Nodegrove VRAM: can I run it?
Can this LLM run on my GPU? VRAM, speed ceiling and what fits instead, for any model and GPU.
我该使用它吗
质量与安全性
A
基于对工具定义和协议合规性的自动分析。
上下文开销
~2,922token 数（工具定义）
~4.2 KB典型响应大小
对注意力有显著影响（占 128k 上下文窗口的 2.28%）
这是每次将服务器的工具加载到模型上下文窗口时所消耗的大致 token 数。数值越高，可用于其他任务的注意力就越少。
安装
一键安装
将以下内容添加到你的 `claude_desktop_config.json` 文件中：
{
"mcpServers": {
"vram-mcp": {
"command": "npx",
"args": [
"@nodegrove/vram-mcp"
]
}
}
}可运行的软件包
npm@nodegrove/vram-mcp
1.1.1stdio远程端点
https://mcp.nodegrove.io/mcpstreamable-http它能做什么
工具清单
工具（6）
🟢 只读🟡 写入🔴 删除⚪ 未知
🟢can_i_run(model, architecture, active_params_b, gpu, vram_gb, ...)
Can this GPU run this open-weight LLM? Returns fits, tight or no, the memory split (weights, KV cache, overhead), a decode-speed ceiling, the longest context that fits and, on a no, every change that would make it fit: quantisation, KV cache, context, another card or a smaller model. Model: a name or id from list_models, any Hugging Face repo id, or its architecture. GPU: a name or id from list_gpus, or vram_gb for any other card.
输入模式
{
"type": "object",
"properties": {
"model": {
"description": "A model from list_models (id or name, e.g. \"llama-3.3-70b\" or \"Llama 3.3 70B\"), or any Hugging Face repo id (e.g. \"Qwen/Qwen3-8B\"), read live from its config.json.",
"type": "string",
"minLength": 1,
"maxLength": 200
},
"architecture": {
"description": "A model described by its config.json values instead of a name.",
"type": "object",
"properties": {
"params_b": {
"type": "number",
"exclusiveMinimum": 0,
"maximum": 10000,
"description": "Total parameters, billions; all experts for a mixture-of-experts model."
},
"layers": {
"type": "integer",
"exclusiveMinimum": 0,
"maximum": 1000,
"description": "num_hidden_layers"
},
"kv_heads": {
"type": "integer",
"exclusiveMinimum": 0,
"maximum": 1024,
"description": "num_key_value_heads"
},
"head_dim": {
"type": "integer",
"exclusiveMinimum": 0,
"maximum": 4096,
"description": "head_dim, or hidden_size ÷ num_attention_heads"
},
"active_params_b": {
"description": "Parameters read per token, billions (mixture-of-experts only).",
"type": "number",
"exclusiveMinimum": 0,
"maximum": 10000
},
"native_context": {
"description": "The context window the model supports, tokens.",
"type": "integer",
"exclusiveMinimum": 0,
"maximum": 9007199254740991
},
"kv_groups": {
"description": "Only for non-standard attention: one entry per group of layers that cache the same way. Replaces layers × kv_heads × head_dim.",
"maxItems": 8,
"type": "array",
"items": {
"type": "object",
"properties": {
"layers": {
"type": "integer",
"exclusiveMinimum": 0,
"maximum": 9007199254740991
},
"values_per_token": {
"type": "number",
"exclusiveMinimum": 0,
"description": "Values each layer caches per token: 2 × KV heads × head dim, or the latent width for MLA."
