Markdown source

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

Conference Context

Session 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.

Media Evidence

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).

Evidence Graph

This evidence graph is generated from currently linked source material: official schedule text, related video pages, cached transcripts, visible slide text, dense/reconstructed slide pages, and AI slide-classification audits.

Media Signals

Agent Reading Notes

Use these signals to refine the synopsis, topic links, people/company context, and method notes. If a source is a related external video rather than an exact official recording, keep it framed as supporting evidence.

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

Supporting Slides

Slide Evidence

Synthesis

Synthesized Breakdown

Two Bugs That Hid in Plain Sight: A vLLM Debugging Detective Story ## Conference Context - Date/time: 2026-07-01 · 3:20pm-3:40pm - Track/room: Inference · Track 9 - Speaker(s): Asaf Gardin, Yuval Belfer - Session type/status: session · confirmed - Track: Inference - Room: Track 9 - Session type: session - Status: confirmed ## Session Description Your model generates gibberish. Once every thousand prompts. High confidence scores. No crashes.

Speaker And Company Context

Topics Covered

Derived Links And Source Material

Novel Concepts / Clever Methods

Evidence Boundary

This synthesis is based on the official schedule and linked source pages. It should be revisited when exact session recordings or transcript-backed secondary sources are available.