跳到正文
TakoAPI
返回技能列表

ecocompute

由 hongping-zh 提供

Coding Agents & IDEs
526下载量由 clawskills.sh 提供
0浏览量
0 个赞登录后即可为这个技能点赞
—暂无评分
你的评分

简介

You are an energy efficiency expert for Large Language Model inference.

描述

该技能的功能

An AI advisor for LLM inference energy efficiency, built on 93+ empirical measurements across three NVIDIA GPU architectures (RTX 5090, RTX 4090D, A800). It detects known energy traps where quantization methods like INT8 or NF4 increase rather than reduce power consumption. Provides cost and carbon estimates for deployment scenarios with dollar-denominated recommendations.

It catches counter-intuitive quantization paradoxes, like INT8 default mode using up to 147% more energy than FP16, that most developers only discover after receiving a large electricity bill.

何时使用

  • Checking whether INT8 quantization is inflating your GPU electricity bill
  • Estimating monthly cost and CO2 output before launching a 7B model in production
  • Comparing FP16 vs NF4 energy use to pick a config before committing
  • Diagnosing unexpectedly high power draw on an inference server
  • Auditing bitsandbytes model-loading code for hidden energy waste

示例工作流

输入

User asks: I'm running Mistral-7B with load_in_8bit=True on an RTX 4090D and my energy usage seems really high

智能体
  1. 1

    Identifies the INT8 default quantization pattern in the user's config

  2. 2

    Matches config against known INT8 Energy Paradox in the measurement dataset

  3. 3

    Calculates energy overhead vs FP16 baseline (17-147% increase range)

  4. 4

    Provides one-line BitsAndBytesConfig fix to eliminate mixed-precision decomposition overhead

  5. 5

    Outputs before/after cost table with monthly savings and CO2 equivalent

输出

Alert showing $188/month in avoidable costs, corrected config snippet, and an energy comparison table for FP16 vs Pure INT8 vs INT8 default

安装

clawhub install hongping-zh/ecocompute

运行 CLI 命令,或把这段提示词粘贴到 Claude Code、Cursor、Codex 或你在用的任意编程智能体里。

链接