biberplus GPU-usage error

[mjgaughan]

Hello! I wanted to flag what I think is an error when I try to use biberplus with the GPU/CUDA on my machine. When I follow the instructions from the README:

from biberplus.tagger import load_config, load_pipeline, calculate_tag_frequencies

config = load_config()
config.update({'use_gpu': True, 'n_processes': 4, 'function_words': False})
pipeline = load_pipeline(config)
frequencies_df = calculate_tag_frequencies("Your sample text goes here", pipeline, config)
print(frequencies_df)

I receive the error

Traceback (most recent call last):
  File "/XXX/git/YYY/biberplus_labeling.py", line 57, in <module>
    preds = biberplus_labeler(docs)
            ^^^^^^^^^^^^^^^^^^^^^^^
  File "/XXX/git/YYY/biberplus_labeling.py", line 39, in biberplus_labeler
    pipeline = load_pipeline(config)
               ^^^^^^^^^^^^^^^^^^^^^
  File "/XXX/.conda/envs/neurobiber/lib/python3.11/site-packages/biberplus/tagger/tagger_utils.py", line 39, in load_pipeline
    spacy.require_gpu()
  File "/XXX/.conda/envs/neurobiber/lib/python3.11/site-packages/thinc/util.py", line 225, in require_gpu
    raise ValueError("Cannot use GPU, CuPy is not installed")
ValueError: Cannot use GPU, CuPy is not installed

Even though I have confirmed that I have installed CuPy 13.5.1 and can access a satisfactory GPU. Not quite sure what this would be, though it may just be an issue of me installing incompatible versions, etc.

For now I'll probably just run biberplus with the CPU but I may return to this issue later.

[joelb-git]

I had a similar error in a different context a while ago, but it's likely the same thing.
spacy or its dependencies is very picky about cuda versions. Here's how I repro'd
my case in isolation:

$ CUDA_HOME=/nas/sadiri/users/joelb/tmp/cuda_11.8.0
$ PATH=$CUDA_HOME/cuda_nvcc/bin:$PATH
$ export CPATH=$CUDA_HOME/cuda_cudart/targets/x86_64-linux/include
$ export LIBRARY_PATH=$CUDA_HOME/cuda_cudart/targets/x86_64-linux/lib
$ export LD_LIBRARY_PATH=$CUDA_HOME/cuda_cudart/targets/x86_64-linux/lib

$ conda create -n pii python=3.12 -y
$ conda activate pii
$ pip install -U pip setuptools wheel
$ pip install -U 'spacy[cuda11x]'
$ python -m spacy download en_core_web_sm
$ python
>>> import spacy
>>> print(spacy.require_gpu())
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/nas/material02/users/joelb/conda_envs/pii/lib/python3.12/site-packages/thinc/util.py", line 230, in require_gpu
    raise ValueError("Cannot use GPU, CuPy is not installed")
ValueError: Cannot use GPU, CuPy is not installed

But this is masking a lower level error:

$ python
>>> import cupy
Traceback (most recent call last):
  File "/nas/material02/users/joelb/conda_envs/pii/lib/python3.12/site-packages/cupy/__init__.py", line 17, in <module>
    from cupy import _core  # NOQA
    ^^^^^^^^^^^^^^^^^^^^^^
  File "/nas/material02/users/joelb/conda_envs/pii/lib/python3.12/site-packages/cupy/_core/__init__.py", line 3, in <module>
    from cupy._core import core  # NOQA
    ^^^^^^^^^^^^^^^^^^^^^^^^^^^
ImportError: libcudart.so.11.0: cannot open shared object file: No such file or directory
...

which points to a failure loading the cuda libs.

Here's how I worked around it:

$ wget https://developer.download.nvidia.com/compute/cuda/12.4.0/local_installers/cuda_12.4.0_550.54.14_linux.run
$ mkdir /nas/sadiri/users/joelb/tmp/cuda_12.4.0
$ sh cuda_12.4.0_550.54.14_linux.run --toolkit --toolkitpath=/nas/sadiri/users/joelb/tmp/cuda_12.4.0

in the install text ui:
* unselect driver
* install

$ export PATH=/nas/sadiri/users/joelb/tmp/cuda_12.4.0/bin:$PATH
$ export LD_LIBRARY_PATH=/nas/sadiri/users/joelb/tmp/cuda_12.4.0/lib64

$ conda create -n pii python=3.12 -y
$ conda activate pii
$ pip install -U pip setuptools wheel
$ pip install -U 'spacy[cuda12x]'
$ python -m spacy download en_core_web_sm
$ python
>>> import spacy
>>> print(spacy.require_gpu())
True

bingo!

[davidjurgens]

Thanks Joel!!! That is some amazing debugging and I hope it helps folks who run into the same cupy issue.

[mjgaughan]

Thank you Joel! This is quite helpful!