知识卡片
资源随候选数与图保留增长
每题多条生成增加采样时间,保存每条有梯度评分图又增加激活内存;训练还占优化器状态和梯度。本章小组、短长度是教学资源取舍,不能从 .6B 参数量或权重文件大小推断完整训练内存。
这张卡片看不懂?
这张卡片已通过人工审查,不再接受失败反馈。
已有反馈 · 0 条
还没有人提交反馈。
在整书笔记中的位置
原书出处 · 1 处
- · 查看原书依据
The number of rollouts (num_rollouts=4) and allowed tokens (max_new_tokens=512) are relatively small to reduce resource requirements. If you experience out-of-memory errors, you can further reduce these values, for example, by setting num_rollouts=2 and max_new_tokens=64 for testing purposes.
参考来源
- 位置:《Build a Reasoning Model (From Scratch)》第06章《强化学习训练推理模型》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Build a Reasoning Model (From Scratch)/_epub-src/OEBPS/Text/chapter-6.html)
- 结论依据:摘录中关于资源随候选数与图保留增长的定义、比较或条件支撑卡片;原书表述不严谨的范围已在覆盖记录逐项限定,不把案例结果外推为普遍结论。
- 原始内容:The number of rollouts (num_rollouts=4) and allowed tokens (max_new_tokens=512) are relatively small to reduce resource requirements. If you experience out-of-memory errors, you can further reduce these values, for example, by setting num_rollouts=2 and max_new_tokens=64 for testing purposes.