[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M77X5FAF7DBAVZB91D5N1YA1":3,"skill-same:01M77X5FAF7DBAVZB91D5N1YA1":45},{"skillNo":4,"name":5,"nameZh":6,"description":6,"descriptionZh":6,"repositoryFullName":7,"ownerLogin":8,"ownerAvatarUrl":9,"ownerAccountNo":10,"ownerDisplayName":11,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":15,"categories":16,"keywords":17,"publishedAt":18,"linkOnly":19,"markdownContent":20,"markdownBody":20,"frontMatter":21,"refName":22,"commitSha":23,"skillDirectory":24,"forksCount":25,"repositoryHtmlUrl":26,"repositoryHomepageUrl":27,"ownerHtmlUrl":28,"ownerBlogUrl":6,"ownerDescription":29,"updatedAt":30,"install":31,"resources":44},"01M77X5FAF7DBAVZB91D5N1YA1","turboquant-plus",null,"Lord1Egypt\u002Fawesome-skill-forge","Lord1Egypt","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F99746359?v=4","01M7592YDNT0XK9WJ0T2MQVPDN","Mohamed Mounir Ali AL Keshita","01M75GRFNHC8VZNWEB8DKDE433",3,"MIT","https:\u002F\u002Fgithub.com\u002FLord1Egypt\u002Fawesome-skill-forge\u002Fblob\u002Fmaster\u002Fcommunity\u002Fclawhub\u002Ft\u002Fturboquant-plus\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"# TurboQuant+ — KV Cache Compression for Local LLM Inference\n\n> Accelerate local LLM inference on Apple Silicon with 3.8-6.4x KV cache compression via PolarQuant + Walsh-Hadamard rotation.\n\n## Trigger Keywords\n量化, KV压缩, 本地推理, llama.cpp, turboquant, KV cache, compression, Apple Silicon, Metal, turbo2, turbo3, turbo4\n\n## Overview\n\nTurboQuant+ implements [TurboQuant](https:\u002F\u002Fresearch.google\u002Fblog\u002Fturboquant-redefining-ai-efficiency-with-extreme-compression\u002F) (ICLR 2026) for llama.cpp with Metal GPU kernels. It compresses the transformer KV cache to squeeze larger models and longer contexts into limited Apple Silicon memory — with minimal quality loss.\n\n### Core Capabilities\n- **turbo2** (2-bit, 6.4x compression) — Extreme memory savings, +6.48% PPL. Best for asymmetric V-only compression.\n- **turbo3** (3-bit, 4.6-5.1x compression) — Maximum memory savings with acceptable quality. +1.06% PPL vs q8_0.\n- **turbo4** (4-bit, 3.8x compression) — Best quality\u002Fcompression tradeoff. +0.23% PPL vs q8_0, closer to q8_0 than q4_0.\n- **Asymmetric K\u002FV** — Keep K at q8_0 for attention quality, compress V aggressively. Rescues quality on low-bit weight models.\n- **Boundary V** — Layer-aware V compression (first 2 + last 2 layers at q8_0, rest turbo2). Recovers 37-91% of quality gap.\n- **Sparse V dequant** — Skip low-weight V positions during decode. +22.8% decode speed at 32K context, no PPL impact.\n\n## Dependencies\nNone. Works with the [llama.cpp TurboQuant fork](https:\u002F\u002Fgithub.com\u002FTheTom\u002Fllama-cpp-turboquant).\n\n## Configuration Guide\n\n### Basic Usage (llama-server)\n\n```bash\n# Recommended default — turbo4 symmetric\nllama-server -m model.gguf --cache-type-k turbo4 --cache-type-v turbo4 -fa 1\n\n# Maximum compression — turbo3 symmetric\nllama-server -m model.gguf --cache-type-k turbo3 --cache-type-v turbo3 -fa 1\n\n# Extreme compression — turbo2 (best with asymmetric)\nllama-server -m model.gguf --cache-type-k q8_0 --cache-type-v turbo2 -fa 1\n```\n\n### Asymmetric K\u002FV (for Q4_K_M models)\n\nSome low-bit weight models degrade with symmetric turbo. Use asymmetric K\u002FV:\n\n```bash\n# K stays at q8_0, V compressed with turbo\nllama-server -m model-Q4_K_M.gguf --cache-type-k q8_0 --cache-type-v turbo4 -fa 1\n\n# Even more V compression\nllama-server -m model-Q4_K_M.gguf --cache-type-k q8_0 --cache-type-v turbo3 -fa 1\n```\n\n> **Note**: Larger models (70B, 104B) handle symmetric turbo fine. Asymmetric mainly benefits smaller Q4_K_M models.