Why it matters
vLLM is the default open-source inference engine for self-hosted LLMs, and Crusoe supplies the GPU capacity many teams route to instead of paying frontier API markups — this is where the real cost-of-serving tradeoffs get argued out.
The tokenmaxxing angle
The agenda covers KV cache offloading and cluster-wide KV caching — the exact techniques that cut redundant token recomputation and lower per-request inference cost — plus GPU orchestration for cheaper self-hosted routing.
From the organizers
Hosted by Delanie Thone, Judy Wu, Simon Mo, and Zachary Xi in China Basin, SF; agenda: 5pm doors, 5:30-7pm talks & panel Q&A, 7-9pm networking; topics include vLLM roadmap, KV cache offloading, and a Rust BPE tokenizer.