Rewinding One Request in a Continuously Batched LLM Server to Block Banned Phrases
AlpinDale · x · 2026-09-08
AlpinDale's new post tackles a subtle serving problem: banning whole phrases (not tokens) from a model's streamed output. Banning token sequences directly has too large a blast radius, and by the time the engine detects a complete phrase, tokens may already be on the user's screen and baked into the KV cache.
- The solution: hold back output that could still complete a banned string; on a match, rewind that individual request and sample a different continuation.
- Implementing this inside an asynchronous scheduler is tricky because workers may already be executing later steps. The author built it in Sonar (formerly Aphrodite Engine).
- Prior art includes ExLlamaV3, KoboldCpp, and AntiSlop Sampler, but not in this async batched setting.
More from Infra
- Anthropic generated $2.08 revenue per $1 of compute spend in Q2 2026, up from $0.63 — cedric_chee · 2026-10-26
- CircuitHub unveils The Grid, a factory-scale robot making PCBs in the US twice as fast as Shenzhen — garrytan · 2026-10-26
- Ubicloud publishes "Fixing Tool-Call Accuracy in vLLM" — petrusenko_max · 2026-10-26
- Meta approves nonvoting Class C shares as 2026 capex guidance hits $135B–$145B — thione · 2026-10-26
- Unsloth Ships MiniMax-M3.1-Flash Quants: 2-bit Retains 91.5% Accuracy, Runs in 150GB RAM — danielhanchen · 2026-10-26
- Engineer with thousands of fabricated devices pours cold water on photonic crystal modulator yields — jwt0625 · 2026-10-26