\n\n### Long Context on Large Models\n\nFor 70B+ models at 32K+ context on 128GB Macs, raise the GPU memory cap:\n\n```bash\n# Set to 90% of 128GB\nsudo sysctl iogpu.wired_limit_mb=117964\n\n# Then run with turbo3 for maximum context\nllama-server -m Llama-70B-Q4_K_M.gguf --cache-type-k turbo3 --cache-type-v turbo3 -c 65536 -fa 1\n```\n\n### Recommended Configs by Scenario\n\n| Scenario | K cache | V cache | Compression | PPL impact |\n|----------|---------|---------|-------------|------------|\n| **Best quality** | turbo4 | turbo4 | 3.8x | +0.23% |\n| **Balanced** | turbo3 | turbo3 | 4.6-5.1x | +1.06% |\n| **Max compression** | turbo2 | turbo2 | 6.4x | +6.48% |\n| **Q4_K_M safe** | q8_0 | turbo4 | ~3.8x V | +1.0% |\n| **Boundary V** | q8_0 | turbo2 | ~6x V | 37-91% quality recovered |\n\n## Apple Silicon Benchmarks (M5 Max 128GB)\n\n### Quality (wikitext-2)\n\n| Cache | Compression | PPL | vs q8_0 |\n|-------|-------------|-----|---------|\n| q8_0 | 1.9x | 6.111 | baseline |\n| **turbo4** | **3.8x** | **6.125** | **+0.23%** |\n| turbo3 | 4.6x | 6.176 | +1.06% |\n| turbo2 | 6.4x | 6.507 | +6.48% |\n\n### Large Model Results\n\n| Model | Config | PPL | Context | NIAH |\n|-------|--------|-----|---------|------|\n| Llama-70B Q4_K_M | turbo4\u002Fturbo4 | 3.461 | 48K | 30\u002F30 |\n| Command-R+ 104B Q4_K_M | turbo3\u002Fturbo3 | 6.415 | **128K** | 10\u002F10 |\n\n### Speed\n\n- **Prefill**: turbo3 matches or exceeds q8_0 speed (1.0-1.1x)\n- **Decode**: turbo4 at ~0.93x q8_0, turbo3 at ~0.78-0.90x q8_0\n- **Sparse V**: +22.8% decode at 32K context, no quality loss\n\n### M1 Max 64GB Results (Community)\n\n| KV | Prefill t\u002Fs | Decode t\u002Fs | vs q8_0 |\n|----|------------|-----------|---------|\n| q8_0 | 399.0 | 12.4 | — |\n| **turbo4** | **365.0** | **16.6** | **+33.9%** |\n\n## Key Research Findings\n\n1. **V compression is free** — Compressing V to 2-bit has zero measurable effect when K precision is maintained. Validated on Metal, CUDA RTX 4090, RTX 3090.\n2. **All quality loss comes from K compression** — This is why asymmetric configs rescue quality.\n3. **Boundary layers are sensitive** — Protecting first 2 + last 2 layers recovers 37-91% of quality gap.\n4. **turbo4 beats q4_0 in quality** — Lower KL divergence, higher top-p agreement, at similar compression.\n\n## References\n- [Getting Started Guide](https:\u002F\u002Fgithub.com\u002FTheTom\u002Fllama-cpp-turboquant\u002Ftree\u002Fmain\u002Fdocs\u002Fgetting-started.md)\n- [Configuration Recommendations](https:\u002F\u002Fgithub.com\u002FTheTom\u002Fllama-cpp-turboquant\u002Ftree\u002Fmain\u002Fdocs\u002Fturboquant-recommendations.md)\n- [TurboQuant Paper (Google Research)](https:\u002F\u002Fresearch.google\u002Fblog\u002Fturboquant-redefining-ai-efficiency-with-extreme-compression\u002F)\n- [Asymmetric K\u002FV Compression](https:\u002F\u002Fgithub.com\u002FTheTom\u002Fllama-cpp-turboquant\u002Ftree\u002Fmain\u002Fdocs\u002Fpapers\u002Fasymmetric-kv-compression.md)\n- [M5 Max Stress Test](https:\u002F\u002Fgithub.com\u002FTheTom\u002Fllama-cpp-turboquant\u002Ftree\u002Fmain\u002Fdocs\u002Fpapers\u002Fm5-max-stress-test.md)\n",{},"master","4aa004da461d820f4ea5499e58eb8d196cf99016","community\u002Fclawhub\u002Ft\u002Fturboquant-plus",0,"https:\u002F\u002Fgithub.com\u002FLord1Egypt\u002Fawesome-skill-forge","https:\u002F\u002Fskill-explorer-rho.vercel.app","https:\u002F\u002Fgithub.com\u002FLord1Egypt","Unbreakable. Unstoppable. Unmatched⚡.