Stop Shipping LLM Settings That Only Win on Paper - Rohan Pandey & Balaji Varadarajan, DigitalOcean

Most LLM serving "wins" come from turning up a single setting, and on production hardware that setting often has little to do with what your users actually get. Rohan Pandey and Balaji Varadarajan walk through the optimization agent DigitalOcean runs on its AMD MI355 inference servers: it proposes a vLLM and Kubernetes configuration, rolls it out, measures throughput under real latency targets, and either keeps the winner or safely restores the old setup. They share what four studies revealed, including a +4% "winner" that turned out to be run-to-run noise, and how they're extending the same agent to disaggregated prefill and decode serving. Rohan Pandey, DigitalOcean: https://x.com/RohanPandey7821 | https://www.digitalocean.com Balaji Varadarajan, DigitalOcean: https://www.digitalocean.com