← 返回 Skills 市场

Hardware LLM Optimizer v2 (llmfit)

Name: Hardware LLM Optimizer v2 (llmfit)
Author: smseow001

作者 SMS · GitHub ↗ · v2.0.0 · MIT-0

cross-platform ⚠ suspicious

总下载

当前安装

版本数

在 OpenClaw 中安装

/install hardware-llm-optimizer-v2

功能描述

AI硬件LLM推荐工具 - 基于llmfit内核。自动检测CPU/GPU/RAM/VRAM → 智能推荐最适合的大模型 + 量化方案 + 速度估算。支持100+模型库，内置TUI界面和硬件模拟。

使用说明 (SKILL.md)

Hardware LLM Optimizer v2.0

基于 llmfit 的智能硬件LLM推荐工具

安装状态

llmfit 已安装在: /usr/local/bin/llmfit

快速使用

当用户问"能跑什么大模型"、"推荐LLM"、"硬件检测"时使用：

1. 查看推荐模型

llmfit recommend

2. 查看所有推荐（JSON格式，便于解析）

llmfit recommend --json

3. 按用途筛选

llmfit recommend --use-case coding
llmfit recommend --use-case chat
llmfit recommend --use-case general
llmfit recommend --use-case embedding

4. 硬件模拟（模拟不同配置）

# 模拟 16GB 显存
llmfit recommend --memory 16G

# 模拟 32GB 显存 + 64GB RAM
llmfit recommend --memory 32G --ram 64G

5. 交互式TUI（需要终端）

llmfit

输出字段说明

字段	含义
`name`	模型名称
`parameter_count`	参数量
`best_quant`	推荐量化方案
`score`	综合评分（越高越好）
`estimated_tps`	预估速度（tok/s）
`memory_required_gb`	所需显存
`run_mode`	运行模式（GPU/CPU/MoE）
`fit_level`	匹配度（Perfect/Good/Marginal）

量化方案参考

量化	质量	速度	适用场景
FP16	最高	最慢	大显存GPU
Q8_0	很高	较快	中等显存
Q6_K	高	快	6-8GB显存
Q4_K_M	中高	最快	4-6GB显存
Q2_K	中	最快	\x3C4GB显存

本地运行命令

安装Ollama模型

ollama run \x3Cmodel-name>

使用llama.cpp

# 下载GGUF后
./llama.cpp -m \x3Cmodel.gguf> --rompt \x3Cprompt>

最低配置参考（来自llmfit）

显存	推荐模型	量化
2GB	Phi-3-mini, Gemma-2B	Q4
4GB	Llama3-8B, Qwen-7B	Q4
6GB	Llama2-13B, Mistral-7B	Q4/Q6
8GB	Llama2-13B, Yi-9B	Q5/Q6
12GB	Llama2-34B	Q4
16GB	Llama2-34B, Qwen-72B	Q4
24GB+	70B大模型	Q4/Q8

安装llmfit（如需）

curl -fsSL https://llmfit.axjns.dev/install.sh | sh

优势对比

功能	v1.0	v2.0 (llmfit)
模型库	手动查表	100+自动匹配
量化推荐	简单估算	智能最优
速度估算	❌	✅
下载源	❌	✅ GGUF
硬件模拟	❌	✅
TUI界面	❌	✅
多GPU	❌	✅
MoE支持	❌	✅

Powered by llmfit | Updated: 2026-04-17

安全使用建议

This skill's detection code (detect.py) appears benign and aligned with the description: it inspects local hardware via psutil and nvidia-smi and prints recommendations. However, the SKILL.md asks you to install 'llmfit' by piping a script from https://llmfit.axjns.dev/install.sh directly into sh — that pattern is high-risk because it runs arbitrary code from an unvetted host. Before installing or running anything: (1) do NOT run curl ... | sh without inspecting the script; fetch the URL and review its contents first; (2) prefer installing llmfit from a known official source (GitHub releases, vendor homepage) or verify the domain and script integrity; (3) run detect.py locally in a restricted environment if you only want hardware info (it has no network calls); (4) ensure you understand any model downloads (GGUF/ollama) and avoid running unfamiliar binaries as root. If the maintainer can provide a verified upstream URL (official project repo/releases) or an explicit install spec using a reputable package host, that would reduce risk and could change this assessment.

功能分析

Type: OpenClaw Skill Name: hardware-llm-optimizer-v2 Version: 2.0.0 The skill bundle provides hardware detection and LLM recommendations. While the Python script (detect.py) is benign and uses standard libraries like psutil to gather system specs, the SKILL.md file contains a high-risk 'curl | sh' command to install a dependency from an external domain (llmfit.axjns.dev). This pattern is a known vector for supply chain attacks and remote code execution; although it is presented as a legitimate installation step, it constitutes a meaningful high-risk behavior that warrants a suspicious classification.

能力评估

ℹ Purpose & Capability

The skill claims to detect hardware and recommend LLMs/quantization, and the included detect.py implements that behavior (psutil, nvidia-smi checks, /proc/version). However, the SKILL.md centers runtime usage on an external tool 'llmfit' (commands like 'llmfit recommend') while the package provides no install mechanism for llmfit and even asserts llmfit is already at /usr/local/bin; relying on an external binary that isn't provided weakens coherence.

✓ Instruction Scope

Runtime instructions and detect.py stay within the stated purpose: they inspect local system state (CPU, RAM, nvidia-smi, /proc/version), produce recommendations, and reference running local model runtimes (ollama, llama.cpp). The SKILL.md does suggest running network-facing installs and model downloads, but it does not instruct reading unrelated secrets or sending detected data to external endpoints.

⚠ Install Mechanism

There is no formal install spec, but SKILL.md recommends installing llmfit with: curl -fsSL https://llmfit.axjns.dev/install.sh | sh. That is a direct download-and-execute from an unrecognized domain (axjns.dev) — high-risk practice. The skill itself does not include code to fetch that URL, but recommending it without provenance is disproportionate and potentially dangerous.

✓ Credentials

The skill requests no environment variables, no credentials, and detect.py only queries local system info. There is no inappropriate credential access requested.

✓ Persistence & Privilege

The skill does not request always:true and is not marked to be force-included. It does not attempt to modify other skills or system-wide configs. Normal autonomous invocation is allowed (platform default).

如何使用

确保已安装 OpenClaw（本地或 Docker 部署）
在对话框中输入安装命令：/install hardware-llm-optimizer-v2
安装完成后，直接呼叫该 Skill 的名称或使用 /hardware-llm-optimizer-v2 触发
根据 Skill 的参数说明提供必要输入，即可获得结构化输出

版本历史

v2.0.0

Hardware LLM Optimizer v2.0 brings a major upgrade with intelligent, automated model recommendations based on your hardware. - Now uses the llmfit core: auto-detects CPU/GPU/RAM/VRAM to recommend optimal large models and quantization. - Expanded to support 100+ model libraries, with built-in TUI interface and hardware simulation features. - Provides speed estimation, usage-based model filtering, and more detailed output (including quantization plans and fit scores). - Supports advanced scenarios: multi-GPU, MoE, and local GGUF model sources. - Improved user experience with clearer guides, command examples, and comparison to previous version.

元数据

Slug hardware-llm-optimizer-v2

版本 2.0.0

许可证 MIT-0

累计安装 0

当前安装数 0

历史版本数 1

常见问题