kbesplv
Re:Show HN: Warmctl – Fast cold starts for vLLM inference
on Tuesday November 10, 2026 @05:36PM (#50447287)
Hi HN!
I’m building this GPU checkpoint & restore tool to optimize vLLM cold start time for GPU neoclouds
It’s dead simple at UX:
- Run warmctl build once in your CI/CD
- warmctl push the warm image to any OCI registry
- warmctl up on every new replica on GPU hosts
That’s it, so you can serve even Deepseek 4.1 or GLM 5.3 < 60 seconds
Re:Ask HN: How do you review code nowadays?
on Thursday October 15, 2026 @05:36PM (#50110672)
We don’t review harder, we make changes smaller. Small blast radius, instant rollback, tested on real envs
Re:Ask HN: What is the black swan event that could bring down GPU costs?
on Thursday October 15, 2026 @05:33PM (#50110631)
Training is 10-100x more compute than inference. The current rush is all training. When it tips over to mostly inference, a lot of that hardware is just going to sit there.
Re:Iosevka is an open-source, sans-serif and slab-serif, monospace font
on Wednesday October 14, 2026 @12:00PM (#50089765)
looks amazing