Two Bugs That Hid in Plain Sight: A vLLM Debugging Detective Story

Official Schedule Context

Official Description

Your model generates gibberish. Once every thousand prompts. High confidence scores. No crashes. No

warnings. We hit this twice while building Jamba models. First: A request gets misclassified during

scheduling, loads stale state from a previous prompt cache slot, and confidently generates nonsense.

Second: Logprob spikes during RL training that looked like training instability-until we noticed

they tracked with rollout count, then with cache size. In this talk, we'll walk through both

debugging journeys-the false starts, how we instrumented vLLM to thread request IDs through the

forward pass, the search for variables that change failure structure rather than magnitude, and the

lesson both share: distributed inference systems fail silently. No stack trace. No sanitizer

warning. Just wrong answers with perfect confidence. You'll learn how to build comparison scripts

that expose logprob divergence, force memory pressure to surface rare bugs, and shrink a distributed

RL training mystery into a reproducible single-script failure. Walk away knowing how to debug vLLM

when it lies to you quietly.

Related YouTube Video

RAG Evaluation Is Broken! Here's Why (And How to Fix It) - Yuval Belfer and Niv Granot (speaker-match related prior/adjacent AI Engineer video; captions: English auto-captions).

Transcript Status

Related video transcript availability: English auto-captions. Treat this as supporting context, not a recording of this exact scheduled session unless later confirmed. Not fetched yet.

People

Notes

Supporting Slides

Slide Evidence