Training time and GPU memory of distilllation stage

[meng-zha]

Thanks for the great work. I use one 4090 gpu with 24G memory to reproduce the project. While the distillation stage, OOM came out when set "multi_process_" False and use the default config. Since then I reduced the "num_envs" and grid sizes, the training time is extremely long to 30 days. So, I would like to know about:

• the nums_envs, num cols, num rows for 24G 4090
• the normal training time of distillation stage