Qwen 团队发布 QwenGyre：面向超长程智能体训练的弹性强化学习框架
Qwen 开源了个新训练框架 QwenGyre，GPU 能在 rollout 和训练间动态调度，跑 70 万 token 的长任务智能体训练比之前快近一半，做 RL 的可以看看。
QwenGyre 是 Qwen 团队推出的强化学习训练框架，包含两个核心设计：Elastic Scheduler 可在 rollout 与训练之间动态重新分配 GPU，训练过程不被中断，相比 Colocate 提速最高 1.85 倍，相比 Async 提速 1.78 倍；Trajectory Processor 将分支执行（压缩、子智能体、重试）转化为前缀共享的轨迹树，去重后得到有界训练样本，超时后的部分进展也能获得评分。该框架已扩展到 Qwen3.8-2.4T-A95B，单次 rollout 约 70 万 token，在 48 步内将 NL2RepoBench 成绩从 52.5% 提升到 58.5%，端到端训练时间较 Async 缩短约 44%。相关论文已发布在 arXiv（2609.33848）。
Introducing QwenGyre, bringing elastic RL to extreme-long-horizon (XLong) agent training. ✨ It launches with two core designs: 🥳 - Elastic Scheduler: reallocates GPUs between rollout & training as demand shifts—live executions never interrupted. Up to 1.85× speedup over Colocate, 1.78× over Async. - Trajectory Processor: turns branching executions (compaction, sub-agents, retries) into prefix-shared trajectory trees with bounded, deduplicated training samples. Partial progress after timeout still gets scored. Scaled to Qwen3.8-2.4T-A95B with ~700K tokens per rollout, QwenGyre lifts NL2RepoBench from 52.5% to 58.5% in 48 steps, cutting end-to-end training time by ~44% vs Async. 🔗 Read the paper arxiv.org/abs/2609.33848 sk 💬 18 🔄 13 ❤️ 138 👀 5780 📊 36 ⚡
- IT之家11:20原文
