What problem does it solve? Optimizing a diffusion model hosted by a third-party framework on Ascend NPUs involves many disjoint steps—environment setup, kernel fusion, multi-card parallelism, quantization/sparsity/cache tuning, and final verification—which are easy to run out of order or declare done without evidence. This Skill provides a single orchestration entry point that routes each stage (S0–S6) to the right capability skill, enforces stage gates, and mandates evidence-backed closure reports. ## Core Features & Use Cases - Staged routing (S0–S6): Routes tasks through environment preparation, DiT kernel fusion, multi-card parallelism (USP/CP/TP), lossy optimization (quantization, sparsity, cache), training-aware distillation, and VAE/host optimization, each with defined acceptance criteria. - Run-state and stage gates: Maintains a single source of truth in run-state.md and blocks stage advancement via scripts/stage_gate.py until evidence checks pass with zero errors. - Mandatory dual reports and claim discipline: Enforces overview_report.md and detail_report.md with measured e2e baselines, step counts, quality gates, and anti-no-op verification counters before any speedup claim. - Use Case: Given a request like "optimize MiniMax-H3 on vLLM-Omni to 8bit + 80% sparsity + cache", the Skill confirms scope, runs S0 environment setup, executes fusion/parallel/lossy stages with per-stage evidence, and closes with audited overview and detail reports. ## Quick Start Ask the assistant to optimize a specific diffusion model on its hosting framework, for example: run the full optimization flow for MiniMax-H3 on vLLM-Omni starting from environment setup.
