提问：能支持单机训练百亿参数的原理是蒸馏吗？那如果用自定义算法是不是也可以单机训练？又如何自定义算法？

chatGpt训练算法从transformer变成ppo以后。可以看到ai的效果和gpt3有了本质的提升。而强化学习中ppo算法还有很多提升空间。也还有不同的强化算法也有提升空间td3,deepnash。那么如果使用这些算法，并且加上集群的功能。那还了得。easyNLP可以自定义算法吗？怎么定义？又支持集群训练吗？怎么处理？以及能支持单机训练百亿参数的原理是蒸馏吗？

The issue names no files or tests. Start by reviewing EasyNLP's Python/PyTorch training entry points and documentation to determine whether custom PPO, TD3-like, or other algorithms and distributed training are supported. Done would require a concrete, documented answer about customization, cluster training, and the basis for single-machine training of very large models.

Written by the indexing model from the issue text.