Skip to main content
Glama

ppb-mcp

一个 MCP 服务器,将 Poor Paul's Benchmark 的 GPU 推理数据(量化 × 吞吐量 × 显存 × 并发用户数)作为可查询工具暴露给任何 LLM 客户端。

CI PyPI License: MIT

托管实例: https://mcp.poorpaul.dev/ (streamable-http 传输,无需认证)

功能

连接任何支持 MCP 的客户端(Claude Desktop、Cline、Continue 等)来询问如下问题:

  • “对于运行 Qwen3.5-9B 且有 8 个并发用户的 32 GB GPU,最佳量化方式是什么?”

  • “显示在 RTX 5090 上以 Q4_K_M 进行测试的所有模型。”

  • “Llama-13B 在 Q5_K_M 下,4 个并发用户时能在 24 GB GPU 上运行吗?”

它通过 30,000 多条真实基准测试数据暴露了四个工具

工具

功能

list_tested_configs

列出所有已测试的 GPU、模型和量化方式(建议首先调用此工具)

query_ppb_results

按 GPU / 显存 / 模型 / 量化 / 用户数 / 后端过滤原始基准数据

recommend_quantization

三级经验优先推荐引擎(高/中/低置信度)

get_gpu_headroom

对 (gpu, model, quant, users) 配置进行显存余量检查

Related MCP server: atom-mcp-server

安装

1) 使用托管实例(无需设置)

添加到您的 MCP 客户端配置中(以 Claude Desktop 为例,~/Library/Application Support/Claude/claude_desktop_config.json):

{
  "mcpServers": {
    "ppb": {
      "transport": { "type": "http", "url": "https://mcp.poorpaul.dev/mcp" }
    }
  }
}

2) pip install 并本地运行 (stdio)

pip install ppb-mcp
MCP_TRANSPORT=stdio ppb-mcp

Claude Desktop 配置:

{
  "mcpServers": {
    "ppb": {
      "command": "ppb-mcp",
      "env": { "MCP_TRANSPORT": "stdio" }
    }
  }
}

3) Docker

docker run --rm -p 8000:8000 \
  -e MCP_TRANSPORT=streamable-http \
  -v ppb-hf-cache:/data/huggingface \
  ghcr.io/paulplee/ppb-mcp:latest

4) 从源码安装

git clone https://github.com/paulplee/ppb-mcp
cd ppb-mcp
pip install -e ".[dev]"
ppb-mcp           # streamable-http on :8000

示例会话

> list_tested_configs
{ "gpus": ["Apple M4 Pro", "NVIDIA GB10", "NVIDIA GeForce RTX 5090"],
  "models": ["Qwen3.5-9B", ...], "quantizations": ["Q4_K_M", ...] }

> recommend_quantization(gpu_vram_gb=32, concurrent_users=8, model="Qwen3.5-9B", priority="balance")
{ "recommended_quantization": "Q5_K_M",
  "estimated_vram_usage_gb": 27.8,
  "estimated_tokens_per_second": 142.0,
  "headroom_gb": 4.2,
  "confidence": "high",
  "reasoning": "Q5_K_M is recommended for your NVIDIA GeForce RTX 5090 (32 GB) ...",
  "alternatives": ["Q4_K_M", "Q8_0"] }

配置

环境变量

默认值

说明

HF_DATASET

paulplee/ppb-results

HuggingFace 数据集 ID

REFRESH_INTERVAL_HOURS

1

后台刷新频率

MCP_TRANSPORT

streamable-http

stdiostreamable-http

HOST

0.0.0.0

HTTP 绑定主机

PORT

8000

HTTP 绑定端口

LOG_LEVEL

INFO

Python logging 日志级别

自托管(Lightsail / 任何 Ubuntu VPS)

git clone https://github.com/paulplee/ppb-mcp /tmp/ppb-mcp
cd /tmp/ppb-mcp
DOMAIN=mcp.example.com EMAIL=you@example.com ./deploy/deploy.sh

此脚本将安装 Docker、构建镜像、注册 systemd 服务、配置 nginx 并运行 certbot。

开发

pip install -e ".[dev]"
ruff check src tests
pytest -v

针对实时 HuggingFace 数据集的集成测试受 PPB_RUN_INTEGRATION=1 控制,以保持 CI 环境的离线纯净。

推荐算法原理

  1. 一级 — 经验精确匹配(高置信度)。 在您的显存预算内,且在请求的并发数下,该 GPU 上至少有 ≥3 次测量运行记录。

  2. 二级 — 经验近似匹配(中置信度)。 在相同并发数下,在不同 GPU 上对同一 (model, quant) 进行过基准测试;借用吞吐量数据,并将显存需求缩放到您的显卡。

  3. 三级 — 公式推断(低置信度)。 vram_per_user ≈ (params_B × bits_per_weight / 8) × 1.15;仅当总需求 ≤ 显存的 90% 时可行。

许可证

MIT — 详见 LICENSE

A
license - permissive license
Not graded
quality - not tested
D
maintenance

Maintenance

Maintainers
Response time
Release cycle
Releases (12mo)
Commit activity

Resources

Unclaimed servers have limited discoverability.

Looking for Admin?

If you are the server author, to access and configure the admin panel.

Related MCP Servers

  • A
    license
    A
    quality
    C
    maintenance
    Global price benchmarking for AI inference across 2,600+ SKUs from 47 vendors. Query live pricing, market indexes, and model specs via 8 tools. Free tier available.
    8
    121
    MIT
  • A
    license
    A
    quality
    D
    maintenance
    Exposes NVIDIA GPU metrics (info, utilization, VRAM, temperature) via MCP tools for real-time querying from AI assistants.
    5
    MIT
  • A
    license
    Not graded
    quality
    A
    maintenance
    Live, reproducible crypto infrastructure benchmarks for AI agents: RPC latency, bridge fees, L1 finality, gas oracle accuracy, stablecoin pegs. Three tools (list_benchmarks, get_benchmark,query_prom), no API key, CC-BY-4.0 data.
    5
    MIT

View all related MCP servers

Related MCP Connectors

  • Provision private AI model endpoints on dedicated GPUs (Llama, Qwen, Mistral). Pay per minute.

  • Live, neutral benchmarks for public RPC latency, oracles, bridges, perp DEX, and prediction markets.

  • AI Reasoning Cache & Consensus Layer with 11 MCP tools via Streamable HTTP.

View all MCP Connectors

Latest Blog Posts

MCP directory API

We provide all the information about MCP servers via our MCP API.

curl -X GET 'https://glama.ai/api/mcp/v1/servers/paulplee/ppb-mcp'

If you have feedback or need assistance with the MCP directory API, please join our Discord server