What problem does it solve? When deploying diffusion models through third-party inference frameworks (vLLM-Omni, DiffSynth-Engine, LightX2V, diffusers) on Ascend NPUs, framework-side features such as quantization, sparsity, caching, and kernel fusion often fail to take effect or are entirely missing. This Skill provides a two-branch workflow to either enable and verify existing framework features (Branch A) or develop missing structural capabilities (Branch B), with rigorous evidence requirements to prevent false speedup claims. ## Core Features & Use Cases - Feature Enablement & Verification (Branch A): Diagnose why quantization/sparsity/cache switches show no effect, fix enablement anomalies, and validate with a counting contract plus three-layer evidence (graph hits, kernel diff, wall-clock A/B). - Gap-Filling Development (Branch B): Add missing structural capabilities to frameworks via injection points (config registries, operator dispatch, model-layer rewrites, _compiled_call_impl) with four merge postures from platform registration to monkey-patch probes. - Use Case: A user reports that enabling torch.compile with MindieSDBackend in vLLM-Omni shows zero pattern hits. The Skill guides checking _compiled_call_impl assignment, backend instance reuse, and kernel-level verification via kernel_details.csv. ## Quick Start Ask the agent to enable and verify a mindiesd feature such as cache or quantization in a specific framework like vLLM-Omni and confirm it with kernel-level evidence.
