TileLang / TileKernels: a tile DSL as the middle layer between hand-written CUDA and Triton

The point of a tile DSL is that the author writes tiles and the compiler decides the instruction selection, register/shared-memory allocation and async-copy pipelining. That means the same source can target NVIDIA SM90/SM100 and Huawei Ascend 950 with identical Python APIs, selected automatically at runtime - a portability claim no hand-written CUDA kernel can make. What it costs is the hardware ceiling: a DSL cannot emit an instruction the abstraction does not model, so the fastest path is still whatever the DSL exposes (here: FP8/FP4 block-scaled casts, fused SwiGLU+quant, MoE top-k routing and scoring, RoPE, fused RMSNorm in the Engram gate). Recorded as custom_kernel with class custom_kernel because the mechanism of interest is the compilation layer, not any single op.

Requirements are specific: Python 3.12+, PyTorch 2.13+, TileLang 0.1.15+, and either NVIDIA SM90/SM100 with CUDA >= 13.1 or Ascend 950 with CANN >= 9.2.0. DeepSeek ships this because FlashMLA's fused-norm-rope-attn-rope-cast kernel test depends on TileLang, TileKernels AND DeepGEMM - i.e. the modern stack is layered: DSL -> kernel library -> GEMM library. The claim 'most kernels achieve performance close to the hardware's compute or memory bandwidth limits' is DeepSeek's own and carries no numbers, so treat it as unquantified. Note also that the repo's own claim of 'dozens of highly optimized kernels' is what makes it a useful proxy for what a well-tuned tile DSL can and cannot reach. HARDWARE-PARAMETER ROWS ADDED 2026-10-05 (schema authority pass, agent w5-schema). `affected_by_hardware` names PRODUCTS (accelerator record ids); these rows name the hardware QUANTITIES that move this class's cost, which had nowhere to live until flop.hardware_parameters was added in this commit. Every effect below is lifted verbatim from this record's own existing prose and no figure here is new - `magnitude` reproduces numbers the record already quoted. `basis` is 'unverified' where the effect was asserted with nothing concrete behind it, rather than being dressed up. The rows were pulled out of free-text notes; they are not a change to what this record claims.