Can anybody enlighten me on the benefit of designating model of device? To be specific, let’s say we compile our model on the NVIDIA Jetson device, and write the following code:
target = tvm.target.Target("cuda -model=tx") # 1
target = tvm.target.Target("cuda") # 2
net, params, input_shape, output_shape = get_network(network, batch_size=1)
with tvm.transform.PassContext(opt_level=3):
lib = relay.build(net, target=target, params=params)
My question here is that how is #1 better than #2 regarding compilation or autotuning operator?