multi gpu train sdxl

I think there is a need for a neat multi GPU documentation now. We cannot access any information about how it should be used, what parameters, how step and epoch should be calculated, and what logic it works on. We are trying to understand through issues and conversations, but we cannot get a healthy result. Please explain this topic a little more specifically and give more support to multi GPU, which is now a great need.

No files, tests, or entry points are named. Start by locating the existing multi-GPU and SDXL training guidance and reviewing the referenced issue discussions; done means documenting usage, parameters, step and epoch calculations, and the underlying logic with clearer multi-GPU support guidance.

Written by the indexing model from the issue text.