TileLang is an open-source domain-specific language for writing high-performance GPU and accelerator kernels in Python โ the kind of attention, GEMM and convolution kernels that decide how fast AI models actually run. The project has collected more than 8,700 stars on GitHub and was actively developed through 2026.
Writing fast kernels normally means dropping into CUDA and managing threads, shared memory and tensor cores by hand. TileLang raises the abstraction level: you describe the computation in tiles with a Pythonic syntax, and the compiler handles layout, pipelining and code generation for NVIDIA and AMD GPUs and other accelerators. It has been used to express state-of-the-art attention kernels (including flash-attention-style and linear-attention variants) with far less code than hand-written CUDA, and is maintained by the Tile-AI community with roots in the Apache TVM ecosystem. Because kernels are written once at the tile level, the same TileLang program can be retargeted across GPU generations instead of being rewritten for each new architecture, and the project is increasingly used inside inference engines that need custom operators for new model architectures.
TileLang is free and open source, installable from PyPI; you supply your own GPU hardware or cloud instances.
TileLang programs are compiled just in time and cached, and the project maintains a gallery of reference kernels โ including matrix multiplication and several attention variants โ that double as performance baselines when bringing up new hardware or comparing against hand-tuned libraries.
It is for ML systems engineers, inference-framework developers and researchers who need custom, fast kernels but want to work at a higher level than raw CUDA.