kvcached brings virtual memory to LLM KV cache, deployed on 10K+ GPUs

techNmak · x · 2026-09-25

The idea: kvcached applies one of the oldest OS concepts—virtual memory—to the biggest bottleneck in LLM serving: KV cache GPU memory allocation.

The problem: vLLM and SGLang can reserve large KV-cache pools per model, and PagedAttention manages blocks efficiently within a model, but physical GPU memory is still hard to redistribute across model instances.

The approach: kvcached decouples the virtual KV address space from underlying physical GPU memory. A model reserves a large virtual space while physical pages are mapped only on demand, then reclaimed and reused by other models.

Traction: the underlying Prism system was published at OSDI '26, and the kvcached balloon driver has been deployed across 10K+ GPUs.

Original post →

More from Infra

Infra channel →