llm-d Put a Fair Queue in Front of the GPUs

llm-d shipped 1.0 on 20 October, and the release notes lead with who gets served rather than how fast the tokens come out.

• Version v1.0.0 landed under Apache 2.0, still a CNCF Sandbox project, built by more than 500 contributors across 25 organizations.
• Flow control adds a probabilistic admitter, fairness scoring, turn priority and controlled load shedding, so a bursty agent loop can’t starve interactive requests.
• Snap reports prefix cache hit rates of up to 75 to 80 percent after putting the router behind an Envoy based service mesh.
• The router configuration API graduates to v1, so configs written against 0.x need rewriting rather than a version bump.
• The project is deprecating its own inference images. The ROCm image went in this release, replaced by a pinned vLLM build.
• DisaggregatedSet treats prefill and decode replicas as one unit, with coordinated rollouts and readiness gated on a pod snapshot.

Admission policy just became an object you own, and a gateway rate limit no longer describes what the cluster accepts. Full analysis here.