Koskamo

MarvinLabs vram_fit MCP

An MCP server that provides tools to determine if a specific QLoRA fine-tuning configuration for the gemma-3-4b-it model will fit on a given GPU. It utilizes measured VRAM data from NVIDIA and AMD hardware to predict peak memory usage, maximum rank, and fit status (fits, tight, or does not fit) for local LLM training.

Visit MarvinLabs

Endpoint and access

No sign-in to connect and list tools. Individual tools may still need a vendor account, credentials, or permissions; check the documented prerequisites.

MarvinLabs vram_fit MCP logo, by MarvinLabsVendor evidenceStreamable HTTPNo sign-inEndpoint checked 2026-11-01Operated by MarvinLabs
MCP endpoint (Streamable HTTP)
https://marvinlabs.org/mcp
Vendor documentation

What you can do with it

  • Determine if a QLoRA config fits a specific GPU
  • Find the maximum LoRA rank for a given card
  • Retrieve measured VRAM probes for AMD and NVIDIA GPUs

Source and maintenance

Vendor evidenceRegistry 0.1.0
mcp-registry: org.marvinlabs/vram-fit@0.1.0Project website

Protocol and capabilities

Server
vram_fit 0.1.0
MCP protocol
2025-06-18
Tools
3 advertised
Resources
Not advertised
Resource templates
Not advertised
Prompts
Not advertised
Server-advertised capabilities
{
  "tools": {
    "listChanged": false
  }
}

Server-provided instructions

Untrusted server output shown for inspection. It is not Koskamo guidance and is excluded from knowledge search.

Answers whether a QLoRA fine-tune of gemma-3-4b-it fits on a GPU, from measured one-step probes on NVIDIA and AMD cards (https://marvinlabs.org/vram). Every answer says whether it is measured, interpolated, inferred or extrapolated. Other models return covered=false rather than a guess.

Tools

3 advertised

vram_fit
Advertised by the server during the last successful anonymous check.
Advertised read-only

Will this LoRA config train on a card with vram_gb GB? Returns verdict fits / tight / does_not_fit, the predicted peak in MB, the margin, how each part was derived, and any measured probe with the same settings.

vram_fit input schema
{
  "type": "object",
  "required": [
    "vram_gb",
    "rank"
  ],
  "properties": {
    "os": {
      "type": "string",
      "default": "linux",
      "description": "windows or linux; Windows spills instead of OOM"
    },
    "dora": {
      "type": "boolean",
      "default": false
    },
    "rank": {
      "type": "integer",
      "minimum": 1
    },
    "model": {
      "type": "string",
      "default": "unsloth/gemma-3-4b-it"
    },
    "quant": {
      "enum": [
        "4bit",
        "bf16"
      ],
      "type": "string",
      "default": "4bit",
      "description": "Base weights: bitsandbytes nf4 or bf16"
    },
    "seq_len": {
      "type": "integer",
      "default": 7177,
      "description": "Longest training example, in tokens"
    },
    "vram_gb": {
      "type": "number",
      "description": "Card memory as sold, in GB (8, 12, 16, 24...)"
    },
    "lora_dropout": {
      "type": "number",
      "default": 0
    },
    "response_len": {
      "type": "integer",
      "default": 237,
      "description": "Tokens the loss covers (the response span)"
    },
    "other_processes_mb": {
      "type": "number",
      "default": 0,
      "description": "VRAM already held by other programs"
    }
  }
}
vram_fit annotations
{
  "readOnlyHint": true,
  "openWorldHint": false
}
vram_max_rank
Advertised by the server during the last successful anonymous check.
Advertised read-only

Largest LoRA rank (multiple of 8) that fits with margin on a card with vram_gb GB, and the largest that is tight (a real run may fit, the one-step probe would not).

vram_max_rank input schema
{
  "type": "object",
  "required": [
    "vram_gb"
  ],
  "properties": {
    "dora": {
      "type": "boolean",
      "default": false
    },
    "model": {
      "type": "string",
      "default": "unsloth/gemma-3-4b-it"
    },
    "quant": {
      "enum": [
        "4bit",
        "bf16"
      ],
      "type": "string",
      "default": "4bit",
      "description": "Base weights: bitsandbytes nf4 or bf16"
    },
    "seq_len": {
      "type": "integer",
      "default": 7177,
      "description": "Longest training example, in tokens"
    },
    "vram_gb": {
      "type": "number",
      "description": "Card memory as sold, in GB (8, 12, 16, 24...)"
    },
    "lora_dropout": {
      "type": "number",
      "default": 0
    },
    "response_len": {
      "type": "integer",
      "default": 237,
      "description": "Tokens the loss covers (the response span)"
    },
    "other_processes_mb": {
      "type": "number",
      "default": 0,
      "description": "VRAM already held by other programs"
    }
  }
}
vram_max_rank annotations
{
  "readOnlyHint": true,
  "openWorldHint": false
}
vram_probes
Advertised by the server during the last successful anonymous check.
Advertised read-only

