[PyTorch 1.6]
I try to torch.jit.trace or torch.jit.script alexnet, the graph displays “prim::DifferentiableGraph”. However, sometimes, it only fuses some small network into a “FusionGroup”.
When prim::DifferentiableGraph and FusionGroup happens?
If I want to use cuda fusion(nvrtc), how to close prim::DifferentiableGraph and turn on “FusionGroup”?
Thanks.