Version incompatibility between Spacy, Cuda, Pytorch and Python

[Project-NSE]

I want run spacy in GPu. The configuration that I installed for the Spacy is defined below
Name: spacy
Version: 3.7.4

The Cuda configuration that have in my Ubuntu 20.04.1 LTS based machine is

nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2019 NVIDIA Corporation
Built on Sun_Jul_28_19:07:16_PDT_2019
Cuda compilation tools, release 10.1, V10.1.243

I have a restriction to upgrade the Cuda. The PyTorch version that I have installed on the machine is "2.2.1+cu121" which have no support for the Cuda version I have.

I tried to downgrade the Pytorch and found that I need to downgrade the Python version as well.

My current python version is Python 3.12.2. If I go for a compatible Python version for pytorch that should be >=3.6,<3.7.0 for pytorch 1.4.0.

If I downgrade both Pytorch and Python that is not compatible with Spacy version.

I need to do a transformer-based NER task in Spacy. Therefore I am not sure what is the minimum requirement of spacy.

How can I handle this version incompatibility?

[svlandeg]

Hi!

So to summarize, you said that:

• You can't upgrade your Cuda and you have 10.1
• You need to downgrade to Python 3.6.x and PyTorch 1.4.0 for this Cuda version

Right?

To get a version of spaCy that is compatible with a specific PyTorch version, you'll have to look at Thinc - this is spaCy's Machine Learning library that is in charge of connecting to PyTorch functionality. The current version 8.2.3 uses torch>=1.6.0 so that won't work for you in combination with PyTorch 1.4.0.

However, it looks like torch 1.7.0 should also be compatible with Cuda 10.1, so I suggest you try that.