npm package discovery and stats viewer.

Discover Tips

  • General search

    [free text search, go nuts!]

  • Package details

    pkg:[package-name]

  • User packages

    @[username]

Sponsor

Optimize Toolset

I’ve always been into building performant and accessible sites, but lately I’ve been taking it extremely seriously. So much so that I’ve been building a tool to help me optimize and monitor the sites that I build to make sure that I’m making an attempt to offer the best experience to those who visit them. If you’re into performant, accessible and SEO friendly sites, you might like it too! You can check it out at Optimize Toolset.

About

Hi, 👋, I’m Ryan Hefner  and I built this site for me, and you! The goal of this site was to provide an easy way for me to check the stats on my npm packages, both for prioritizing issues and updates, and to give me a little kick in the pants to keep up on stuff.

As I was building it, I realized that I was actually using the tool to build the tool, and figured I might as well put this out there and hopefully others will find it to be a fast and useful way to search and browse npm packages as I have.

If you’re interested in other things I’m working on, follow me on Twitter or check out the open source projects I’ve been publishing on GitHub.

I am also working on a Twitter bot for this site to tweet the most popular, newest, random packages from npm. Please follow that account now and it will start sending out packages soon–ish.

Open Software & Tools

This site wouldn’t be possible without the immense generosity and tireless efforts from the people who make contributions to the world and share their work via open source initiatives. Thank you 🙏

© 2026 – Pkg Stats / Ryan Hefner

dsh-model-deploy

v1.0.1

Published

DeepSeek Harness plugin: LLM model selection & deployment analysis (deployability, VRAM, TTFT, latency, throughput, power for 38 models × 20 GPUs/NPUs)

Downloads

187

Readme

dsh-model-deploy

🌐 语言切换 / Language: 简体中文 · English

DeepSeek Harness(DSH)插件:LLM 模型选型部署分析

给它模型、GPU/NPU 型号、节点数、精度、上下文长度、并发数与互联带宽,它用一阶解析模型给出可审计的估算:

  • ✅/⚠️/❌ 能否部署(显存/精度/互联检查)+ 修复建议
  • 显存:单卡与集群拆解(权重 / KV 缓存 / 激活 / 运行时)
  • 时延:首 token(TTFT)与整体时延(空闲 / 满载两种口径)
  • 吞吐:单请求与系统 tok/s、req/s、预填充 tok/s
  • 功耗:空闲 / 典型 / 峰值(含 PUE)、每 token 能耗
  • 自动 TP × PP × DP 并行策略与选择理由

按 DeepSeek Harness 架构文档(docs/architecture.md)实现:模型可见能力注册在 ctx.tools,以 bundle 行的形式挂载(cordis.patch.yml)。

安装

dsh plugin --profile web add dsh-model-deploy

或从 GitHub 检出安装:

dsh plugin --profile web add github:YOUR_ACCOUNT/dsh-model-deploy

--profile 必填。安装后重启会话(或 profile),工具 schema 才会进入 prompt 组装。

工具

model_deploy_analyze

主分析工具。示例调用:

{
  "model": "deepseek-v3",
  "gpu": "ascend910b",
  "gpusPerNode": 8,
  "nodes": 1,
  "precision": "int4",
  "ctx": 131072,
  "batch": 4
}

字段:model(目录 id,或自定义模型 JSON 字符串)、gpu(目录 id)、gpusPerNode (1–8)、nodes (1–64)、precisionfp32|bf16|fp16|fp8|int8|int4|fp4)、ctx/batch/ prompt/output(tokens)、intraModeauto|custom)+ intraGBsinterModenone|ib400|ib800|roce100|roce200|custom)+ interGBs

返回结构化报告(statusissuesstrategymemoryperformancepowerassumptions),渲染输出为简洁中文文本报告。

model_deploy_catalog

列出支持的模型(id、参数量、层数、上下文上限)与 GPU/NPU(id、显存、带宽、 FP16/FP8/INT8/FP4、互联、TDP),支持 query/vendor 过滤。先用它查 id,再调分析工具。

覆盖范围

  • 38 个模型 — Qwen3 全系(含 Qwen3-Next/Coder/VL)、Llama 3.1/3.3/4、 DeepSeek-V3/V3.1/V3.2/V4-Flash/V4-Pro/R1、Kimi-K2/-K2-Thinking/-K3、 GLM-4.5-Air/4.6/5、Hunyuan-A13B、Baichuan-M2、Seed-OSS、GPT-OSS、Mixtral、 MiniCPM4、QwQ、Qwen2.5 …
  • 20 款 GPU/NPU — NVIDIA(H100/H200/H20/A100/B200/L40S/RTX 3090/4090/5090)、 AMD MI300X、华为昇腾(910B/910C/950PR)、平头哥含光800、海光 Z100、昆仑芯 P800、 寒武纪 MLU590、天数天垓150、沐曦 C500、摩尔线程 S5000

模型架构数据来自各模型公开的 config.json(ModelScope/HuggingFace);硬件数据来自 官方规格表与公开规格汇总,每个条目在 dsh/data.js 中带有来源链接。官方未公开的 字段(如含光800 显存带宽、950 灵衢互联)标注 ~估算,报告里会提示。

估算模型(可信度所在)

公式与系数全部公开,结论可复核:

  • 权重 = 参数量 × 精度字节;KV = 2 × 层数 × KV头 × head_dim × 字节(GQA), MLA 为 (kv_lora_rank + rope) × 层数 × 字节,滑动窗口层按窗口折算(Qwen3-Next)
  • 解码每 token ≈ max((权重+KV)/带宽, 计算) + TP allreduce(2·(TP−1)/TP × hidden × batch × 2B,走 NVLink/HCCS/PCIe/节点上行)
  • TTFT ≈ 2 × 激活参数量 × 输入 tokens / (算力 × TP × PP × MFU 0.5) + 通信
  • 功耗 = TDP × 负载比例 + 22% 平台开销,PUE 1.25
  • MoE 专家路由通信 ×1.35 修正;PP 每级气泡 10%

引擎由 22 项回归测试锁定,包含真实部署锚点(70B BF16 在 8×H100 ≈ 150 tok/s; DeepSeek-V3 FP8 放不进 8×80G、INT4 可以;Qwen3-8B INT4 单 4090 ≈ 170 tok/s), 以及全部 760 个模型×GPU 组合的无 NaN 扫描。

国产 NPU 按规格表理想值估算,工具会提示 CANN/XPU/MUSA 软件栈实际利用率有差异, 正式采购前请用目标框架实测。

开发

npm test        # node --test,零依赖

dsh/ 为无运行时依赖的纯 ESM。估算引擎与配套网站(见同仓库 README 英文版 “companion site”说明,可选)共用同一实现。

发布(npm / GitHub / 插件目录)

  1. npmnpm publish(需 npm adduser 登录)
  2. GitHub:建仓 dsh-model-deploygit push,目录站会同步元数据
  3. DSH 插件目录:到 https://dshmarketplace.dev/submit 提交收录(人工审核), 或在 DshMarketPlace/dshmarketplace 提 issue
  4. (可选)发到 LINUX DO 社区可获得 verified 徽章

License

MIT