菜单
Adapt full-attention language models to sliding window attention without expensive retraining. Combine five synergistic strategies (full decode, sink tokens, interleaved layers, chain-of-thought, fine-tuning) achieving 30-100% speedups while maintaining 90-100% accuracy.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/ADu2021/skillXiv --skill sliding-window-attention-adaptation命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
Convert arXiv papers into ready-to-use agent skills using category-aware extraction. First classifies the paper into one or more of 11 research categories, then applies a specialized extraction pipeline for each category — because different types of papers produce different types of usable knowledge. A single paper can yield multiple skills if it spans categories. Use this skill whenever the user wants to turn a paper into a skill, extract practical techniques from research, build a skill library from papers, convert arXiv papers into reusable agent instructions, or batch-process multiple papers into skills. Also trigger when someone asks about extracting actionable knowledge from papers, making research practical for LLM agents, or systematically converting academic contributions into structured agent capabilities.
Establish regret bounds for behavior cloning with discretized actions combining statistical error and quantization error terms. Prove smoothness requirements for safe quantizer design, show that learning-based quantizers fail these requirements, and propose model-based augmentation to reduce error dependence from H² to H.
Dynamically allocate LoRA ranks per-layer during fine-tuning instead of using fixed uniform ranks. Learn optimal rank for each layer and subject via variational framework with discretized exponential distribution, reducing memory footprint while maintaining fidelity and text-alignment.
基于 SOC 职业分类
正在显示 SKILL.md
| name | sliding-window-attention-adaptation |
| title | Practical Sliding Window Attention Adaptation for Pre-trained Language Models |
| version | 0.0.2 |
| engine | skillxiv-v0.0.2-claude-opus-4.6 |
| license | MIT |
| url | https://arxiv.org/abs/2512.10411 |
| keywords | ["sliding-window-attention","long-context","adaptation","inference-speedup","language-models"] |
| description | Adapt full-attention language models to sliding window attention without expensive retraining. Combine five synergistic strategies (full decode, sink tokens, interleaved layers, chain-of-thought, fine-tuning) achieving 30-100% speedups while maintaining 90-100% accuracy. |
Rather than retraining from scratch, this method systematically combines five complementary adaptation strategies to convert full-attention LLMs to sliding window attention (SWA). The key insight is that "no single ingredient suffices"—specific synergistic combinations of these methods effectively recover original performance. The approach identifies configurations achieving substantial speedups while preserving accuracy, depending on whether efficiency or quality takes priority.
The approach systematically combines five complementary strategies:
1. Full Attention Decode Apply SWA only during prefilling (context encoding), switch to full attention during generation when sequence lengths are short.
2. Keep First k Tokens Preserve attention to initial "sink" tokens while using SWA elsewhere, maintaining access to document-level context.
3. Interleaving FA/SWA Layers Alternate between full-attention and SWA layers throughout the model, balancing efficiency and global context awareness.
4. Chain-of-Thought Enable explicit reasoning during decoding to compensate for limited context windows, allowing the model to compensate for SWA context constraints.
5. Fine-tuning Lightweight SWA-aware supervised fine-tuning on long-context data to adapt the model to the new attention pattern.
Start with your full-attention model. Experiment with combinations of these five strategies:
Recommended: Test recommended configurations on your specific task to find the right accuracy-efficiency trade-off.
