[Question] Best Practices for Multi-Turn Long-Context GRPO Training

Checklist / 检查清单

• I have searched existing issues, and this is a new question or discussion topic. / 我已经搜索过现有的 issues，确认这是一个新的问题与讨论。

Question Description / 问题描述

Hi,

I’m currently working on multi-turn GRPO training with Qwen3.5 series models under long-context settings (64K–128K+ tokens).

In my previous attempts at long-context training, I kept running into CUDA Out-of-Memory (OOM) errors, so I’d like to ask whether there are any recommended configurations or best practices for this scenario.

I’m particularly interested in the following:

• For 64K–128K+ context lengths, what kind of hardware configuration would you recommend? Are there any recommended memory optimization techniques?
• If using vLLM for rollout, how should it be configured to reduce GPU memory usage while maintaining good training/rollout throughput?
• Are there any existing configurations, training scripts or best-practice documentation that I could use as a reference?

If there is already a recommended setup or best-practice guide for Qwen + long-context GRPO, I would really appreciate any pointers.

Thanks.

The issue names no files, tests, or entry points. Start by checking the repository for existing GRPO, Qwen, long-context, or rollout configuration examples; done would mean identifying or documenting a supported setup for 64K–128K+ training, including hardware and memory guidance.