[mikasenghaas]

Summary

• Bump vllm-router from 0.2.1 to 0.2.2 (amd64 and arm64 release wheels).
• The new router picks up fix: refresh worker models every 20 seconds by default router#58. The default health-check interval drops from 60s to 20s, and every health-check cycle refreshes /v1/models. Workers that are still loading at startup no longer stay under unknown and miss model-filtered traffic for minutes.

Verification

• uv lock resolves with vllm-router v0.2.1 -> v0.2.2.
• Installed router reports version 0.2.2 and RouterArgs.health_check_interval_secs == 20.
• Ran 5 steps of configs/basic/reverse-text/rl.toml single-node on SLURM (H200) through the router: 5/5 steps, 0% rollout errors, no tracebacks.

🤖 Generated with Claude Code

---

Note

Medium Risk
Touches the disaggregated inference routing dependency only, but the new router changes health-check timing and worker/model discovery behavior on the critical vLLM serving path.

Overview
Upgrades the pinned vllm-router wheels from 0.2.1 to 0.2.2 for Linux x86_64 and aarch64 in pyproject.toml, with matching uv.lock updates (version, download URLs, and wheel hashes).

Installs that use the disagg extra (and related GPU/disagg resolution paths) will pull the new router release instead of 0.2.1. No in-repo application code changes—only dependency resolution.

The 0.2.2 router (PrimeIntellect-ai/router#58) shortens the default health-check interval (60s → 20s), refreshes /v1/models on each health cycle, and avoids workers staying unknown while still loading at startup (so model-filtered traffic is not skipped for long periods).

^{Reviewed by Cursor Bugbot for commit f8087ad. Bugbot is set up for automated code reviews on this repo. Configure here.}