TileLang

TileLang is an open-source domain-specific language (DSL) built on TVM that enables developers to write high-performance GPU/CPU kernels (e.g., GEMM, FlashAttention) using Pythonic syntax with explicit tile-level primitives for memory and parallel scheduling. It reduces kernel implementation code by up to 90% compared to manual CUDA/HIP implementations and achieves up to 5–6× speedup over Triton, with adoption by frontier labs such as DeepSeek.