Writing High-Performance Kernels in TileLang, from GEMM to MLA

TileLang is presented as a middle-ground GPU kernel framework: more explicit and controllable than Triton, but far simpler than CUTLASS/CuTe.

The post explains TileLang’s programming model around tiling, memory placement, layout inference, and software pipelining for efficient kernels.

It walks through building a high-performance GEMM kernel and then extends the ideas to DeepSeek’s MLA decode case.

The main takeaway is that TileLang gives kernel authors fine-grained performance control without forcing them into the full complexity of lower-level CUDA libraries.