Merged PR #925 breaks CPU-Only installs

[aatifjiwani]

While running a Github CD job, it attempted to install Apex, but failed with the following traceback:

Traceback (most recent call last):
      File "<string>", line 1, in <module>
      File "/tmp/pip-req-build-f1vgc77y/setup.py", line 35, in <module>
        _, bare_metal_major, _ = get_cuda_bare_metal_version(cpp_extension.CUDA_HOME)
      File "/tmp/pip-req-build-f1vgc77y/setup.py", line 14, in get_cuda_bare_metal_version
        raw_output = subprocess.check_output([cuda_dir + "/bin/nvcc", "-V"], universal_newlines=True)
    TypeError: unsupported operand type(s) for +: 'NoneType' and 'str'

It seems that cuda_dir is None, which is intuitively expected when installing Apex on a CPU only machine. But even when installing on a CPU only machine, the setup is still looking for the Nvidia Cross-Compilers.

The most recent commit as of 8/4/2020 (hash: 5b53121...) adds this line into setup, which breaks the install.

The corresponding line within the current release of Apex:

apex/setup.py

Line 14 in 5b53121

raw_output = subprocess.check_output([cuda_dir + "/bin/nvcc", "-V"], universal_newlines=True)

[matthaeusheer]

I can confirm this issue.

[sugeeth14]

I am facing the same issue as well but in my case I am facing the issue when trying to install apex in a GPU machine. Was anyone able to fix it ?

[zifuwanggg]

same here

[aatifjiwani]

I am facing the same issue as well but in my case I am facing the issue when trying to install apex in a GPU machine. Was anyone able to fix it ?

@raghava14 The fix that we were able to make was install Apex using the commit before the one that broke the installation. Obviously this is not good in the long run but its a fine temporary fix

[ptrblck]

Thanks for reporting this issue. The get_cuda_bare_metal_version has to be moved inside the MHA code.
I'll provide a fix tomorrow for it.

[mcarilli]

Please do not consider the fix an endorsement of Apex as a future-proof source of mixed precision. torch.cuda.amp is the truth. It's much easier to use. The modularity and flexibility makes it almost fun.

[vince62s]

Even if torch.cuda.amp must be the ground truth :) if we need fusedadam, Apex is still required, and can confirm #937 does not fix the issue. Needs to be reopen.

[monk1337]

Run this command export TORCH_CUDA_ARCH_LIST="compute capability" before installing apex from source code.