https://warmctl.io/blog/ Benchmarks, deep dives and field notes on warm restores for vLLM. en Tue, 13 Oct 2026 00:00:00 +0000 https://warmctl.io/blog/boosting-dedicated-inference/ https://warmctl.io/blog/boosting-dedicated-inference/ Tue, 13 Oct 2026 00:00:00 +0000 hello@warmctl.io (Kirill Bespalov) warmctl restores GLM-5.3-Flash on 8×H100 in 58 s from a Nebius shared filesystem, against a 650 s vLLM cold start. Kubernetes setup, 15 cold starts, storage and vLLM benchmarks. benchmarkNebiusKubernetes8×H100