What problem does it solve? Decode-stage inference on Ascend NPU suffers from high task scheduling overhead because each step launches many small operators. This Skill guides the adaptation of SuperKernel operator binary fusion, which merges operators within a marked scope to reduce kernel launch and scheduling costs. ## Core Features & Use Cases - Structured Four-Step Workflow: Analyzes model structure, determines fusion scope (Attention-only, MoE-only, full Decoder, or custom), modifies config.yaml and modeling code with superkernel_scope markers, and runs mandatory baseline-vs-optimized performance verification. - Constraint Enforcement: Validates prerequisites before acting—Atlas A3 hardware, ge_graph execution mode, PyTorch framework, and decode-only activation—while flagging dynamic shapes, Python control flow, and unsupported Tiling operators. - Use Case: A user running DeepSeek-R1 inference on Atlas A3 notices high decode latency. The Skill analyzes the MoE architecture, marks the MoE expert computation with superkernel_scope, enables enable_superkernel in config.yaml, and produces a superkernel_optimization_report.md with measured latency and throughput improvements. ## Quick Start Enable SuperKernel binary fusion for my model's decode stage and verify the performance improvement against a baseline.
