What problem does it solve? Once profiling has identified a DiT compute bottleneck in a diffusion model, engineers still face the hard question of which optimization tier to enable, at what setting, and how to prove it actually took effect. This Skill turns a located DiT compute bottleneck into a concrete, verifiable feature-tier combination for MindIE-SD on Ascend NPU. ## Core Features & Use Cases - Tier Selection: Chooses concrete quantization tiers (W8A16, W4A16, W8A8 series, W4A4, MXFP8, FA quantization), sparsity (rf_v2, ada_bsa), caching (DiTCache, AttentionCache, timestep optimization), and compilation options (MindieSDBackend, pattern fusion, ACLGraph) based on docs/zh/features/* as the single source of truth plus the framework support matrix. - Enablement Verification: Enforces the rule that "enabled does not mean effective" by requiring graph-hit, kernel-diff, or feature-active-count evidence, and treats byte-identical output to the lossless baseline as proof of non-activation rather than near-zero gain. - Combination Search Protocol: Provides a disciplined protocol for combining two or more lossy dimensions, including seam conflict checks, mandatory pairwise and ternary test coverage, single-variable stacking, frontier retention, and layer fallback. - Use Case: A profiling report shows MatMul takes 58% of DiT time with a quantization direction. The Skill looks up quantization.md for the API, checks the support matrix, selects W8A8_MXFP8, verifies activation with kernel evidence, and revalidates with a same-window A/B run. ## Quick Start Ask the assistant to pick and enable the right quantization, sparsity, or cache tier for an already-located DiT compute bottleneck and verify it actually took effect.
