概览

AI 生成的概览

估算某个开放权重的大语言模型能否装进指定显卡的显存，并给出速度上限和可行的替代方案。

功能: 提供只读工具，用于计算在特定显卡上运行开放权重大语言模型所需的内存。工具包括 can_i_run（是否装得下、内存分配、速度上限、可容纳的最长上下文）、what_fits（单张显卡能装下的模型）、estimate_vram（各量化下的权重、KV 缓存与开销）、estimate_from_hf_repo（读取任意 Hugging Face 仓库的 config.json），以及 list_models、list_gpus。回答以字段形式给出数值，并附模型与显卡页面链接和各项假设。
适用场景: 适合在规划本地大模型推理时使用：想知道某个模型能否装进自己的显卡、能支持多长上下文，或该改选哪个模型或显卡。也适合询问量化方式、显存预算和单流速度的大致预期。
运行要求: 远程方式：连接 streamable HTTP 端点，无需账号或密钥。本地方式：需要 Node.js 20 或更高版本，通过 npx 运行 npm 包。未声明任何环境变量或请求头。远程端点和 Hugging Face 仓库查询都需要网络访问。

安装前请注意

只读且免费，不索取任何凭据。远程服务器无状态、不保留请求日志，但托管方会保留短期的边缘日志，且按地址限流。查询 Hugging Face 仓库时会把仓库名发送给服务器，由其获取该仓库公开的 config.json。速度数值是单流理论上限，并非实测结果。

安装

在 SourceWeft 中

打开控制台中的 Nodegrove VRAM: can I run it?，将其添加到工作区。
为需要使用其工具的对话启用该服务。

Web executable，通过 Streamable HTTP。远程服务在工作区中配置后即可从网页运行时运行。

其他 MCP 客户端

把它添加到你客户端的 mcpServers 配置中。

{
  "mcpServers": {
    "vram-mcp": {
      "type": "http",
      "url": "https://mcp.nodegrove.io/mcp"
    }
  }
}

README

[nodegrove]

Can I run it?

An MCP server that tells your AI assistant whether an open-weight LLM fits your GPU, and what would fit instead.
The arithmetic behind nodegrove.io, for any model on Hugging Face and any card.

[npm] [CI]

Can my RTX 4090 run Llama 3.3 70B?

No: Llama 3.3 70B at Q4_K_M with 8,192 tokens of context needs 45.8 GB, and the RTX 4090 holds 22.8 GB after headroom. Short by 23 GB. What would work instead:

No context length helps: the weights alone are 43.1 GB before a single token of conversation.

RTX 6000 Ada (48 GB) is within a whisker: 45.8 GB against 45.6 GB after headroom. With the KV cache at Q8 it needs 44.4 GB and fits.

The smallest card here that runs it exactly as asked: A100, 80 GB usable.

The biggest model your card does run at these settings, counting mixture-of-experts models at their dense equivalent: Qwen3 32B, 22.4 GB at ~37 tokens/s.

That is the server's own answer, word for word. Every figure is a stated formula over the model's config.json and the card's published specs: no benchmarks, no guesses. It is free, read-only, and needs no account or key.

Connect

Remote: nothing to install

[Install in Cursor] [Install in VS Code] [Add to LM Studio]

Claude (claude.ai and the desktop app)

Customize → Connectors → Add → Add custom connector. Name it Nodegrove VRAM, paste the URL, choose No sign-in, then turn it on in a chat from + → Connectors.

Claude Code

Add --scope user to have it in every project.

Cursor

~/.cursor/mcp.json:

VS Code

.vscode/mcp.json:

Devin Desktop (formerly Windsurf)

LM Studio

Program → Install → Edit mcp.json:

Open WebUI

Admin Settings → Integrations → External Tool Servers → Add Connection. Type MCP (Streamable HTTP), the URL above, authentication None.

Cline

The type must be stated, or Cline treats a URL as the older SSE transport:

json

{ "mcpServers": { "nodegrove-vram": { "type": "streamableHttp", "url": "https://mcp.nodegrove.io/mcp" } } }

{ "mcpServers": { "nodegrove-vram": { "type": "streamableHttp", "url": "https://mcp.nodegrove.io/mcp" } } }

Codex CLI

Gemini CLI

Zed

settings.json:

ChatGPT

Settings → Security and login → turn on Developer mode. At chatgpt.com/plugins, add one with the URL and No Authentication, then pick it in a chat from + → Developer mode.

