cannot get spacy properly installed to use GPU because of conflict with pytorch or cupy or thinc

[euricocovas]

Hi

I have a Windows 10 PC with card is NVIDIA GeForce RTX 2060
Compute capability 7.5
CUDA SDK 11.0 – 11.2 support for compute capability 3.5 – 8.6

nvcc --version
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2020 NVIDIA Corporation
Built on Wed_Jul_22_19:09:35_Pacific_Daylight_Time_2020
Cuda compilation tools, release 11.0, V11.0.221
Build cuda_11.0_bu.relgpu_drvr445TC445_37.28845127_0

I had to do a fresh install, so removed anaconda totally, reboot, reinstalled anaconda Anaconda3-2020.11-Windows-x86_64.exe, then did

conda activate
conda update -c conda-forge --all --yes
conda update -c conda-forge anaconda --yes
conda install pytorch torchvision torchaudio cudatoolkit=11.0 -c pytorch -c conda-forge --yes
conda install -c conda-forge thinc --yes
conda install -c conda-forge spacy --yes

Then I see this:

import torch
torch.cuda.is_available()
replies False

import spacy
spacy.require_gpu()
replies False

Then I did

conda install -c conda-forge cupy --yes

Then I see this:

import spacy
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "C:\ProgramData\Anaconda3\lib\site-packages\spacy\__init__.py", line 10, in <module>
    from thinc.api import prefer_gpu, require_gpu, require_cpu  # noqa: F401
  File "C:\ProgramData\Anaconda3\lib\site-packages\thinc\__init__.py", line 5, in <module>
    from .config import registry
  File "C:\ProgramData\Anaconda3\lib\site-packages\thinc\config.py", line 14, in <module>
    import srsly
  File "C:\ProgramData\Anaconda3\lib\site-packages\srsly\__init__.py", line 4, in <module>
    from ._msgpack_api import read_msgpack, write_msgpack, msgpack_dumps, msgpack_loads
  File "C:\ProgramData\Anaconda3\lib\site-packages\srsly\_msgpack_api.py", line 3, in <module>
    from . import msgpack
  File "C:\ProgramData\Anaconda3\lib\site-packages\srsly\msgpack\__init__.py", line 13, in <module>
    from ._msgpack_numpy import encode_numpy as _encode_numpy
  File "C:\ProgramData\Anaconda3\lib\site-packages\srsly\msgpack\_msgpack_numpy.py", line 19, in <module>
    import cupy
  File "C:\ProgramData\Anaconda3\lib\site-packages\cupy\__init__.py", line 46, in <module>
    import cupyx as _cupyx  # NOQA
  File "C:\ProgramData\Anaconda3\lib\site-packages\cupyx\__init__.py", line 8, in <module>
    from cupyx import linalg  # NOQA
  File "C:\ProgramData\Anaconda3\lib\site-packages\cupyx\linalg\__init__.py", line 2, in <module>
    from cupyx.linalg import sparse  # NOQA
  File "C:\ProgramData\Anaconda3\lib\site-packages\cupyx\linalg\sparse\__init__.py", line 3, in <module>
    from cupyx.linalg.sparse._solve import lschol  # NOQA
  File "C:\ProgramData\Anaconda3\lib\site-packages\cupyx\linalg\sparse\_solve.py", line 7, in <module>
    from cupyx.scipy import sparse
  File "C:\ProgramData\Anaconda3\lib\site-packages\cupyx\scipy\sparse\__init__.py", line 4, in <module>
    from cupyx.scipy.sparse.coo import coo_matrix  # NOQA
  File "C:\ProgramData\Anaconda3\lib\site-packages\cupyx\scipy\sparse\coo.py", line 13, in <module>
    from cupyx.scipy.sparse import csr
  File "C:\ProgramData\Anaconda3\lib\site-packages\cupyx\scipy\sparse\csr.py", line 589, in <module>
    @cupy._util.memoize(for_each_device=True)
AttributeError: partially initialized module 'cupy' has no attribute '_util' (most likely due to a circular import)

while this works:

>>> import torch
>>> torch.cuda.is_available()
True
>>>
>>> import spacy
>>> spacy.require_gpu()
True
>>>
>>> from thinc.api import prefer_gpu
>>> is_gpu = prefer_gpu()
>>> is_gpu
True

So spacy only works if I import torch first? Why? In my previous installation all was working, pytorch, thinc, spacy, but I forgot the conda/pip sequence I used then before rebuilding the PC.

So question is what is the right sequence of installs for a Windows 10 PC with card is NVIDIA GeForce RTX 2060 and Cuda compilation tools, release 11.0, V11.0.221
Build cuda_11.0_bu.relgpu_drvr445TC445_37.28845127_0 ?

