omm

omm tune

Recommend a starting context length, GPU offload, thread count, and batch size for a model on this machine.

01 / 06

Overview

Reach for tune once you've picked a model, installed or not, and want conservative starting values before you configure a runner by hand. It never benchmarks anything itself — it's a prediction based on this machine's hardware and the model's size, meant as a first guess you'd then verify with omm benchmark.

02 / 06

Options

Every flag this command accepts, and what it defaults to when you leave it out.

  • <name>—Default: required

    An installed model's filename, a curated name, or a repo reference. Not-yet-installed models work too, as long as their size can be resolved.

03 / 06

Examples

From a plain search to something you'd put in a script.

Recommended settings for an installed model.

$ omm tune qwen2.5-0.5b-instruct-q4_k_m.gguf

The same profile as JSON.

$ omm tune qwen2.5-0.5b-instruct-q4_k_m.gguf --json

Works on a model that isn't installed yet, if its size can be resolved.

$ omm tune mistral-7b-instruct-q4

04 / 06

A real run

Real omm tune qwen2.5-0.5b-instruct-q4_k_m.gguf capture, 2026-08-24, this dev machine. The negative headroom reflects this machine's real memory pressure at capture time — a busier or freer machine will show a different number.

06 / 06

If something goes wrong

Every message below is one this command actually prints. Find yours, read why it happened, then do the last line.

  1. Unknown model 'zzzz-totally-fake-model-name-xyz'. Use a curated name (tinyllama-1.1b-q4, llama3.1-8b-instruct-q4, mistral-7b-instruct-q4), an 'org/repo:file.gguf' ref (optionally prefixed 'hf:' or 'ms:'), or a direct URL.
    why
    This name doesn't match anything in the curated catalog, and it isn't a repo reference or URL tune recognizes either.
    what to do
    Try omm search first to find the exact name or reference.
    source
    src/omm/hub.py:371

Still stuck? Open an issue with the exact message you saw.

All commands