},
"window_tokens": {
"description": "Sliding window: these layers keep only this many tokens.",
"type": "integer",
"exclusiveMinimum": 0,
"maximum": 9007199254740991
}
},
"required": [
"layers",
"values_per_token"
]
}
},
"fixed_state_gb": {
"description": "Fixed recurrent state of linear-attention or Mamba layers, GB.",
"type": "number",
"minimum": 0,
"maximum": 100
}
},
"required": [
"params_b",
"layers",
"kv_heads",
"head_dim"
]
},
"active_params_b": {
"description": "Parameters read per token, billions, for a mixture-of-experts model read from Hugging Face (from its model card). Sets the speed ceiling.",
"type": "number",
"exclusiveMinimum": 0,
"maximum": 10000
},
"gpu": {
"description": "A GPU from list_gpus (id or name, e.g. \"rtx-4090\", \"4090\" or \"M4 Max\").",
"type": "string",
"minLength": 1,
"maxLength": 100
},
"vram_gb": {
"description": "Memory of a card not in list_gpus, GB. For a Mac, its unified memory with apple_silicon: true.",
"type": "number",
"exclusiveMinimum": 0,
"maximum": 4096
},
"bandwidth_gb_s": {
"description": "Memory bandwidth from the maker's spec, GB/s, for a speed ceiling.",
"type": "number",
"exclusiveMinimum": 0,
"maximum": 100000
},
"apple_silicon": {
"description": "vram_gb is Apple unified memory; the GPU can use about 75% of it by default.",
"type": "boolean"
},
"quant": {
"default": "q4",
"description": "Weight quantisation: fp16 (FP16 / BF16), q8 (Q8_0), q6 (Q6_K), q5 (Q5_K_M), q4 (Q4_K_M), q3 (Q3_K_M). q4 is the common default.",
"type": "string",
"enum": [
"fp16",
"q8",
"q6",
"q5",
"q4",
"q3"
]
},
"context": {
"default": 8192,
"description": "Tokens held in context: prompt plus conversation.",
"type": "integer",
"minimum": 1,
"maximum": 10000000
},
"kv_cache": {
"default": "fp16",
"description": "KV cache precision. fp16 is what most runtimes use; q8 halves the cache.",
"type": "string",
"enum": [
"fp16",
"q8"
]
}
},
"$schema": "https://json-schema.org/draft/2020-12/schema"
}🟢what_fits(gpu, vram_gb, bandwidth_gb_s, apple_silicon, quant, ...)
Which open-weight LLMs fit this GPU: every model in list_models checked at one quantisation and context, with a recommended everyday model (the biggest class that fits with room for context at conversational speed), the largest that fits, the best at Q8 and the first out of reach. GPU: a name or id from list_gpus, or vram_gb for any other card.
输入模式
{
"type": "object",
"properties": {
"gpu": {
"description": "A GPU from list_gpus (id or name, e.g. \"rtx-4090\", \"4090\" or \"M4 Max\").",
"type": "string",
"minLength": 1,
"maxLength": 100
},
"vram_gb": {
"description": "Memory of a card not in list_gpus, GB. For a Mac, its unified memory with apple_silicon: true.",
"type": "number",
"exclusiveMinimum": 0,
"maximum": 4096
},
"bandwidth_gb_s": {
"description": "Memory bandwidth from the maker's spec, GB/s, for a speed ceiling.",
"type": "number",
"exclusiveMinimum": 0,
"maximum": 100000
},
"apple_silicon": {
"description": "vram_gb is Apple unified memory; the GPU can use about 75% of it by default.",
"type": "boolean"
},
"quant": {
"default": "q4",
"description": "Weight quantisation: fp16 (FP16 / BF16), q8 (Q8_0), q6 (Q6_K), q5 (Q5_K_M), q4 (Q4_K_M), q3 (Q3_K_M). q4 is the common default.",
"type": "string",
"enum": [
"fp16",
"q8",
"q6",
"q5",
"q4",
"q3"
]
},
"context": {
"default": 8192,
"description": "Tokens held in context: prompt plus conversation.",
"type": "integer",
"minimum": 1,
"maximum": 10000000
}
},
"$schema": "https://json-schema.org/draft/2020-12/schema"
}🟢estimate_vram(model, architecture, active_params_b, quant, context, ...)
How much memory an LLM needs: weights + KV cache + overhead at each quantisation (or one), at a given context, and the smallest common card class that holds each. Model: a name or id from list_models, any Hugging Face repo id, or its architecture (params_b, layers, kv_heads, head_dim).
输入模式
{
"type": "object",
"properties": {
"model": {
"description": "A model from list_models (id or name, e.g. \"llama-3.3-70b\" or \"Llama 3.3 70B\"), or any Hugging Face repo id (e.g. \"Qwen/Qwen3-8B\"), read live from its config.json.",
"type": "string",
"minLength": 1,
"maxLength": 200
},
"architecture": {
"description": "A model described by its config.json values instead of a name.",
"type": "object",
"properties": {
"params_b": {
"type": "number",
"exclusiveMinimum": 0,
"maximum": 10000,
"description": "Total parameters, billions; all experts for a mixture-of-experts model."
