Miles 在 NVIDIA Vera 上训练智能体RL
RadixArk 与 NVIDIA 合作运行其开源 Miles 后训练框架,端到端训练的每个隔离 Firecracker 沙箱均托管在 NVIDIA Vera CPU 上。该运行以同步和全异步模式,对 GLM-4.7-Flash 训练了 68 步、覆盖逾 2,400 个 Terminal-Bench 和 SWE-bench 任务回合,其中一个 gRPC 任务的通过率由前半程的 40%升至后半程的 82%。RadixArk 还将其与 AMD EPYC Turin 和 Intel Xeon 61 系统比较,报告称在测试的繁忙服务器条件下,Vera 上的任务回合快 12%。
该报告表明,CPU 沙箱吞吐量会显著影响大规模智能体强化学习训练,并为评估 Vera 的基础设施团队提供了具体训练与对比数据。
On this page
- End-to-end training with every sandbox on Vera
- Episodes finish sooner in real RL training
- Vera stays fast at full load
- What this means for RL training
- Conclusion
In agentic RL, the model does more than write text: it runs commands in a sandbox, compiles code and runs tests, and decides its next step from the results. A training run can keep hundreds of sandboxes or more open at once, and the CPUs that run them can affect the speed and scale of training.
RadixArk partnered with NVIDIA to run Miles, our open-source RL post-training framework for LLMs and VLMs, with every sandbox on NVIDIA Vera CPUs. To our knowledge, this is the first public result of end-to-end agentic RL training with its sandboxes on Vera CPU servers. We also compared Vera systematically with the current-generation AMD EPYC Turin and Intel Xeon 61. The conclusion is clear: Vera runs sandboxes fast even at full load, and its advantage grows as training gets larger and inference gets faster.
#Miles#NVIDIA Vera#RadixArk#GLM-4.7-Flash#AgentENV#Firecracker#SGLang#AMD EPYC Turin#Intel Xeon 61#Terminal-Bench#SWE-bench#NVIDIA