Repository navigation
Replies: 2 comments 4 replies
|
Good question regarding |
|
I think the That matters for Thor since newer PyTorch releases have the newer CUDA 13 builds, and there are already Thor setups running PyTorch with CUDA 13.x successfully. I'd check what actually got installed with: import torch
print(torch.__version__)
print(torch.version.cuda)
print(torch.cuda.is_available())
print(torch.cuda.get_device_name(0) if torch.cuda.is_available() else "no cuda")If you're getting a 2.10.x PyTorch from the current package and the newer PyTorch you tried recognizes Thor, I'd try relaxing the Would be interesting to see what |
Uh oh!
There was an error while loading. Please reload this page.
Curious if the samples won't work on Thor using GPU. Cuda version I have is 13.2, but it seems like the reasoning-from-scratch package is forcing a torch downgrade (from latest), which in turn makes torch not support my gpu - it thunks down to cpu. Fwiw the latest torch recognizes 130 and not 132, which might be good enough.
I could in the meantime just copy the sources into a working directory and continue from there - but more curious if there is a fundamental challenge with the sample code and the cuda version, which is forcing the torch pin to < 2.11 per
reasoning-from-scratch/pyproject.toml
Line 27 in 0acaa28
thanks
All reactions