What problem does it solve? Developing high-performance fused operators (Attention, MatMul variants, Norm, Sort/TopK) for Ascend NPUs requires coordinating block-level task partitioning, tile-level implementation, and measurement-driven performance tuning, which is error-prone when done ad hoc. ## Core Features & Use Cases - Orchestrated Design Workflow: Coordinates design methodology and coding-convention references to produce block_level/ designs, tile_level/ TileLang kernels, and a model_new_tilelang.py implementation from a PyTorch reference model. - Mandatory Pattern Routing: Enforces gated checklists for Attention, reduction, shuffle, Sort/TopK, and Norm-fusion operator families before any design code is written. - Verification and Performance Iteration: Ships scripts for TileLang accuracy verification, PyTorch-regression static detection, and msprof-based performance tuning with a geomean >= 0.6x target versus PyTorch reference kernels. - Use Case: Given a BatchMatmulMaxSum PyTorch model, produce a fused single-kernel TileLang design with CV-fusion pipelining, pass accuracy validation, and iterate on performance before handing off to the AscendC translator. ## Quick Start Run this skill on my output_dir containing model.py to generate the block-level and tile-level TileLang design plus model_new_tilelang.py for my fused operator.
