Setup.py introduced sm100/sm120 regardless of cuda version being used.
When a user is using cuda 12.6 still, they would encounter failures due to the presence of sm100/sm120.

"nvcc fatal : Unsupported gpu architecture 'compute_100'"

Creating this issue to track the fix.

@ptrblck pointed out that https://github.com/NVIDIA/apex/blame/master/setup.py#L534 needs to be indeed improved. cc @crcrpar