DeepSeek has released open-source infrastructure components specifically designed for Huawei Ascend computing platforms. The new suite includes the TileLang high-level programming language compiler, alongside specialized computing and distributed communication libraries.
The company stated that these tools mirror the functionality of their existing open-source components for Nvidia platforms. TileLang is designed to simplify programming compared to CUDA, while enabling developers to reach the hardware performance limits of the Ascend chips.
The release features several core modules: DeepGEMM for matrix operations, DeepEP for large-scale cross-device communication, and TileKernels for vector computation. Additionally, it includes FlashMLA for sparse attention and DeepSelect for data filtering.
According to DeepSeek, these components have been validated through their own model training processes. The company reports that in key test cases, the performance of these tools on Ascend hardware is close to the theoretical maximum.