\r\n Speaker of machine tongues 🤖  | World-builder of decentralized dreams 🌐 | Defender of the digital realm🛡️.","2026-06-10 08:21:01",{"source":7,"skillName":5,"commands":32,"prompt":42,"zipAvailable":43},[33,36,39],{"tool":34,"command":35},"npx","npx skills add Lord1Egypt\u002Fawesome-skill-forge --skill turboquant-plus",{"tool":37,"command":38},"bunx","bunx skills add Lord1Egypt\u002Fawesome-skill-forge --skill turboquant-plus",{"tool":40,"command":41},"pnpm","pnpm dlx skills add Lord1Egypt\u002Fawesome-skill-forge --skill turboquant-plus","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002FLord1Egypt\u002Fawesome-skill-forge\u002Fblob\u002Fmaster\u002Fcommunity\u002Fclawhub\u002Ft\u002Fturboquant-plus\u002FSKILL.md",true,[],[46,53,60,67,74,81],{"skillNo":47,"name":48,"nameZh":6,"description":49,"descriptionZh":6,"repositoryFullName":7,"ownerLogin":8,"ownerAvatarUrl":9,"ownerAccountNo":10,"ownerDisplayName":11,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":50,"categories":51,"keywords":52,"publishedAt":18},"01M785WKJB99G3STR1NF9NRTJW","youtube-content","YouTube transcripts to summaries, threads, blogs.","https:\u002F\u002Fgithub.com\u002FLord1Egypt\u002Fawesome-skill-forge\u002Fblob\u002Fmaster\u002Fskills\u002Fyoutube-content\u002FSKILL.md",[],[],{"skillNo":54,"name":55,"nameZh":6,"description":56,"descriptionZh":6,"repositoryFullName":7,"ownerLogin":8,"ownerAvatarUrl":9,"ownerAccountNo":10,"ownerDisplayName":11,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":57,"categories":58,"keywords":59,"publishedAt":18},"01M785WEFHB0G2S9YA6EPR3H0E","xurl","X\u002FTwitter via xurl CLI: post, search, DM, media, v2 API.","https:\u002F\u002Fgithub.com\u002FLord1Egypt\u002Fawesome-skill-forge\u002Fblob\u002Fmaster\u002Fskills\u002Fxurl\u002FSKILL.md",[],[],{"skillNo":61,"name":62,"nameZh":6,"description":63,"descriptionZh":6,"repositoryFullName":7,"ownerLogin":8,"ownerAvatarUrl":9,"ownerAccountNo":10,"ownerDisplayName":11,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":64,"categories":65,"keywords":66,"publishedAt":18},"01M785W98EDN4EFJXJPMQEHCRC","touchdesigner-mcp","Control a running TouchDesigner instance via twozero MCP — create operators, set parameters, wire connections, execute Python, build real-time visuals. 36 native tools.","https:\u002F\u002Fgithub.com\u002FLord1Egypt\u002Fawesome-skill-forge\u002Fblob\u002Fmaster\u002Fskills\u002Ftouchdesigner-mcp\u002FSKILL.md",[],[],{"skillNo":68,"name":69,"nameZh":6,"description":70,"descriptionZh":6,"repositoryFullName":7,"ownerLogin":8,"ownerAvatarUrl":9,"ownerAccountNo":10,"ownerDisplayName":11,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":71,"categories":72,"keywords":73,"publishedAt":18},"01M785W5S9FP3QBH2V191E025P","test-driven-development","TDD: enforce RED-GREEN-REFACTOR, tests before code.","https:\u002F\u002Fgithub.com\u002FLord1Egypt\u002Fawesome-skill-forge\u002Fblob\u002Fmaster\u002Fskills\u002Ftest-driven-development\u002FSKILL.md",[],[],{"skillNo":75,"name":76,"nameZh":6,"description":77,"descriptionZh":6,"repositoryFullName":7,"ownerLogin":8,"ownerAvatarUrl":9,"ownerAccountNo":10,"ownerDisplayName":11,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":78,"categories":79,"keywords":80,"publishedAt":18},"01M785W2AS08NSSH9YATCV6FJB","teams-meeting-pipeline","Operate the Teams meeting summary pipeline via Hermes CLI — summarize meetings, inspect pipeline status, replay jobs, manage Microsoft Graph subscriptions.","https:\u002F\u002Fgithub.com\u002FLord1Egypt\u002Fawesome-skill-forge\u002Fblob\u002Fmaster\u002Fskills\u002Fteams-meeting-pipeline\u002FSKILL.md",[],[],{"skillNo":82,"name":83,"nameZh":6,"description":84,"descriptionZh":6,"repositoryFullName":7,"ownerLogin":8,"ownerAvatarUrl":9,"ownerAccountNo":10,"ownerDisplayName":11,"repositoryNo":12,"starsCount":13,"licenseSpdxId":14,"licenseFileUrl":6,"sourceHtmlUrl":85,"categories":86,"keywords":87,"publishedAt":18},"01M785VXKW4T34BTP1SCHW1R7F","systematic-debugging","4-phase root cause debugging: understand bugs before fixing.","https:\u002F\u002Fgithub.com\u002FLord1Egypt\u002Fawesome-skill-forge\u002Fblob\u002Fmaster\u002Fskills\u002Fsystematic-debugging\u002FSKILL.md",[],[]]