Thanks

Eurico

[euricocovas]

Futher to this: I uninstalled python from my windows 10 PC, by removing anaconda totally, then rebooted, then run Anaconda3-2020.11-Windows-x86_64.exe, then did

conda activate

conda install -c conda-forge spacy --yes 
conda install -c conda-forge cupy --yes  
conda update -c conda-forge --all --yes

the update did this (among many other things):

scipy pkgs/main::scipy-1.5.2-py38h14eb087_0 --> conda-forge::scipy-1.6.2-py38he847743_0

then

import spacy
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "C:\ProgramData\Anaconda3\lib\site-packages\spacy\__init__.py", line 10, in <module>
    from thinc.api import prefer_gpu, require_gpu, require_cpu  # noqa: F401
  File "C:\ProgramData\Anaconda3\lib\site-packages\thinc\__init__.py", line 5, in <module>
    from .config import registry
  File "C:\ProgramData\Anaconda3\lib\site-packages\thinc\config.py", line 14, in <module>
    import srsly
  File "C:\ProgramData\Anaconda3\lib\site-packages\srsly\__init__.py", line 4, in <module>
    from ._msgpack_api import read_msgpack, write_msgpack, msgpack_dumps, msgpack_loads
  File "C:\ProgramData\Anaconda3\lib\site-packages\srsly\_msgpack_api.py", line 3, in <module>
    from . import msgpack
  File "C:\ProgramData\Anaconda3\lib\site-packages\srsly\msgpack\__init__.py", line 13, in <module>
    from ._msgpack_numpy import encode_numpy as _encode_numpy
  File "C:\ProgramData\Anaconda3\lib\site-packages\srsly\msgpack\_msgpack_numpy.py", line 19, in <module>
    import cupy
  File "C:\ProgramData\Anaconda3\lib\site-packages\cupy\__init__.py", line 46, in <module>
    import cupyx as _cupyx  # NOQA
  File "C:\ProgramData\Anaconda3\lib\site-packages\cupyx\__init__.py", line 3, in <module>
    from cupyx._runtime import get_runtime_info  # NOQA
  File "C:\ProgramData\Anaconda3\lib\site-packages\cupyx\_runtime.py", line 37, in <module>
    import scipy
  File "C:\ProgramData\Anaconda3\lib\site-packages\scipy\__init__.py", line 130, in <module>
    from . import _distributor_init
  File "C:\ProgramData\Anaconda3\lib\site-packages\scipy\_distributor_init.py", line 59, in <module>
    WinDLL(os.path.abspath(filename))
  File "C:\ProgramData\Anaconda3\lib\ctypes\__init__.py", line 373, in __init__
    self._handle = _dlopen(self._name, mode)
FileNotFoundError: Could not find module 'C:\ProgramData\Anaconda3\lib\site-packages\scipy\.libs\libbanded5x.3OIBJ6VWWPY6GDLEMSTXSIPCHHWASXGT.gfortran-win_amd64.dll' (or one of its dependencies). Try using the full path with constructor syntax.

fixed by
pip install scipy --force-reinstall

the

conda install scipy --force-reinstall

did not fixed it.

[euricocovas]

Also

pip install spacy-transformers 
pip install spacy-lookups-data 

installs pytorch but the CPU version I think and does not allow spacy to access GPU anymore, as spacy seems to stop using thinc and start using pytorch as backend?

So this combination does not work for GPU:

	conda activate

	conda install -c conda-forge spacy --yes 
	conda install -c conda-forge cupy --yes  
	
	pip install spacy-transformers 
	pip install spacy-lookups-data 
	
	python -m spacy download en_core_web_sm
	python -m spacy download en_core_web_md
	python -m spacy download en_core_web_lg
	python -m spacy download en_core_web_trf

[polm]

Sorry you're having trouble with this. In general versioning around CUDA is unfortunately pretty brittle everywhere in Python, especially on Windows.

We've had reports of the order of imports of torch and spaCy making a difference (like here), but we don't completely understand what's going on and haven't been able to reproduce it. It may be that there are multiple versions of some DLL on your system and the import order affects which one is loaded.

In general, I would recommend avoiding mixing pip and conda if at all possible. There are some rare cases where it can be a workaround for a nasty version issue, but it's the source of a lot of headaches.

Maybe you could try using a non-Anaconda Python version?

[euricocovas]

Thanks. Yes, mixing conda and pip is not good, but I think spacy-transformers and spacy-lookups-data doesn't exist in conda yet. I may try to do the following whenever I redo my installations of python from fresh, forcing pytorch cuda11.2 before I actually install the rest:

    conda activate

conda install -c conda-forge spacy --yes 
conda install -c conda-forge cupy --yes  

pip install --pre torch torchvision torchaudio -f https://download.pytorch.org/whl/nightly/cu112/torch_nightly.html --force-reinstall

pip install spacy-transformers 
pip install spacy-lookups-data 

python -m spacy download en_core_web_sm
python -m spacy download en_core_web_md
python -m spacy download en_core_web_lg
python -m spacy download en_core_web_trf