Qingguo Hu

> I also encountered a similar error. My solution was to switch to deepspeed2. I hope my suggestions can help you. @zfw-cv Thank you, but using `deepspeed_stage_2` will cause an...

> 你不是用到了deepspeed_stage_zero3，换zero2应该可以跑 [@qiji2023](https://github.com/qiji2023) @Rookienovice 您好。请问为何 stage_3 不能用？stage_2 对单卡显存要求太高了，顶不住。

The "/cluster/balrog/jtang/" seem to be written in the .pkl file, which you would not like to modify. `ThreedFutureDataset.from_pickled_dataset` creates a dataset directly without `__init__`, so modifying `ThreedFutureModel.__init__` won't work either....

Hello. How do you choose the `epochs` and `steps_per_epoch`? In the paper the authors mentioned they used `100k iterations`. But in the config file they wrote `10000 epochs` and `500...