DeepSeek-V4 post-training on Ascend SuperPOD reaches 34.22% MFU
Dongfang Li · hf · 2026-07-23
What it reports
This paper describes an end-to-end post-training optimization stack for the DeepSeek-V4 family on an Ascend NPU SuperPOD.
- It tackles trillion-parameter MoE post-training bottlenecks such as memory pressure, communication overhead, and inefficient kernels.
- The authors build a hierarchical optimization framework across parallelism, compute/communication orchestration, and kernel execution.
- The system reaches 34.22% MFU, a 2.93× improvement over the open-source baseline, while keeping training stable.
- On top of that infrastructure, they create a CPT/SFT workflow for complex Operations Research tasks using solver-verified synthetic data.
- The resulting specialized model gets 71.81% zero-shot Pass@1, beating GPT-5.4-Mini by 3.98 points and the base DeepSeek-V4-Flash by 11.27 points.
Related event: Ascend SuperPOD Achieves 34.22% MFU for DeepSeek-V4 Training(3 posts)→
More from Infra
- Open-source Latch compiles PyTorch into training megakernels, 2.75x faster on H100 — sguada · 2026-10-27
- NVIDIA Claims HGX Rubin NVL8 Delivers 10x More Tokens per Megawatt Than B200 — thursdai_pod · 2026-10-27
- Commoditization dilemma: either human labor or GPUs become the commodity — _AustinCalvert_ · 2026-10-27
- DeepSpeed shows AutoEP/AutoTP/AutoSP parallelism gains, hosts office hours — PyTorch · 2026-10-27
- Running a local 9B as a coding agent: brief the big model to plan, let the small model execute — Electrical_Ninja3805 · 2026-10-27
- Cloudflare Durable Objects validated as top architecture for AI agent WebSocket servers — tristanbob · 2026-10-27