Local: over stdio

Needs Node.js 20 or newer:

json

{  "mcpServers": {    "nodegrove-vram": {      "command": "npx",      "args": ["-y", "@nodegrove/vram-mcp"]    }  }}

{  "mcpServers": {    "nodegrove-vram": {      "command": "npx",      "args": ["-y", "@nodegrove/vram-mcp"]    }  }}

Tools

Tool	Ask it	It answers with
`can_i_run`	Can my RTX 4090 run Llama 3.3 70B with 32k of context?	Fits, tight or no; the memory split; a speed ceiling; the longest context that fits; and on a no, every change that would make it fit
`what_fits`	What is the best model for my 16 GB card?	Every model checked on one card, with a recommended everyday pick, the largest that fits, the best at Q8 and the first out of reach
`estimate_vram`	How much VRAM does Qwen3 32B need at 64k?	Weights, KV cache and overhead at each quantisation, and the smallest common card that holds each
`estimate_from_hf_repo`	How much memory does `Qwen/Qwen3-Next-80B-A3B-Instruct` need?	Any Hugging Face repo, read from its `config.json`: the attention layout, the cost of each 1,000 tokens of context, and memory at every quantisation
`list_models`, `list_gpus`	Which GPUs do you know?	The models and cards with their specs, ids and pages

Models can be named the way people type them ("llama 3.3 70b", Llama-3.3-70B-Instruct) or given as any Hugging Face repo id. Cards can be named ("4090", "M4 Max") or described by their memory and bandwidth. Every answer carries the numbers as fields, links to the model and card pages on nodegrove.io, and the assumptions behind each figure. All six tools are read-only.

How the numbers are made

Memory is weights + KV cache + overhead. Weights are parameters × bytes per parameter at the quantisation: FP16 2.00, Q8_0 1.06, Q6_K 0.82, Q5_K_M 0.71, Q4_K_M 0.58, Q3_K_M 0.47. Overhead is 0.5 GB plus 4% of the weights.
The KV cache is counted the way each model caches. Sliding-window layers stop at their window, hybrid models (linear attention or Mamba) grow a cache only on their few full-attention layers, and latent attention stores one compressed vector per layer. A standard-transformer formula would overstate these models several times over at long context.
Fit means at most 95% of the memory a runtime can address; above 85% it is tight. Apple silicon gives the GPU about 75% of unified memory.
Speed is 0.7 × memory bandwidth ÷ bytes of active weights read per token. It is a single-stream ceiling, not a measurement, and the faster the figure, the further real runtimes fall below it.
The model table was read from each model's config.json and checked against Hugging Face. Any other repo is read live by the same rules, which reproduce every row of the table; anything the reader cannot model is named in the answer, never guessed.

The full method, with every constant, is at nodegrove.io/data. The same figures are an open dataset under CC BY 4.0, DOI 10.5281/zenodo.22966137.

The math as a library

The server is a thin layer over @nodegrove/llm-math, the package nodegrove.io's pages and calculators are built on:

ts

import { modelById, estimate } from '@nodegrove/llm-math';
estimate({ ...modelById('llama-3.3-70b')!, quant: 'q4', context: 8192 }).totalGb; // 45.77

import { modelById, estimate } from '@nodegrove/llm-math';
estimate({ ...modelById('llama-3.3-70b')!, quant: 'q4', context: 8192 }).totalGb; // 45.77

Privacy

The remote server runs on Cloudflare Workers. Nodegrove keeps no request logs and no record of what you ask: each request is answered by a fresh, stateless instance and forgotten. Cloudflare keeps standard edge logs for a short period, as for any website. Requests are rate-limited to 120 a minute per address. When you ask about a Hugging Face repo, the server fetches that repo's public config.json and metadata; only the repo name is sent. The local version contacts nothing but huggingface.co, and only when you ask about a repo there. Details: nodegrove.io/privacy.

Run your own

The remote endpoint is mcp/src/worker.ts. To deploy a copy to your Cloudflare account, change the route in mcp/wrangler.jsonc, then:

Development

llm-math/ is the package of formulas and data; mcp/ is the server, with the stdio entry (src/stdio.ts), the Worker (src/worker.ts) and the tools (src/tools.ts). This repository is published from Nodegrove's main repository, where the site uses the same package. Issues and pull requests are welcome here; accepted changes are applied upstream and credited.

Found a figure that disagrees with its source? Open an issue with the model or card and the link. To report a security problem, see SECURITY.md.

Licence

The code is MIT. The model and GPU data, llm-math/src/models.ts and llm-math/src/gpus.ts, is CC BY 4.0: use it for anything, and credit Nodegrove (nodegrove.io).

Model and GPU names are trademarks of their owners.

来源：README.md，提交 98c4f60

工具

0

工具元数据尚未被收录。

版本历史

1

v1.0.0最新Oct 3, 2026

Nodegrove VRAM: can I run it?

概览

安装