I don’t think that’s the case, as the 1.10.0+cu113 pip wheel ships with these compute capabilities:
>>> import torch
>>> torch.__version__
'1.10.0+cu113'
>>> torch.cuda.get_arch_list()
['sm_37', 'sm_50', 'sm_60', 'sm_70', 'sm_75', 'sm_80', 'sm_86']
while yours returns:
sm_37 sm_50 sm_60 sm_61 sm_70 sm_75 compute_37
which is the CUDA 10.2 support.
If you get stuck, try to create a new virtual env, reinstall the right version, and verify the right versions in the install logs.