Spacy 2.1 bug when running with GPU: 'gpu_ops' is not defined

[01alex]

How to reproduce the behaviour

!pip install -U spacy
!python -m spacy download en

import spacy

gpu = spacy.prefer_gpu()
print('GPU:', gpu)

nlp = spacy.load("en_core_web_sm")

doc = nlp("This is a sentence. This is another sentence.")
for sent in doc.sents:
    print(sent.text)

Output

GPU: True
---------------------------------------------------------------------------
NameError                                 Traceback (most recent call last)
<ipython-input-3-1313493bc47b> in <module>()
      6 nlp = spacy.load("en_core_web_sm")
      7 
----> 8 doc = nlp("This is a sentence. This is another sentence.")
      9 for sent in doc.sents:
     10     print(sent.text)

(...)

ops.pyx in thinc.neural.ops.CupyOps.hash()

NameError: name 'gpu_ops' is not defined

This exact same code worked yesterday with the previous release.

Note

If I do:

!pip uninstall thinc
!pip install thinc==7.0.3

It works correctly:

GPU: True
This is a sentence.
This is another sentence.

Your Environment

• spaCy version: 2.1.0
• Platform: Linux-4.14.79+-x86_64-with-Ubuntu-18.04-bionic
• Python version: 3.6.7
• Models: en
• Environment Information: Google Colaboratory (Python 3 Google Compute Engine backend (GPU)), CUDA 10.0

[ines]

Did you install the new version in the same environment as the old one, and did you specify your CUDA version as described here? It's possible that you ended up with stale files or packages in your site-packages, or Thinc didn't correctly pull in the thinc_gpu_ops.

[01alex]

So I ran two new Python 3 Google Compute Engine backend (GPU).
The first with the default spaCy 2.0.18.

The example:

import spacy

nlp = spacy.load("en_core_web_sm")
doc = nlp("This is a sentence. This is another sentence.")
for sent in doc.sents:
    print(sent.text)

It works fine. Then I required the gpu after import spacy, it returns true and throws the error.
In another runtime, I started with:

!pip install -U spacy[cuda100]
!python -m spacy download en

The update output:

Installing collected packages: srsly, numpy, wasabi, blis, thinc, spacy
  Found existing installation: numpy 1.14.6
    Uninstalling numpy-1.14.6:
      Successfully uninstalled numpy-1.14.6
  Found existing installation: thinc 6.12.1
    Uninstalling thinc-6.12.1:
      Successfully uninstalled thinc-6.12.1
  Found existing installation: spacy 2.0.18
    Uninstalling spacy-2.0.18:
      Successfully uninstalled spacy-2.0.18
Successfully installed blis-0.2.4 numpy-1.16.2 spacy-2.1.0 srsly-0.0.5 thinc-7.0.4 wasabi-0.1.4

Checked version 2.1.0. Ran the example snippet. It works fine on cpu. Required gpu and got the same error.
It was working fine yesterday with spaCy 2.0.18.

[honnibal]

Thanks, this was a regression introduced when fixing #3430 .

Thinc v7.0.3 required the thinc_gpu_ops module, which caused problems if CUDA was installed but no GPU was available (or if no compiler was installed).

The fix is to make the thinc_gpu_ops module an extras_require. This was set correctly in Thinc v7.0.4, but I didn't set it in spaCy. Have triggered a build for v2.1.1 to fix this.

[lock]

This thread has been automatically locked since there has not been any recent activity after it was closed. Please open a new issue for related bugs.