Does LibMTL support DeepSpeed for distributed training?

[ZeliaYu]

Hi, thanks for the great work on LibMTL!

I’m currently experimenting with large-scale multi-task training, and I’m wondering whether DeepSpeed integration is supported or recommended with LibMTL.

Specifically, I would like to know:

1. Does LibMTL officially support DeepSpeed for distributed or memory-efficient training?
2. If not directly supported, is there any recommended way to integrate DeepSpeed?
3. Are there any examples or plans to support DeepSpeed in the future?

Since DeepSpeed is commonly used for large-scale model training and memory optimization, it would be very helpful for scaling LibMTL to larger models and datasets.

Thanks!