vLLM Lab

Hands-on, source-level studies of vLLM internals — reading the latest main and turning findings into upstream contributions.

Source @ latest main KV cache · offloading Contribution-oriented

vLLM Lab collects deep, contribution-oriented studies of vLLM subsystems, focused on the moving frontier.

Labs

KV OFFLOADING

The v0.19→v0.23 KV-offloading rewrite

How the native OffloadingConnector became a hybrid-aware, multi-tier subsystem, with a timeline, open issues, and two contribution scaffolds.

vllm/v1/kv_offload · main @183a430 · 2026-06-21  
PLANNED

GPUDirect-Storage tier benchmark

Reproducing the cuFile NVMe secondary-tier scaffold on H100/H200, measured against the filesystem tier.

In progress