Content Comparison

TL;DR: This is a guide for using the University at Albany's DGX On-Prem cluster's free tier. Key points:

1. Free tier limits: Max 4 concurrent jobs, up to 4 GPUs/256 CPUs, 7-day job duration limitplease refer to the Service Level Agreement (SLA).

2. Jobs can be preempted (stopped and requeued) if paid-tier users need resources. When this happens, all progress is lost unless properly saved.

3. The solution is checkpointing - regularly saving your work's state (like autosave in a document). The guide provides PyTorch code examples for:

   • Saving/loading model training progress

   • Tracking training metrics

   • Managing inference tasks

The main message is: Always implement checkpointing in your jobs, or you risk losing hours/days of work when preemption occurs.