Dear all.
I am trying to run yolov ultralytics package with GPU for training pattern recognition of images that uses tensorflow. This is for ATLAS phase-II upgrade of ITk-Pixels
I did this already in the past. But apparently there were been changes in the packages and GPUs as cuda for the Jupyter Notebooks are now in version 12.4.
I basically went back and forward with compatibilities will all packages and got this jupyter command:
!python -m pip install pip idna==2.10 psutil==5.9.8 urllib3==1.26.20 swanportallocator 2to3 traitlets==5.6.0 idna==2.10 protobuf==3.20.3 numpy==1.26.4 ultralytics torch==2.5.0 torchvision==0.20.0 torchaudio==2.5.0 --extra-index-url https://download.pytorch.org/whl/cu124 --upgrade --target=/eos/home-i03/m/mdacunha/.local/lib/python3.11/site-packages
This puts my eos .local at 5.6G (which is why I am targeting EOS and not home).
Is is possible to install ultralytics package in the GPU machines ?
we moved to another stack by default (LCG 106a) which has Python 3.11 and CUDA 12.3.
Since it runs a new Python version, this explains why you needed to reinstall packages (it’s also mentioned in the OTG announcement).
To have ultralytics installed in an LCG release, you can open a ticket to the EP-SFT librarians that maintain the LCG stacks (which we make available in SWAN). Their JIRA is:
You can mention in that ticket that you are a SWAN user, you can explain your use case and request any package you need. Please add me as watcher of the ticket so I can follow up the progress.
Dear Enric,
I saw another case with such a request and I have created already yesterday a ticket: https://its.cern.ch/jira/browse/SPI-2697
Meanwhile (after a day lost) I realised I could just skip SWAN altogether and run directly in lxplus.
If I understand it correctly the problem of not be able to run with SWAN GPU was the version of cuda:
The 12.4 is the supported version of the driver, that can be above the CUDA version of the toolkit (12.3), so that is fine.
As for your use of SWAN, please note you can keep using LCG 105a CUDA (the software stack you used to use). It’s in the list, when you select the software stack. That will keep things as they were before the change of default release.
