approach: **Core Benefits** 1. **👀 Unified Community Visibility** → Centralizes support for DeepEP's users → Avoids ecosystem fragmentation from org-specific forks 2. **🛡️ DeepEP Integrity Preservation** → Isolates Ascend-specific logic from NVIDIA-centric CI pipelines → Maintains single-source truth for EP semantics **Why `deepseek-ai` > `Ascend/`?** 1. vllm-ascend (**hosted by vllm-project**) demonstrates **higher adoption** and **cleaner abstraction** than hardware-vendor-owned repos. 2. DeepEP interfaces **deserve shared community stewardship** — a goal best realized under **deepseek-ai**, where hardware innovations and core standards evolve together. **Let’s collaborate!** ✨ We’d welcome a technical discussion via **email (948800153@qq.com)** or synchronous meeting to align on next steps.

bsakaguchi commented

2026-10-26 09:57:08 +00:00

Author

Owner

@sphish commented on GitHub (Aug 5, 2025):

Thanks a lot for your proposal and for the work you’ve put in!

After some discussion, we don’t think it’s a good idea to host the Ascend EP backend under deepseek-ai. Mainly because:

The EP backend for Ascend is totally different from the NVIDIA one, so there’s no real need for unified maintenance.
DeepEP’s interface is already quite stable. If you want to adapt it for Ascend, it’s straightforward for your team to do it independently.
Deepseek-ai isn’t an open community like vllm; putting Ascend team’s code here might be confusing.

We’re happy to keep discussing APIs and ways to work together, but it’s better for now to keep things separate.

@sphish commented on GitHub (Aug 5, 2025): Thanks a lot for your proposal and for the work you’ve put in! After some discussion, we don’t think it’s a good idea to host the Ascend EP backend under deepseek-ai. Mainly because: 1. The EP backend for Ascend is totally different from the NVIDIA one, so there’s no real need for unified maintenance. 2. DeepEP’s interface is already quite stable. If you want to adapt it for Ascend, it’s straightforward for your team to do it independently. 3. Deepseek-ai isn’t an open community like vllm; putting Ascend team’s code here might be confusing. We’re happy to keep discussing APIs and ways to work together, but it’s better for now to keep things separate.

bsakaguchi commented

2026-10-26 09:57:08 +00:00

Author

Owner

@Yael-X commented on GitHub (Aug 7, 2025):

Thank you for your thoughtful review and clear feedback. We fully respect DeepEP’s architectural direction and appreciate your commitment to maintaining repository independence. While we’re naturally a bit disappointed that a unified effort under deepseek-ai isn’t feasible, we completely support your decision to preserve the integrity of your project. We remain committed to advancing the EP ecosystem alongside you—albeit on separate tracks. Your guidance has been truly appreciated. 🙏

@Yael-X commented on GitHub (Aug 7, 2025): Thank you for your thoughtful review and clear feedback. We fully respect DeepEP’s architectural direction and appreciate your commitment to maintaining repository independence. While we’re naturally a bit disappointed that a unified effort under deepseek-ai isn’t feasible, we completely support your decision to preserve the integrity of your project. We remain committed to advancing the EP ecosystem alongside you—albeit on separate tracks. Your guidance has been truly appreciated. 🙏

Rows
Columns

[GH-ISSUE #332] ​​Proposal: Create DeepEP-Ascend Extension for Seamless NPU Integration While Exploring Future Upstream Path​​ #234

Official Ascend Backend for DeepEP

1. Background & Motivation

2. Gradual Integration Proposal

3. Why Start with Dedicated Repo?

4. Call to Action

5. Huawei’s Commitments

[GH-ISSUE #332] Proposal: Create DeepEP-Ascend Extension for Seamless NPU Integration While Exploring Future Upstream Path #234