Can pytorch still work on old GPU if build from source?

1

My environment is

What works:

  • compare running cuda-sample:nbody container on GPU vs CPU shows GPU acceleration is available:
    ** GPU 450.378 single-precision GFLOP/s
    ** CPU 0.626 single-precision GFLOP/s
  • torch.cuda.is_available() returns True.

What doesn’t work:

NVIDIA GeForce 930MX which is of cuda capability 5.0.
PyTorch no longer supports this GPU because it is too old.
The minimum cuda capability supported by this library is 5.2.

I heard build from source can make old GPU work, so I tried:

python setup.py develop

It took a few hours to build, surprisingly there was no error, pytorch is successfully built, thanks to docker.
However, pytorch still does not work on the old GPU.

image
image992×440 22.8 KB

Can pytorch still work on old GPU if build from source?
Would recompile with TORCH_USE_CUDA_DSA make it work? What other flags should I enable/disable when recompile?

2

How did you install PyTorch as the latest nightly with CUDA 12.1 should support compute capabilities 5.0+ and the nightly with 11.8 should support 3.7+:

torch.cuda.get_arch_list()
['sm_50', 'sm_60', 'sm_70', 'sm_75', 'sm_80', 'sm_86', 'sm_90']
3

For me, it returns:
['sm_52', 'sm_60', 'sm_61', 'sm_70', 'sm_75', 'sm_80', 'sm_86', 'sm_90', 'compute_90']
Guess I should use cuda toolkit 11.8 instead.
Thank you very much.