independent inference research
Coconut Labs
Schedulers, systems notes, and reproducible measurements for shared inference.
KVWarden Gate 2: 1.14× of solo TTFT under load. 26× better than FIFO.
Coconut Labs works on the shared layer of inference: scheduling, fairness, cache pressure, and the measurements that keep claims honest.
The lab is small by design. Fewer abstractions between the benchmark, the note, and the code.
The quiet tenant should still have a name.
Projects
Live
KVWarden
1.14x of solo, 26x better than FIFO
Tenant fairness on shared inference. A quiet tenant stays visible when a flooder arrives.
Read moreIn research
Weft
Apple Silicon scheduler experiments
A local inference thread focused on load, correctness, and tenants that do not politely wait their turn.
Read morepeople
Two people, close to the work.
Coconut Labs is Shrey Patel and Jay Patel, building public research around inference scheduling, fairness, and systems measurement.
Building something at this layer? Write us.
latest note · 2026-04-19 (1 week ago)14 commits this weekkvwarden gate 2 · 1.14× solo · 26× better than fifo3 repos tracked12 rfc openupdated 1h ago