Start
warmctl docs
warmctl snapshots a warmed-up vLLM engine and restores it in seconds on any host with the same GPUs and NVIDIA driver branch. One static binary that builds, pushes, pulls and serves a warm engine the way you run a container.
$ curl -fsSL https://warmctl.io/install.sh | sh $ warmctl doctor # check this host $ warmctl catalog # models this host can restore $ warmctl up qwen3-8b --listen 0.0.0.0:8000 # pull, restore, serve
The install script ships with the public release. Until then, early-access users install from the release archive: see Install.
How these docs are organized
Four kinds of pages, each for a different moment. Pick the one that matches what you need now.
Tutorials
Learn by doing, step by step, with a result at the end.
How-to guides
Get one task done on a system you already run.
Five words
| snapshot | A warmed engine frozen to disk: the process tree (CRIU), the GPU state of every worker (cuda-checkpoint) and the processed weights. One directory with a manifest.json. |
| profile | A few lines of YAML that say which model, on which GPUs, with which vllm serve flags and warmup. build turns a profile into a snapshot. |
| platform | What a snapshot needs from a host: GPU model and count, NVIDIA driver branch, topology and CPU features. Its tag reads like 8xh100-80gb-sxm-580.173.02. |
| registry | Any OCI registry. A model is one repository with one snapshot per platform, listed in latest like a multi-arch image. |
| instance | A restored engine serving on this host, behind warmctl's OpenAI-compatible gateway. ps lists them, down stops one. |
Where to start
New here: install warmctl, then serve your first model from the public catalog. Running your own models: build and publish a snapshot. Questions and early access: hello@warmctl.io.