Built a VRAM scheduler (vramble) to automate model swapping on my local AI workstation
To fix this, I built a request dispatching layer called vramble.
Proof1 community thread
Key points
- I set up a dedicated workstation for local AI experimentation with various models and tools.
- It automatically loads and unloads models and manages the queue using a priority system.
- Now I can run different workflows without manually managing models (or wasting Claude tokens to orchestrate the process).
- Side note: Qwen 27B at 200+ t/s is performing way better than I expected. opeMIT license, available on GitHub: github.com/ValerioDolci/vramble
Sources (1)
- [1]Built a VRAM scheduler (vramble) to automate model swapping on my local AI workstationr/LocalLLaMA (top, daily) · Oct 13, 05:28 AM
To fix this, I built a request dispatching layer called vramble.
I set up a dedicated workstation for local AI experimentation with various models and tools.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 13, 2026How far are we from a local model being able to optimize it's own runtime stack?
- Oct 13, 2026Hybrid Reasoning Agents
- Oct 12, 2026Qwen 3.8 Flash Next Q3 S - Benchmark creation
- Oct 12, 2026LSearch - Spotlight for windows with embedding
- Oct 12, 2026Inside the AI Economy: What Stripe's Data Reveals — Nilofer Rajpurkar, Stripe
- Oct 12, 2026AReA: Adaptive Receptive-Field Attention for Efficient High-Resolution Diffusion Transformers
