训练对齐器，显存溢出

[angelOnly]

看了训练流程图，我理解这个对齐器是不是在全参微调，我跑百川7b的模型，4090 24G的显卡，跑不起来，显存满了，只能换更大的显存吗？多大的显存合适？

[PKU-Aligner]

是的，对齐器也是在全参数微调，通过Q-A-C来进行残差微调。 训练一个7B的模型，和SFT所需要的资源是一致的，你可以考虑lora或者更小尺寸的aligner，比如可以拿qwen1.5-2B来训练对齐器，是可以直接在3090上训练了。

Yes, aligners are also fine-tuned with full-parameter fine-tuning, using Q-A-C for residual fine-tuning. Training a 7B model requires the same resources as SFT, you can consider using lora or a smaller size aligner, such as training an aligner with qwen1.5-2B, which can be trained directly on 3090s.