layout() does not seem to respect spacy.require_cpu()

[marioangst]

I am trying to use spacy-layout (0.0.12) and spacy (3.8.4) on CPU for testing on an older machine. I am getting an OOM memory "CUDA out of memory (...)" when running (with a small test set of 14 files):

import spacy
from spacy_layout import spaCyLayout

spacy.require_cpu()
nlp = spacy.blank("en")

filepaths = <list of 14 filepaths>

docs = layout.pipe(filepaths)
for doc in docs:
    print(doc.text)

returns: "torch.OutOfMemoryError: CUDA out of memory. Tried to allocate (...)"

Does spacy-layout always require a GPU? https://explosion.ai/blog/pdfs-nlp-structured-data talks about how "Docling runs at 1-3 pages per second on CPU (...)", which is why I thought this might work on CPU.

[6801318d8d]

How did you obtain CUDA accelleration in the first place?

I'm trying to obtain that.

Looks like you need to specify it as an option in docling's DocumentConverter (see here), but we don't:

spacy-layout/spacy_layout/layout.py

Line 72 in 1e6a51a

self.converter = DocumentConverter(format_options=docling_options)