torch.compile errors on torch.autograd.backward #125287
Description
Min repro:
import torch
@torch.compile(backend='aot_eager')
def f(x):
y = x.sin().sin()
torch.autograd.backward([y], [torch.ones_like(y)])
x = torch.ones(4, requires_grad=True)
f(x)
print(x_ref.grad)
gives:
RuntimeError: Cannot backprop through mirrored meta, file a bug in PyTorch
It looks like we are not properly skipping autograd.backward in dynamo's trace rules
cc @ezyang @msaroufim @anijain2305 @chauhang @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @chenyang78 @kadeng