[AutoTVM] Call set_cuda_target_arch by default

Questions
1

Hi,

We are currently tuning some models on g4 instance (T4 GPU) and noticed that the following setting is able to boost the performance (from @trevor-m ):

from tvm.autotvm.measure.measure_methods import set_cuda_target_arch
set_cuda_target_arch('sm_75')

It seems always better to enable this by default, but we still need to set it manually, so I have some questions:

  1. Why don’t we enable it by default?

  2. If we set this flag for AutoTVM, how should we reproduce the result when building the model after tuning?

Thanks in advance.

cc @merrymercy @vinx13 @Laurawly