},
"layers": {
"type": "integer",
"exclusiveMinimum": 0,
"maximum": 1000,
"description": "num_hidden_layers"
},
"kv_heads": {
"type": "integer",
"exclusiveMinimum": 0,
"maximum": 1024,
"description": "num_key_value_heads"
},
"head_dim": {
"type": "integer",
"exclusiveMinimum": 0,
"maximum": 4096,
"description": "head_dim, or hidden_size ÷ num_attention_heads"
},
"active_params_b": {
"description": "Parameters read per token, billions (mixture-of-experts only).",
"type": "number",
"exclusiveMinimum": 0,
"maximum": 10000
},
"native_context": {
"description": "The context window the model supports, tokens.",
"type": "integer",
"exclusiveMinimum": 0,
"maximum": 9007199254740991
},
"kv_groups": {
"description": "Only for non-standard attention: one entry per group of layers that cache the same way. Replaces layers × kv_heads × head_dim.",
"maxItems": 8,
"type": "array",
"items": {
"type": "object",
"properties": {
"layers": {
"type": "integer",
"exclusiveMinimum": 0,
"maximum": 9007199254740991
},
"values_per_token": {
"type": "number",
"exclusiveMinimum": 0,
"description": "Values each layer caches per token: 2 × KV heads × head dim, or the latent width for MLA."
},
"window_tokens": {
"description": "Sliding window: these layers keep only this many tokens.",
"type": "integer",
"exclusiveMinimum": 0,
"maximum": 9007199254740991
}
},
"required": [
"layers",
"values_per_token"
]
}
},
"fixed_state_gb": {
"description": "Fixed recurrent state of linear-attention or Mamba layers, GB.",
"type": "number",
"minimum": 0,
"maximum": 100
}
},
"required": [
"params_b",
"layers",
"kv_heads",
"head_dim"
]
},
"active_params_b": {
"description": "Parameters read per token, billions, for a mixture-of-experts model read from Hugging Face (from its model card). Sets the speed ceiling.",
"type": "number",
"exclusiveMinimum": 0,
"maximum": 10000
},
"quant": {
"type": "string",
"enum": [
"fp16",
"q8",
"q6",
"q5",
"q4",
"q3"
],
"description": "Weight quantisation: fp16 (FP16 / BF16), q8 (Q8_0), q6 (Q6_K), q5 (Q5_K_M), q4 (Q4_K_M), q3 (Q3_K_M). Omit it for all 6."
},
"context": {
"default": 8192,
"description": "Tokens held in context: prompt plus conversation.",
"type": "integer",
"minimum": 1,
"maximum": 10000000
},
"kv_cache": {
"default": "fp16",
"description": "KV cache precision. fp16 is what most runtimes use; q8 halves the cache.",
"type": "string",
"enum": [
"fp16",
"q8"
]
}
},
"$schema": "https://json-schema.org/draft/2020-12/schema"
}🟢estimate_from_hf_repo(repo, active_params_b, context, kv_cache)
Reads any Hugging Face model repo's config.json and parameter count and estimates its memory: the attention layout found (standard, sliding-window, hybrid or latent), how much each 1,000 tokens of context costs, and weights + KV cache + overhead at every quantisation. For models nodegrove.io has not reviewed; anything the reader cannot model is listed in warnings.
输入模式
{
"type": "object",
"properties": {
"repo": {
"type": "string",
"minLength": 3,
"maxLength": 200,
"description": "Hugging Face repo id, e.g. \"Qwen/Qwen3-8B\", or its huggingface.co URL."
},
"active_params_b": {
"description": "Parameters read per token, billions, for a mixture-of-experts model read from Hugging Face (from its model card). Sets the speed ceiling.",
"type": "number",
"exclusiveMinimum": 0,
"maximum": 10000
},
"context": {
"default": 8192,
"description": "Tokens held in context: prompt plus conversation.",
"type": "integer",
"minimum": 1,
"maximum": 10000000
},
"kv_cache": {
"default": "fp16",
"description": "KV cache precision. fp16 is what most runtimes use; q8 halves the cache.",
"type": "string",
"enum": [
"fp16",
"q8"
]
}
},
"required": [
"repo"
],
"$schema": "https://json-schema.org/draft/2020-12/schema"
}🟢list_models(search)
The open-weight LLMs nodegrove.io has verified against their config.json (data version 2026-10-06): id, size, attention design, native context, licence, memory at Q4 with 8k context and each model's page.
输入模式
{
"type": "object",
"properties": {
"search": {
"description": "Words to filter by, e.g. \"qwen\" or \"24 GB\".",
"type": "string",
"maxLength": 100
}
},
"$schema": "https://json-schema.org/draft/2020-12/schema"
}🟢list_gpus(search)
The GPUs and machines nodegrove.io covers: memory, the memory a runtime can use and bandwidth, from the makers' specs, with each one's page.
输入模式
{
"type": "object",
"properties": {
"search": {
"description": "Words to filter by, e.g. \"qwen\" or \"24 GB\".",
"type": "string",
"maxLength": 100
}
},
"$schema": "https://json-schema.org/draft/2020-12/schema"
}社区
证据
最近观测
已验证未记录版本6 个工具