NameError: name 'gpu_ops' is not defined

[Caramazula]

Hi,
I am currently trying to set up an Ubuntu machine in order to train a spaCy model via GPU. Here is how I set everything up:

• I started on a clean Ubuntu 18 machine with Python 3.7 and cuda 10.0
• I created a venv
• I installed spaCy via pip install spacy[cuda10] as described on https://spacy.io/usage/#gpu which should install cupy and thinc_gpu_ops, but it didn't install either. Furthermore I got this warning spacy 2.0.16 does not provide the extra 'cuda10'. I run into a similar warning when installing thinc with pip install thinc[cuda10]
• I installed cupy

And then I just tried running the script and ran into this error:

  File "/home/cuda/venv/lib/python3.7/site-packages/thinc/neural/_classes/hash_embed.py", line 51, in begin_update
    keys = self.ops.hash(ids, self.seed) % self.nV
  File "ops.pyx", line 821, in thinc.neural.ops.CupyOps.hash
NameError: name 'gpu_ops' is not defined

Which has been discussed over here: explosion/spaCy#1745 . But none of the answers helped me with the problem. Which is probably because the newest thinc version (which I use) had its gpu ops moved into a sperate package.
Thats why I tried to install thinc-gpu-ops manually with pip install thinc-gpu-ops, but I run into an error while compiling:

thinc_gpu_ops/gpu_ops.cpp:7861:13: error: ‘PyThreadState {aka struct _ts}’ has no member named ‘exc_traceback’; did you mean ‘curexc_traceback’?
         tstate->exc_traceback = local_tb;
                 ^~~~~~~~~~~~~
                 curexc_traceback
    error: command 'gcc' failed with exit status 1

I tried downgrading my gcc to 6, but to no avail.

My Environment

• spaCy version: 2.0.16
• Platform: Linux-4.15.0-36-generic-x86_64-with-debian-buster-sid (Ubuntu 18.04.1 LTS)
• Python version: 3.7.0
• cupy 4.5.0
• thinc 6.12.0
• cuda 10.0
• g++ 7
• GPU GeForce GTX 1050

I am fairly certain that all of this is due to some version mismatch, but I don't want to try all combinations, can anyone tell me which setup of versions is working?

Many thanks!

[honnibal]

Hmm, it seems cupy doesn't support cuda 10 yet. I wasn't aware of this. Unfortunately you'll have to reinstall with CUDA 9.2.

[Caramazula]

Thank you for pointing that out. I tried it again using cuda 9.2. Furthermore I had to downgrade to Python 3.6.6 in order to install cupy-cuda92. And now it works!
The first few tests suggest a speedup of about 70%. That said, my GPU is only running at 12.2% right now.