- Who it helps
- Teams operating vLLM on Linux GPU servers
- Availability
- Public Apache-2.0 beta; requires Docker, NVIDIA runtime, root and matching GPU model/count and driver branch.
What is saved
The snapshot includes loaded weights, compiled kernels and captured CUDA execution graphs. The maker reports 37 seconds versus a 650-second cold start for GLM-5.3-Flash on eight H100s with the snapshot in page cache; cold network storage took 56.8 seconds.
Where it applies
Current support covers one to eight GPUs on one host. Different hardware needs a matching snapshot; multi-host engines and SGLang are not supported yet. Host checks, fingerprints and a comparison prompt run before serving.
What you can do
- Compare your exact GPU model, count and driver branch with the snapshot requirements. Open reference ↗
Where the evidence stops
Documentation and repository count checked November 11, 2026; no installation or restore executed. Reported timings depend on model, storage, cache state and host configuration. GPU and snapshot storage costs remain.
Sources read. Public source text checked November 11, 2026 UTC. No third-party project or complete workflow executed.
Sources & further reading
GitHub · implementation
warmctl repository and limits ↗
Beta checkpoint and restore documentation read. Timing depends on cache state and matching hardware.
TensorRails · page checked 2026-11-11
33 stars · observed 2026-11-11 00:24 UTC
Coverage note: No restore, storage-cost comparison or independent timing run performed.
The record over time (1)
Dated source checks and published explanations. An observation date can differ from the original release date.
- · Added to Current
warmctl saves an already initialized vLLM engine so a matching GPU host can resume it without repeating every startup stage.