Each iteration takes about 15 seconds. Does it support multi-GPU training? Is this project only run on cuda:0？

It's a great honor to study the code you've open-sourced! I just got it running on an RTX 4090:
Total timesteps: 10813440
Iteration time: 13.94s
Total time: 1559.45s
ETA: 140223.4s

training world model time: 2.0059447288513184
acquiring depth image time: 2.2802934646606445
acquiring depth image time: 2.250629425048828
acquiring depth image time: 2.3184289932250977
acquiring depth image time: 2.3598833084106445
acquiring depth image time: 2.4244799613952637

Is this duration normal? Each iteration takes about 13.94 seconds in total.

The issue names no files, tests, or entry points, and reports timing from training on an RTX 4090. Start by locating the training entry point and checking how the device is selected and whether multi-GPU execution is documented. Done means establishing whether the duration is expected and whether multi-GPU training is supported.

Written by the indexing model from the issue text.