Import broken for CPU-only machines: AttributeError: 'NoneType' object has no attribute 'split'

[k8si]

I'm trying to run some code that imports apex on my laptop (without a GPU) for debugging purposes, but the import throws an error. Version info and stacktrace below:

mac osx 10.13.6 (high sierra)
python 3.7.3
torch 1.1.0.post2

git clone https://github.com/NVIDIA/apex && cd apex
pip install -v --no-cache-dir ./

$ python3
Python 3.7.3 (default, Apr  9 2019, 13:13:38)
[Clang 10.0.0 (clang-1000.11.45.5)] on darwin
Type "help", "copyright", "credits" or "license" for more information.
>>> import apex
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/Users/kfsilvers/apex/apex/__init__.py", line 5, in <module>
    from . import amp
  File "/Users/kfsilvers/apex/apex/amp/__init__.py", line 1, in <module>
    from .amp import init, half_function, float_function, promote_function,\
  File "/Users/kfsilvers/apex/apex/amp/amp.py", line 3, in <module>
    from .lists import functional_overrides, torch_overrides, tensor_overrides
  File "/Users/kfsilvers/apex/apex/amp/lists/torch_overrides.py", line 77, in <module>
    if utils.get_cuda_version() >= (9, 1, 0):
  File "/Users/kfsilvers/apex/apex/amp/utils.py", line 9, in get_cuda_version
    return tuple(int(x) for x in torch.version.cuda.split('.'))
AttributeError: 'NoneType' object has no attribute 'split'

Seems like the case where torch.version.cuda returns None isn't handled properly in two places:

1. apex/amp/utils.py

 def get_cuda_version():
     return tuple(int(x) for x in torch.version.cuda.split('.'))

Seems like we could change this to:

def get_cuda_version():
    if torch.version.cuda is not None:
        return tuple(int(x) for x in torch.version.cuda.split('.'))
    else:
        return None

2. apex/amp/lists/torch_overrides.py

if utils.get_cuda_version() >= (9, 1, 0):
    FP16_FUNCS.extend(_bmms)
else:
    FP32_FUNCS.extend(_bmms)

We'll also need to add a check for None here:

cuda_version = utils.get_cuda_version()
if cuda_version is not None:
    if cuda_version >= (9, 1, 0):
        FP16_FUNCS.extend(_bmms)
    else:
        FP32_FUNCS.extend(_bmms)

Here's where I confess I know nothing about apex. Should there still be a call to FP32_FUNCS.extend(_bmms) in the case where cuda_version is None i.e. there is no GPU available, or no?

[ptrblck]

Hi @k8si,

thanks for reporting this!
amp is not usable without a GPU.
The reason for not throwing an error but just a warning, if no GPU was found during the setup, is that it might be sometimes useful for cross-compilation.

Anyway, I think the error message should be clearer in case you are trying to run amp without a GPU.

[DannyDannyDanny]

I've manually changed mine to:

def get_cuda_version():
    return tuple(int(x) for x in torch.__version__.split('.'))

Should I submit a PR?

[ptrblck]

@DannyDannyDanny Your workaround will not work, as explained here.