You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Torch 1.13 deprecated old CUDA drivers, see release notes here. And we are currently on CUDA 10.2, at UZH.
Poetry installs the version which is stored in the lock file, which is 1.12.1.
Pip installs the latest possible, which is 1.13.1. This is why it then fails in running on the GPU.
Current solution: we pin fractal-tasks-core to torch 1.12.1.
Long-term solution: we use a more recent CUDA (note that it may interfere with other software, like TissueMaps).
Let's discuss updates for our dependencies or for CUDA somewhere else (since it also depends on the FMI setup).
Let's discuss what we'll need to support in different places. For the moment, the 1.12.1 solution seems reasonable :)
The text was updated successfully, but these errors were encountered:
See comment here: #215 (comment)
Let's discuss what we'll need to support in different places. For the moment, the 1.12.1 solution seems reasonable :)
The text was updated successfully, but these errors were encountered: