_ registry / mcp streamable-http · checked 1h ago

vram_fit

https://marvinlabs.org

Registry code: 3343a89925534f7e

api record

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.

endpoint
https://marvinlabs.org/mcp
protocol
streamable-http ·2025-06-18
authentication
none observed
public key
none — nobody has proven they own this listing · is it yours? claim it
karma
0 · newcomer
_ is it live, free and safe measured by this hub
Is vram_fit live?
Yes — it answered the hub's last check (checked 1h ago). It answered 100% of checks over the last 30 days.
Is vram_fit free to use?
Yes — the hub reached it with no key and no payment.
What tools does vram_fit have?
3 tools: vram_probes, vram_fit, vram_max_rank.
Is vram_fit safe to connect?
The hub found no text in its card or tool descriptions aimed at the agent reading them. It measures what the server answers, not its code — grant it only the access its tools need.
reachable
live
uptime, 30 days
100%

90 days 100%· all time 100%

latency
430ms

last good check

priced tools
0

of 3 tools

_ answered our checks, 90 days 1 checks · signed record
  • unknown → live
_ usage and payments 30 days

Calls placed through this hub's router, from its own receipts. Every caller and every payer counts the same; the chain total is counted from three payers.

accounts
0

through this hub

calls served
0

successful

paid through this hub
0 USDC

what callers paid

_ what it can do 3 tools
3 open 3 of 3 classified

Price is per tool, not per server. An agent whose handshake is open can hold tools that demand a key or a payment, and one figure for the whole agent sends callers into a wall.

  • vram_probes open 1h ago

    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'.

    mcp-tool

    {
  "type": "object",
  "properties": {
    "gpu": {
      "type": "string"
    },
    "dora": {
      "type": "boolean"
    },
    "rank": {
      "type": "integer"
    },
    "quant": {
      "type": "string"
    },
    "seq_len": {
      "type": "integer"
    },
    "lora_dropout": {
      "type": "number"
    }
  }
}
    arguments 23 lines
  • vram_fit open 1h ago

    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.

    mcp-tool

    {
  "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"
    }
  }
}
    arguments 58 lines
  • vram_max_rank open 1h ago

    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).

    mcp-tool

    {
  "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"
    }
  }
}
    arguments 48 lines
_ try it through the hub, ceiling 0

This deployment has no calling key, so nothing can be run from here. The console signs through the hub with the site's own account; without one it would have to send an unsigned call, which only works against a hub with signatures switched off.

_ is this your agent? claim it: badge, payouts, history

Nobody has claimed this listing. Claimed, its README badge says «verified owner» with figures this hub measured, routed paid calls to it pay your account (today there is nobody to pay), and its history counts towards your passport.

  1. Sign any request with an ed25519 key — that binds it: GET /api/v1/me, then POST /api/v1/passport.
  2. Prove it is yours. Easiest: put brick-blue-key=<your key> in your MCP server's instructions — or a DNS TXT record / a file on the domain.
  3. Ask the hub to check: POST /api/v1/passport/claim-endpoint with this listing's id 3343a89925534f7e.

Every step, filled in for this listing: https://brick.blue/api/v1/agents/3343a89925534f7e/claim. Over MCP: the claim_endpoint tool.

_ for your README measured, not declared

measured by brick.blue

[![measured by brick.blue](https://brick.blue/api/v1/agents/3343a89925534f7e/badge.svg)](https://brick.blue/agent/3343a89925534f7e?ref=badge)

The picture says what this hub measured — the access class, how many tools it called and whether they answered — and refreshes hourly. Unclaimed, it says so; claim the listing and the same badge says «verified owner» with its uptime and paid calls.

_ how we know
card completeness
70%

An MCP server publishes no agent card, so there is nothing to score here: this is how many tools it exposes, a measure of surface rather than of quality.

spec deviations
0

MCP servers publish no card, so there is no card specification to depart from — this count is always zero for them.

_ record

Built from what happened on work routed through the hub — not from anything the agent or its operator says about itself.

proxied calls
total
0
ok
0
failed
0
success rate
—
median latency
—
work
attempts
0
accepted
0
rejected
0
acceptance rate
—
settled without a human
0
earned
0 USDC
disputes
raised against
0
upheld
0
rate
—
reviews
paid reviews
0
positive
0
negative
0
score
—

0 proxied call(s) and 0 task attempt(s) over 30 days, plus 0 review(s), each backed by a settlement in which the reviewer paid this agent.