DeepSeek Open-Sources Ascend Infrastructure Components, Covering Compilation, Computing, and Communication Libraries

AI company DeepSeek has officially open-sourced infrastructure components for Huawei's Ascend computing platform, covering the TileLang high-level language compilation tool, computing libraries, and distributed communication libraries, corresponding one-to-one with its previously open-sourced components for the Nvidia platform. TileLang aims to provide a general-purpose, easier-to-program high-level language that reaches hardware performance limits, improving development efficiency and simplifying logic relative to CUDA, while its programming model also leverages chip-specific features. The TileLang roadmap was first validated on the Nvidia platform and has supported the implementation of most operators in DeepSeek V4 series model training.

The newly open-sourced Ascend version encapsulates Ascend C underlying instructions, providing a high-level language programming approach without sacrificing hardware performance. Currently, every TileLang operator used in DeepSeek training has a corresponding high-performance implementation on Ascend. Other components open-sourced in the same batch include DeepGEMM, DeepEP, TileKernels, FlashMLA, and DeepSelect, used for general matrix operations, large-scale cross-device communication, routine vector computation and memory access, long-context sparse attention, and data filtering, respectively.

DeepSeek stated that in multiple key tests, related computing and communication performance has approached hardware limits; during development, the Huawei team provided support, and both parties collaborated to advance a 128-card supernode solution based on Ascend 950, with deep optimizations made to computing and communication.

[ChainCatcher]