What problem does it solve? Enabling PyTorch operators on an accelerator that is not CUDA-compatible requires hand-writing hundreds of vendor kernel bindings, which is intractable. This Skill provides a category-based codegen approach that reduces the work to a handful of templates plus a one-line-per-operator mapping table. ## Core Features & Use Cases - Category-based kernel generation: Groups operators into categories (unary, binary, reduce, gemm) so one template covers many operators, modeled on the Ascend ACLNN backend that reached 138 operators from 63 categories. - Vendor support layer guidance: Walks through writing the op_api_common/op_preparation headers, the EXEC macro hiding workspace query and execute sequences, and dtype enum mapping. - Backend slot and routing setup: Covers adding the k<Vendor> enum, the conf file that routes ops at runtime, and CMake gating via FLAGOS_BUILD_VENDOR. - Use Case: After runtime bring-up passes on a new Ascend-like chip and CUDA compatibility has been ruled out by measurement, use this Skill to generate kernels, wire routing, and verify each operator against CPU with per-op comparison tests. ## Quick Start Enable the unary operator category on my non-CUDA-compatible accelerator by generating native vendor kernels and verifying them against CPU comparison tests.
