DeepSeek open-sources full Ascend infra stack: TileLang and five components mirroring its NVIDIA tooling
机器之心 · wechat · 2026-09-30
DeepSeek has open-sourced its infrastructure components for Huawei Ascend, including TileLang (high-level kernel language), DeepGEMM-Ascend, DeepEP-Ascend, TileKernels, FlashMLA and DeepSelect — a one-to-one counterpart to its previously released NVIDIA stack.
Key facts:
- TileLang simplifies kernel programming versus CUDA/AscendC; it already powers most kernels in DeepSeek V4 training, and every TileLang kernel used in training now has a high-performance Ascend counterpart.
- Joint optimization on Ascend 950 and the 128-card UBL128 supernode: DeepEP covers EP/CP/PP/FSDP parallelism, with Dispatch bandwidth at 375GB/s and Combine at 347GB/s, near hardware limits.
- Huawei open-sourced joint results in the CANN community — deployment, large-scale training, long-context KVCache pooling, and AgenticRL — and disclosed DeepSeek-V4.1-Flash offline inference throughput under EP32, 128K context, and 0.85 speculative acceptance rate.
This gives developers reusable, battle-tested components from real frontier-model training and signals progress for domestic AI compute.
Related event: DeepSeek Open-Sources Full Ascend Toolchain, Taking Aim at CUDA(17 posts)→
More from Infra
- Mobile AI Factory: agents auto-build on-device SLMs with 19x less GPU time — pmttyji · 2026-11-16
- Stratechery's Fictional Obituary: Imagining Jensen Huang's 2026 Exit From Nvidia — Gregory_C_Allen · 2026-11-16
- vLLM serves DeepSeek V4.1-Flash at 954K tokens/s per GPU, $38.96/GPU-hour on GB300 NVL72 — bookwormengr · 2026-11-16
- Colossus 2's 1.2M B200s could generate ~$400B/year in tokens at DeepSeek Flash rates — bookwormengr · 2026-11-16
- BIS issues rule blocking remote access to US AI chips via third-country data centers — teortaxesTex · 2026-11-16
- Qdrant 1.20.1: 50% fewer cold-storage reads, ~2x faster quantization, agent-ready setup — qdrant_engine · 2026-11-16