Thinc GPU capability on Windows - fails to compile

[ivmreg]

It looks like Thinc has some hard time making GPU computation available if used on Windows.

Similar issues were raised in spaCy and thinc:
explosion/spaCy#1745
#50

It seems related in how the build is set-up in setup.py:
https://github.com/explosion/thinc/blob/master/setup.py#L56

Based on quick search, others are struggling with this as well:
rbgirshick/py-faster-rcnn#11

Some more info (they came to a working solution there):
cudamat/cudamat#46 (comment)

Will research a bit more later.

[honnibal]

Sorry this is hard. I don't use Windows, so am not of much help. You might try the wheels here? https://www.lfd.uci.edu/~gohlke/pythonlibs/ Some advertise CUDA linkage. It might solve your problem.

[pythonBerg]

Has there been a resolution to this? Right now using linux through VM so I do not have access to GPU. Windows solution would be extremely welcome...

[ericstarr]

I'm also having trouble getting GPU working on Windows. I'm not sure what issue you are having, but in my case, it's the fact that thinc.neural.gpu_ops isn't getting created.

I should note that I'm trying to do this using CUDA v9, so some of the issue may be due to that. Also, I just started digging into this today, so I may be way off.

I think the issue is in the locate_cuda function in setup.py. I think there are two separate issue with the function.

First, the following snippet fails to see nvcc, even though it's there.

for k, v in cudaconfig.items():
    if not os.path.exists(v):
        print('Warning: The CUDA %s path could not be located in %s' % (k, v))
        return None

I think this is because
os.path.exists('[CUDA_HOME]\bin\nvcc) returns False,
but
os.path.exists('[CUDA_HOME]\bin\nvcc.exe) returns True.

I'm guessing this is a platform dependent issue.

Also, the code is looking for a lib64 directory under the CUDA_HOME directory. In my case, this doesn't exist, though there is a lib directory with an x64 directory under that. I assume that is the directory it should be pointing to. I'm not sure whether this is a platform specfic issue or an issue due to the fact that I'm using CUDA v9.