"Array device must be same as the current device" issue "with nlp.pipe()"

[fatihbozdag]

Greetings,

I am running Spacy on Colab. However, I got the error 'Array device must be same as the current device: array device = -2 while current = 0' when using nlp.pipe()

code as follows:

!pip install -U pip setuptools wheel
!pip install -U spacy
!pip install spacy_transformers -U
import spacy
import spacy_transformers
spacy.require_gpu()
!python -m spacy download('en_core_web_trf')
nlp = spacy.load('en_core_web_trf')

TIME = []
TIME_LEMMA = []
TIME_START = []
TIME_NATIVE_LANG = []

for doc, context in nlp_pipe(CORPUS as_tuples =True):
    for token in doc:
        if token.dep_ == 'advmod' and token.lemma_ in TIME_SUBORDINATORS:
            TIME.append(token.sent)
            TIME_LEMMA.append(token.lemma_)
            TIME_START.append(token.is_sent_start)
            TIME_NATIVE_LANG.append(context['Native_language'])

cupy/_core/core.pyx in cupy._core.core.ndarray.copy()
cupy/_core/core.pyx in cupy._core.core.ndarray.copy()
cupy/_core/core.pyx in cupy._core.core.ndarray.astype()
cupy/_core/_kernel.pyx in cupy._core._kernel.ufunc.__call__()
cupy/_core/_kernel.pyx in cupy._core._kernel._preprocess_args()
cupy/_core/_kernel.pyx in cupy._core._kernel._check_array_device_id()
ValueError: Array device must be same as the current device: array device = -2 while current = 0

However, without for doc, context in nlp_pipe(CORPUS, as_tuples =True): code simply works.

    for token in doc:
        if token.dep_ == 'advmod' and token.lemma_ in TIME_SUBORDINATORS:
            TIME.append(token.sent)
            TIME_LEMMA.append(token.lemma_)
            TIME_START.append(token.is_sent_start)
            TIME_NATIVE_LANG.append(context['Native_language'])

So, I assume I should define GPU on nlp.pipe() ? I haven't had this issue on my local computer, though. Also without importing spacy_transformers, I got no GPU detected

[fatihbozdag]

Solved! by replacing code from

spacy.require_gpu()
nlp = spacy.load('en_core_web_trf')

to

nlp = spacy.load('en_core_web_trf')
spacy.require_gpu()

I am not sure about the logic here and do not know if it was mentioned before, but accidentally changing the lines worked!

[polm]

Glad you figured it out, thanks for sharing the solution! I am also not sure why that would work but it may have to do with some of the special handling for GPU support colab has - we'll look into it!

[fatihbozdag]

Sure thing! I've added some details here for another user having the same problems. I guess it is more to do with Jupyter than Colab GPU handling.