amp + eager backend + training failing for some timm models #97382
Description
🐛 Describe the bug
python benchmarks/dynamo/timm_models.py --backend eager --only mnasnet_100 --accuracy --amp --training
Results in
[2023-03-22 21:57:03,055] torch._dynamo.utils: [ERROR] RMSE (res-fp64): 0.09759, (ref-fp64): 0.01743 and shape=torch.Size([8, 1000])
for me.
interestingly, switching to the aot_eager backend allows it to pass.
python benchmarks/dynamo/timm_models.py --backend eager --only mnasnet_100 --accuracy --amp --training
The cause of this breakage is #95416.
cc: @yanboliang @davidberard98
Versions
N/A
cc @ezyang @soumith @msaroufim @wconstab @ngimel @bdhirsh @voznesenskym @penguinwu @anijain2305 @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @Xia-Weiwen @wenzhe-nrv @jiayisunx @desertfire