Test multi-gpu setup

[BrownianNotion]

Description

Spin up an instance on the cloud with GPU and test that the training/eval pipeline works with multiple GPUs.
Start by changing --num_gpus=1 to 2. Understand and attempt to fix any basic errors encountered, otherwise close the issue and report the errors.

Multi-GPU training would allow us to:

• train faster (at the cost of paying more)
• train larger models like 3/7B potentially

Checklist before resolving issue

• I have included the issue number in my PR description
• I have rebased/pulled in changes from main and fixed any conflicts
• If I have changed the models/training code, I have run dry_run.sh on a CUDA device successfully