What slows down your PyTorch training runs?

#1
by iparaco - opened

We’re developing Virtual LPU at IPARA to help identify what slows down PyTorch training. Has your GPU ever spent more time waiting for the next batch than you expected?

Share your GPU model, training workload, and what helped you find the cause. If you tested a change, such as adjusting the number of DataLoader workers, we’d love to see the throughput before and after in samples per second.

We’ll use these examples to decide which measurements to add to Virtual LPU next.

Sign up or log in to comment