TileLang README and autotuning guide (tile-ai/tilelang)

TileLang 'is a concise domain-specific language designed to streamline the development of high-performance GPU/CPU/NPU kernels (e.g., GEMM, Dequant GEMM, FlashAttention, LinearAttention). By employing a Pythonic syntax with an underlying compiler infrastructure on top of TVM'. Two facts that make it interesting for a planner rather than another Triton clone. (1) It is the 2025-2026 DSL that crossed from research into a shipped stack: the repo's own dated news records an Ascend 950 backend with 'native code generation, automatic scheduling and synchronization' (2026-09-30), a multi-backend language dialect over static CUDA, ROCm and Metal dialects (2026-07-24), Metal 4 cooperative-tensor GEMM for Apple M5 (2026-07-28), and SM120 NVF4 block-scaled MMA (2026-07-30). (2) Autotuning is real and built in: '@tilelang.autotune(configs=...) stacked on @tilelang.jit' searches a config space, 'compiles candidates in parallel, validates correctness, benchmarks them, and caches the best result for reuse', with warmup/rep/timeout arguments. The caching behaviour is the deployment-relevant part - tuning is paid once per config and reused. Carries the DeepSeek TileKernels dependency noted in the repo's existing flops/tilelang record. Latest release v0.1.15; MIT, but the LICENSE records 'additional collaboration terms with Microsoft Corporation' for 2024-12-01 to 2025-03-14, a window that has passed.