TileLang

TileLang has been validated and proven mature on the NVIDIA platform, and carries the implementation of most operators in the training of the DeepSeek V4 series models.

DeepSeek has open-sourced infrastructure components for the Ascend platform, covering the TileLang compiler, compute library, and communication library.

On September 30, 2025, DeepSeek open-sourced a batch of underlying infrastructure components, including operators, a programming language, compute libraries, and communication libraries

CUDA Tile并非要取代现有的CUDA编程方式，而是在其基础上提供更高层次的选择

TileLang is a domain-specific language (DSL) for tensor computing that lets operator developers write one piece of code and compile native implementations for different hardware backends, sitting between low-level programming models like CUDA/HIP and high-level frameworks like PyTorch.

In DeepInfer, Ascend and NVIDIA are first-class citizens, with 30 TileLang operators on each side corresponding one-to-one, plus Ascend versions of DeepGEMM and DeepEP.

For other domestic chip vendors (Cambricon, Hygon, Kunlunxin, Moore Threads, MetaX, Biren, Enflame, and Iluvatar CoreX) to fully adapt DeepInfer, they would roughly need to overturn their existing engines and fill in about 30 TileLang backend operators, FP8×FP4 GEMM fusion, Mega MoE, sparse attention, and TopK components.

TileLang targets GPUs, CPUs, and various accelerators simultaneously, supporting operator logic reuse across hardware platforms.

TileLang has made dedicated optimizations for the operator requirements of the large-model era, such as Flash Attention variants and MoE routing.

Compared to Triton, TileLang operates at a more explicit level of abstraction, emphasizing exposing blocking strategies explicitly to developers in order to preserve fine-grained control space beyond automatic optimization.

TileLang's most direct application scenarios are concentrated on core deep learning operators such as FlashAttention, GEMM (General Matrix Multiplication), and quantization and dequantization.

TileLang is a domain-specific language (DSL) that aims to simplify high-performance kernel development targeting GPUs, CPUs, and various accelerators, using Python as the development language and tile-based blocking as its core abstraction.

TileLang's GitHub repository address is tile-ai/tilelang.

TileLang's compilation toolchain automatically translates high-level computation descriptions into corresponding TBE primitives or AscendC code, and handles the timing and synchronization of data movement.

TileLang is positioned similarly to Triton in the CUDA ecosystem or TVM/TIR in AI compilers, seeking to strike a balance between development efficiency and generating high-performance code.

The supporting resources developed around TileLang for Ascend include the TileLang main repository, the TileKernels operator library, an Ascend-oriented operator library, LearningHub development courses, and AI-assisted development tools.

TileLang's support for the Ascend 950 makes it possible to rapidly validate new operator ideas and test performance on Ascend hardware.

TileLang is a compiled DSL whose code goes through stages including frontend parsing, intermediate representation optimization, and backend code generation, ultimately producing low-level code that can run on specific hardware.

The TileLang development workflow for Ascend is roughly divided into three steps: expressing computation logic and data movement in TileLang, compiling via the toolchain to generate code targeting the Ascend 950, and running and benchmarking on the hardware.

TileLang uses Python-style syntax to express computation logic and data movement, which is then mapped to Ascend hardware through a compilation toolchain.