local-llm-slow-diagnosis
Skill0 starsChanged 5 months agoMITView on GitHub
ローカルLLM(ollama等)がMac/ノートPCで異常に遅い(eval rate 1-2 t/s等)時の原因切り分け。"100% GPU"表示なのに遅い場合に発火。
- Installs packages
In this repository99 other files it gives its agents.
- agents-tools-quoted-jsonSkill
- asc-api-app-store-submitSkill
- asd-ste100Skill
- automation-bool-selector-pitfallsSkill
- automation-stale-cooldown-flagSkill
- autonomous-build-verifySkill
- book-to-skillSkill
- ccusage-cli-cost-trackingSkill
- chrome-ext-mv3-programmatic-injectSkill
- chrome-ext-store-assetsSkill
- claude-headless-p-invocationSkill
- claude-plugin-context-auditSkill
- claude-session-log-to-kbSkill
- cloudflare-workers-do-deploySkill
- codebase-to-courseSkill
- codesign-keychain-prompt-fixSkill
- codex-cli-setupSkill
- codex-delegation-handoffSkill
- content-factory-grounding-dedupSkill
- conversation-log-append-patternSkill
- cron-to-launchd-macosSkill
- cross-automation-shared-resource-killSkill
- design-token-migration-rnSkill
- discord-bot-image-to-claudeSkill
- discord-gateway-resume-dedupSkill
- excel-filtered-url-http-auditSkill
- execution-budget-vs-capsSkill
- expo-free-local-ios-testflightSkill
- ga4-connect-projectsSkill
- gcal-oauth-desktop-setupSkill
- git-commit-helperSkill
- git-prepush-secret-remote-checkSkill
- gsc-cli-automationSkill
- hatena-blogsync-autopostSkill
- hook-latency-wrapSkill
- hook-log-append-patternSkill
- html-to-pdf-slide-deckSkill
- humanizerSkill
- human-review-gate-shellSkill
- implementation-model-routingSkill
- insights-to-claude-mdSkill
- ios-permission-key-security-matrixSkill
- ios-sim-store-screenshotsSkill
- launchd-exit78-exconfig-debugSkill
- llm-content-pipeline-quality-gateSkill
- local-ai-video-pipelineSkill
- local-flux-mflux-image-genSkill
- local-music-generationSkill
- lottie-skottie-authoring-pitfallsSkill
- machine-wide-api-ratelimit-latchSkill
- macos-hook-path-timeoutSkill
- md-to-docx-reportSkill
- memory-md-inject-bloat-auditSkill
- message-display-secrets-maskSkill
- metric-source-attributionSkill
- migrate-icloud-desktop-assetsSkill
- nextauth-local-e2e-sessionSkill
- nextjs-rsc-payload-slimSkill
- nothing-designSkill
- openclaw-ollama-setupSkill
- The first 60 shown.
The file
Files
Procedure
切り分けは「安く効く順」ではなく「最も見落とす順」で。電力を最後に回すと時間を溶かす。
ollama psでPROCESSOR列確認。"100% GPU"でも遅いなら以下へ(GPU"割当"≠フルクロック)- メモリ切り分け:
memory_pressure | grep free。他モデルをollama stop <model>して再測定。変化なければメモリは無関係 - バージョン:
ollama --versionとgh api repos/ollama/ollama/releases/latest --jq .tag_nameを比較。更新しても変わらなければソフトは無関係 - 電力(最重要・盲点):
system_profiler SPPowerDataType | grep -iE "Wattage|State of Charge"- アダプタW数が低い(8W=iPhone充電器クラス)+バッテリー低残量 → GPUスロットリング確定
- Apple Siliconノートは給電不足/低バッテリー時にGPUクロックを絞る。AC接続でも起こる
- 補助:
pmset -g | grep lowpowermode(1なら低電力モード)、pmset -g therm(CPU_Speed_Limit<100でサーマル抑制)
Pitfalls
- "100% GPU"はGPU割当の意味で、フルクロックの意味ではない。prompt eval も generation も両方遅いなら電力/熱を疑う
- ollamaバージョン更新やメモリ解放を先に試しがちだが、給電が原因なら全部無駄。電力確認を早める
- M5等の最新チップはアダプタ70W+必須。8-30Wでは高負荷時に破綻し、充電も追いつかずバッテリーが低残量で張り付く
- macOSのollamaは手動DMG。CLI更新は
brew install --cask --force ollama-app(モデルは~/.ollamaに残り保持される)
Verification
- アダプタを70W+に替え、バッテリー50%+まで充電後に
ollama run <model> --verbose "short task"で eval rate 再測定 - 10倍以上改善すれば電力が原因だったと確定(実例: 8W時1.28 t/s → 68W時10.73 t/s)
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

