HKUST+Huawei routing boosts DeepSeek decoding throughput 15% on 96 NPUs
机器之心 · wechat · 2026-10-16
HKUST's Zhou Zijie team and Huawei's 2012 Lab propose BalanceRoute for the "lockstep" problem in DeepSeek-style MoE deployments: worker groups must sync every decode step, so lighter groups idle waiting for the busiest one, and equal request counts still mean unequal KV loads — accumulated waiting quietly erodes scaling gains.
Approach
- Mathematical modeling of load gaps yields an imbalance metric I(k), turning routing into a well-defined online optimization;
- BR-H uses an H-step prediction window estimating each running request's completion probability and remaining steps;
- Assignment weighs filling idle capacity against creating a new peak; requests stay put once assigned (KV migration too costly);
- A stateful agent atop vLLM-Ascend schedules with 1.20ms median overhead, well under the 60ms per-step decode budget.
Results: on 96 Ascend 910 NPUs running DeepSeek-V3 671B (W8A8), replaying enterprise and Azure-2024 traces, BR-H lifts decoding throughput 15.4% / 14.5% over the strongest baselines while cutting tail latency. Accepted at NeurIPS 2026 and deployed in the joint MathBot-Pie MaaS system.
More from Infra
- SGLang Summit sponsors grow from 11 to 36 in one week; Lip-Bu Tan, Lilian Weng to speak — BanghuaZ · 2026-10-18
- Yacine jokes his datacenter KPI is gigawatt-hours: get paged if not burning enough power — yacineMTB · 2026-10-18
- OpenAI reportedly handled ~15T tokens daily in April, likely higher now — JoshPurtell · 2026-10-18
- 1.68x inference speedup drops, directly accelerating GRPO post-training — yacineMTB · 2026-10-18
- OpenRouter is only a fraction of the total addressable market, poster argues — sloppenheimer · 2026-10-18
- Developer builds change alerts for LLM prices and retirements across Bedrock, Azure and Vertex — jsandhu0780 · 2026-10-18