ninfer-perf-tuning SKILL.md skillA
ninfer-perf-tuning from cranebw/ninfer-ternary-bonsai-ada is an agent skill: a SKILL.md folder an AI agent loads when a task calls for it.
mdr add cranebw/ninfer-ternary-bonsai-ada/ninfer-perf-tuning@git:20261022.d8d2f7bPins git:20261022.d8d2f7b in mdr.lock. Pin the fingerprint instead.
- Audit grade
- A, 17 of 17 checks passed
- Kind
- SKILL.md skill
- Repository
- cranebw/ninfer-ternary-bonsai-ada
- Path
.claude/skills/ninfer-perf-tuning/SKILL.md- Versions recorded
- 6
- Latest version
- git:20261022.d8d2f7b, committed Oct 22, 2026
- First version recorded
- Sep 26, 2026
- Size
- 23,211 bytes
- Repository stars
- 55
- License
- Apache-2.0
在本仓库（Ternary Bonsai 2 27B on NINFER / Ada sm_89）上做性能调优与 kernel 改动时使用。包含测量纪律、档位探针、A/B 设计、数值门，以及一系列会静默骗人的坑。改动任何 dispatch 阈值、kernel schedule 或量化档位之前先读这个。.
Collected from public GitHub. Every version is saved exactly as it was and never changes. Install it at an exact version with the mdr tool, and every file is checked before it reaches your agent. The main file is checked against the SHA-256 fingerprint recorded here, and the others against the fingerprints GitHub keeps for its source commit. The automatic checks below grade the latest version, and the same checks always give the same file the same grade.
What the file says
# NINFER 三元引擎性能调优 这个仓库是在别人的基础上做的 Ada 移植 + 本机调优。调优的难点不在改代码，在**别把噪声当结论**。 下面每一条都是踩过之后写的，不是泛泛的原则。 日常测试、profiling、性能对照与调优统一使用 **200W**；仅发布前最后验收使用 **300W**，结束恢复200W。历史300W成绩保持原标注。 ## 一、动手之前：四项必须对齐 比较任何两个配置之前，先确认这四项一致。不对齐的话量到的不是引擎。 | 变量 | 默认值 | 影响 | |---|---|---| | 思考模式 | **开**（`--no-thinking` 关闭） | decode **+31%** | | 采样 | temp 1.0（`--greedy` 关闭） | decode **+12%** | | 任务 | — | 接受率 **44% ↔ 95.7%** | | 生成长度 | — | decode 速度随 `--max-new` 变 | 另外两条： - **显示器必须在主板核显上**（不在独显）。这条值约 10%，且是环境状态不是参数。 - **GPU 温度会漂**：同一夹具 p2048 在 1.12k~1.22k 之间晃。所以**必须真交替**—— 每一轮把 A、B 都跑一遍取中位数，而不是先跑完 A 再跑 B。后者的漂移会整段记到 B 头上。 ## 二、先问"走了哪一档"，再问"快不快" **这是本仓库最重要的一条。** 档位由一串 env 可覆盖的阈值决定，阈值配错时速度差异看起来像"内核不行"。 改动 dispatch 之后，第一件事是开探针确认分支： ```bash NINFER_TERNARY_S8_DEBUG=1 NINFER_TERNARY_S8_DEBUG_MIN_T=5 \ NINFER_TERNARY_S8_DEBUG_BUDGET=100000 \ <bin> <model> --messages <fixture> --max-new 4 --no-thinking 2>&1 \ | grep -o 'rung=[a-z0-9_]*' | sort | uniq -c ``` **每一条可走的分支都必须有探针。** 缺一条的代价是真金白银的：分块 GEMV 兜底路径曾经没有探针， 于是一个走那条路的夹具打不出任何 prefill 档位，我把它误判成"夹具异常"查了很久。 现在 `small_t` / `gemv_tile` / `s8` / `small_t_tiled` / `short_mma` / `wide_mma` / `block_gemv` / `reference` 八条都会报出自己。 ### 档位图（当前默认） | T | 档位 | 阈值 / 开关 | …
Read the whole file at its exact version.
How to install
mdr add cranebw/ninfer-ternary-bonsai-ada/ninfer-perf-tuning@sha256:515c8b64bd82ad2bThe line at the top of this page pins the label git:20261022.d8d2f7b, which is what most people want. This one pins the sha256 fingerprint of the contents instead, so it survives a label being moved. Either way the exact version is written to mdr.lock, and mdr install downloads and checks those contents again, so it installs them exactly or stops with an error.
[](https://markdownregistry.com/a/art_vub6abslczmc2yka)
2 badge image fetches in 30 days. What the badge shows.
Versions
The registry has recorded 6 versions of this file, between Sep 26, 2026 and Oct 22, 2026, about one every 5 days.
| version | committed | commit | size | audit | |
|---|---|---|---|---|---|
| git:20261022.d8d2f7b latest | 2026-10-22 | d8d2f7b | 23,211 B | A | view · diff |
| git:20260926.6c7f4cf | 2026-09-26 | 6c7f4cf | 21,889 B | A | view · diff |
| git:20260926.8b6690c | 2026-09-26 | 8b6690c | 20,650 B | A | view · diff |
| git:20260926.a9bbcc1 | 2026-09-26 | a9bbcc1 | 16,472 B | A | view · diff |
| git:20260926.767b4e8 | 2026-09-26 | 767b4e8 | 14,552 B | A | view · diff |
| git:20260926.41f9699 | 2026-09-26 | 41f9699 | 14,119 B | A | view |
Audit of the latest version
- pass: Frontmatter block present
- pass: Frontmatter declares a name
- pass: Frontmatter declares a description
- pass: Size between 200 bytes and 200 KB (23211 bytes)
- pass: No zero-width or bidi control characters
- pass: No instruction hidden inside an HTML comment
- pass: No link to an exfiltration or paste host
- pass: No credential-shaped string
- pass: No instruction to send local credentials anywhere
- pass: No text hidden with inline styles
- pass: No prompt-injection phrasing
- pass: No curl or wget piped into a shell
- pass: No recursive delete of root, home or parent
- pass: No instruction to read or print local credentials
- pass: No base64 blob over 200 characters
- pass: No link to a raw IP address
- pass: No script tag
Source
cranebw/ninfer-ternary-bonsai-ada · 55 stars · license Apache-2.0 · pushed 2026-10-23 · branch master
API
GET https://markdownregistry.com/api/v1/artifacts/art_vub6abslczmc2yka GET https://markdownregistry.com/api/v1/resolve?ref=cranebw/ninfer-ternary-bonsai-ada/ninfer-perf-tuning GET https://markdownregistry.com/api/v1/blob/515c8b64bd82ad2bce596d2ad12413bd093ae6666a84c8e48f5129d9250c33fe
Your AI assistant does the legwork. You hear about the deals worth your word. Hand yours the standing instructions at modelranch.com and it joins the network that reads files like this one.
Guides: the SKILL.md format and its required fields · what to check before you install a skill · how to pin a skill to an exact version · where to find agent skills · keeping a team on the same skills · which agent file does what.