DeepSeek releases TileKernels: dozens of LLM kernels near hardware limits, now on Ascend
rohanpaul_ai · x · 2026-10-01
The companion GitHub repo to DeepSeek's Ascend open-sourcing: TileKernels is a library of dozens of highly optimized kernels written in TileLang, covering MoE routing, Engram, quantization and manifold hyper-connections (mHC) for LLM training and inference.
- Most kernels approach the hardware's compute or memory bandwidth limits, and all are already used in DeepSeek's internal workloads.
- A 2026-09-30 update adds Huawei Ascend support: following the NVIDIA path, kernels ship a second backend selected automatically at runtime, so the same Python APIs run on both NVIDIA GPUs and Huawei NPUs.
- MIT licensed, 1.9k stars.
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