/algoperf_results

Primary LanguagePython

Case Studies

criteo1tb

1. aten::index

reference: pytorch/pytorch#41162 (anonymous namespace)::indexing_backward_kernel from 30.5s to 3.80s

ogbg

1. aten::index

the same optimization as cirteo1tb.

fastmri

1. dataloader thread tuning

core number = 12, tune parallel workers from 16 to 8, got 7-8s improvement.

2. dataloader prefetching

3. NCHW data transform

before optimization: cudnn::ops::nchwToNhwcKernel, 2.11s, 15.4%

wmt

1. loss_fn(also available in ogbg)

we reduce loss function time by torch.compile. before torch.compile: model_fn 21.9s, loss_fn 7.36s. after torch.compile: model_fn 21.9s, loss_fn 0.89s.

ViT

1. cudnn deprecated