The current release of Haystack (deepset.ai) supports GPU's but it is not using a new enough CUDA support (>= 11.1) to work with my RTX3090. Can I just uninstall the current version of torch and reinstall with cu111 support? Or are other components also tied to an earlier version of the CUDA library?
I have torch 1.9.0 installed with CUDA 11.2 installed and it works fine with the master branches of FARM and haystack. Let me walk you through the steps how to get there.
The problem is that the latest release of haystack 0.9.0 has FARM 0.8.0 as dependency, which fixes the torch dependency to <1.9. However, if you install haystack from its master branch you could remove the fixed FARM 0.8.0 dependency and install FARM from its master branch as well. On the current master branch of FARM, the torch dependency is relaxed to <1.10.
I have in mind the following steps to first install FARM and then haystack from their master branches:
git clone https://github.com/deepset-ai/FARM.git
cd FARM
pip install -r requirements.txt
pip install --editable .
cd ..
git clone https://github.com/deepset-ai/haystack.git
cd haystack
#now edit requirements.txt and remove line with farm==0.8.0
pip install --editable .
Note that all of this won’t be necessary when the 1.0.0 version of haystack is released, which will be within the next 3 weeks in late September 2021.
Maybe there will also be a minor 0.9.1 even before that, which should solve your problem so that you can simply use pip install farm-haystack.