QwenGyre: Qwen Cuts Long-Horizon Agent RL Training Time by 44%
量子位 · wechat · 2026-10-16
The Qwen team built QwenGyre to tackle GPU underutilization in long-horizon agent RL, where rollouts can run for hours with million-token contexts. Colocate mode idles waiting for the slowest agent; Async mode starves either inference or training.
Two components fix this. An Elastic Scheduler splits the cluster into role-switchable Cells with water-level monitoring: idle GPUs switch to training on the fly, with lossless migration via request re-routing and RDMA KV-Cache transfer (8.52s to switch to training, 3.46s back). A Trajectory Processor handles the branching trajectory trees left by black-box agents (Claude Code): TITO with prefix-tree sharing, partial credit from an evaluator for timed-out tasks, main-agent-priority sampling with a budget of 5, and masking of non-model tokens.
On Qwen3.8 2.4T with NL2RepoBench, 48 RL steps took 75.42 hours vs 134.55 for Async (1.78×); pass rate rose from 52.48% to 58.54%. A trajectory budget of 5 matches full-trajectory results while cutting compute 25.2%. The framework will be open-sourced.
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