Installing with pytorch 2.0

[samar-khanna]

Thanks to the authors for making their very helpful code publicly available. This is for those of you working with newer versions of torch.

Currently, if you try to run python setup.py install with torch 2.0 (python3.8, cuda 11.8), it will fail to compile. To solve this, you need to add just one line in setup.py on L30:

        nvcc_flags += ['-std=c++14']

EDIT: This actually didn't work, it seemed to compile for me because it clashed with an existing build but fails if I try to build from scratch.

[bytesnake]

replacing dynamic dispatching with restricted version from ATen works for me (I also had to downgrade gcc to 10 NVlabs/instant-ngp#119)

#undef AT_DISPATCH_FLOATING_AND_COMPLEX_TYPES
#define AT_DISPATCH_FLOATING_AND_COMPLEX_TYPES(TYPE, NAME, ...)                                  \
  AT_DISPATCH_SWITCH(TYPE,NAME,                                           \
      AT_DISPATCH_CASE(at::ScalarType::ComplexFloat, __VA_ARGS__) \
      AT_DISPATCH_CASE(at::ScalarType::Float, __VA_ARGS__))