model-optimize

What problem does it solve?

This Skill provides a staged, disciplined approach to optimize models for NPU inference, enforcing staged validation between KVCache, fusion, and graph-mode adaptation, and coordinating with subagents to ensure traceable progress and quality.

Core Features & Use Cases

• KVCache静态化、FA替换、Cache metadata准备等阶段的有序执行
• 融合算子优化与非 Attention 模块聚焦
• Graph-mode adaptation is applied only in Decode stage, Prefill remains eager
• Use Case: Single-card optimization to produce auditable reports and progress records

Quick Start

Start Stage 0 by analyzing the target model and creating a baseline optimization plan