The raw measured probe rows (RX 7900 XTX and RTX 3060 Ti), filtered by any of the given settings. gpu matches part of the card name, e.g. '7900' or '3060'.

vram_probes input schema
{
  "type": "object",
  "properties": {
    "gpu": {
      "type": "string"
    },
    "dora": {
      "type": "boolean"
    },
    "rank": {
      "type": "integer"
    },
    "quant": {
      "type": "string"
    },
    "seq_len": {
      "type": "integer"
    },
    "lora_dropout": {
      "type": "number"
    }
  }
}
vram_probes annotations
{
  "readOnlyHint": true,
  "openWorldHint": false
}

Prompts

Not advertised

No prompt definitions available.

Resources

Not advertised

No resource definitions available.

30-day availability

Aggregated from read-only remote checks; this is not a tool-execution or security guarantee.

Successful checks
1 of 1 (100%)
Average latency
415 ms
95th percentile latency
415 ms

Maintenance history

Public check and source-refresh events. Editorial and ownership records remain private.

  • · probe · passed · 415 ms · metadata changed · 3 tool(s)

Connect from your client

Each snippet registers the same endpoint; pick the client you use.

This server is operated by the named third party. Listing and endpoint checks are not an endorsement or a security audit. Review the operator's permissions, terms, and data handling before connecting. How listings are checked and corrected.

Claude Code

Run this once in a terminal. It registers the remote server. Use --scope project to write .mcp.json for the repo.
Claude Code configuration
claude mcp add --transport http marvinlabs-vram-fit-mcp https://marvinlabs.org/mcp

Codex

Run this once. Codex stores MCP servers in ~/.codex/config.toml (or .codex/config.toml in a trusted project).
Codex configuration
codex mcp add marvinlabs-vram-fit-mcp --url https://marvinlabs.org/mcp

Cursor

Merge this server into .cursor/mcp.json in your project, then enable marvinlabs-vram-fit-mcp in your MCP settings.
Cursor configuration
{
  "mcpServers": {
    "marvinlabs-vram-fit-mcp": {
      "url": "https://marvinlabs.org/mcp"
    }
  }
}

Cline

Merge this into cline_mcp_settings.json (Cline → MCP Servers → Configure).
Cline configuration
{
  "mcpServers": {
    "marvinlabs-vram-fit-mcp": {
      "type": "streamableHttp",
      "url": "https://marvinlabs.org/mcp",
      "disabled": false
    }
  }
}

VS Code

Merge this server into .vscode/mcp.json. Use MCP: List Servers to start marvinlabs-vram-fit-mcp and inspect its tools.
VS Code configuration
{
  "servers": {
    "marvinlabs-vram-fit-mcp": {
      "type": "http",
      "url": "https://marvinlabs.org/mcp"
    }
  }
}

Gemini CLI

Run this once. It writes .gemini/settings.json in the current project; pass --scope user for ~/.gemini/settings.json.
Gemini CLI configuration
gemini mcp add --transport http marvinlabs-vram-fit-mcp https://marvinlabs.org/mcp

Claude Desktop

Merge this into claude_desktop_config.json (macOS: ~/Library/Application Support/Claude/, Windows: %APPDATA%\Claude\), then restart Claude Desktop.
Claude Desktop configuration
{
  "mcpServers": {
    "marvinlabs-vram-fit-mcp": {
      "url": "https://marvinlabs.org/mcp"
    }
  }
}

Windsurf

Merge this into ~/.codeium/windsurf/mcp_config.json, then refresh MCP in Cascade. Windsurf remote servers use serverUrl, not url.
Windsurf configuration
{
  "mcpServers": {
    "marvinlabs-vram-fit-mcp": {
      "serverUrl": "https://marvinlabs.org/mcp"
    }
  }
}

Let your agent find its tools

Find servers like MarvinLabs vram_fit MCP from your agent. Add Koskamo once to search the index for endpoints, access requirements, and client setup. Public discovery needs no key.

Connect Koskamo Explore workflows
See all 4,175

Do you run MarvinLabs vram_fit MCP?

Prove ownership with a DNS record, a well-known file, or your Official MCP Registry entry, and your edits publish immediately. Proofs are re-checked weekly; admins can still hide or remove a listing.

Claim this listing