tile-ai/tilelang +839 ★ this week + Follow

What it actually is

TileLang (tile-ai/tilelang) is a Python-embedded domain-specific language for writing GPU/CPU/accelerator kernels — GEMM, dequantized GEMM, FlashAttention, linear attention and similar workloads — without hand-writing CUDA/HIP/assembly. It compiles down through an Apache TVM-based infrastructure, exposing a Pythonic kernel-authoring surface (T.gemm, T.gemm_sp, T.copy, T.print, kernel/thread-binding primitives in tilelang/language/kernel.py) backed by a JIT pipeline (tilelang/jit/__init__.py: compile, tilelang/jit/kernel.py: get_profiler) and a benchmarking harness (tilelang/profiler/bench.py: do_bench). It targets practitioners building high-performance kernels for LLM/attention workloads who want tensor-core-level performance from a few hundred lines of Python.

Why it's growing

The 839-star jump in a week coincides with a dense, continuous release cadence (v0.1.11 through v0.1.15 between June and September 2026, roughly monthly) and active feature expansion rather than a single viral event — recent commit titles show new hardware backend support (Ascend 950, ROCm GLM-5.3, Metal 32-bit atomics, SM120 block-scaled GEMM) landing within the last 30 days. The README's running "Latest News" log (CuTeDSL backend, Z3 theorem-prover integration, Apple Metal support, Ascend/NPU backends) suggests sustained multi-backend ambition, which plausibly explains organic interest, though the provided facts don't directly confirm what triggered this specific week's spike.