`compile_autograd`

compiler
1

What is the purpose of the compile_autograd argument in the dynamo benchmarking` script?

I thought that forward / backwards graphs were automatically compiled for a nn.Module when using torch.compile (with default inductor backend)?