What problem does it solve? When optimizing diffusion model inference on MindIE-SD, teams often jump straight into tuning without knowing which component (DiT compute, DiT communication, VAE decoding, or host overhead) is actually the bottleneck, leading to wasted effort on the wrong target. This Skill acts as the optimization domain entry point: it takes an already-identified bottleneck label and dispatches the task to the correct specialized optimization module, while enforcing prerequisites and acceptance criteria. ## Core Features & Use Cases - Bottleneck-based dispatch: Maps bottleneck labels (DiT compute-bound, DiT communication-bound, non-DiT decode stage, non-DiT host stage, consistency failure) to the right module: dit-perf-opt, dit-parallel-opt, vae-opt, or host-opt. - Prerequisite gating: Requires a working environment, a completed end-to-end run, and baseline numbers before any optimization starts; tasks missing these are returned to model-auto-optimization. - Anchor enforcement: Rejects user-declared bottlenecks without a measured anchor (stage breakdown row or kernel percentage) to prevent optimizing the wrong component. - Unified acceptance criteria: Performance numbers follow perf-gate same-window A/B comparison, differences under 3% are treated as noise, and lossy features must pass accuracy-gate three-level verification. - Use Case: A profile shows MatMul kernels taking 62% of DiT time. The Skill maps this anchor to the DiT-compute-bound label and dispatches the task to dit-perf-opt for feature tier selection, with results validated via perf-gate. ## Quick Start Tell the AI your measured bottleneck anchor, for example: profile shows DiT MatMul at 62% with no exposed communication, and ask it to route the optimization task to the right module.
