In docker image nvcr.io/nvidia/l4t-pytorch:r35.1.0-pth1.13-py3, although the installed pytorch’s version is 1.13, when I import it and run torch.version, it shows ‘1.12.0a0+02fb0b0f.nv22.06’
Hi @shuyw, the actual PyTorch wheel that was installed into that container is from here: https://developer.download.nvidia.com/compute/redist/jp/v50/pytorch/torch-1.13.0a0+340c4120.nv22.06-cp38-cp38-linux_aarch64.whl
Those newer PyTorch wheels get built from PyTorch master (top-of-tree), and the version that’s reported from the code doesn’t always seem to match the git versioning (depending on when PyTorch releases new tags).
In pytorch 1.13.0+, there should be get_default_qconfig_mapping when I import from torch.ao.quantization:
from torch.ao.quantization import get_default_qconfig_mapping
However, I got a not found error.
By the way, it there a patch for torch1.13.0 when I try to build from source?
Sorry, I believed you emailed me about the patch, you would need to take one of my older patches and apply it by hand (which I normally do anyways) to 1.13. I’m not actually sure what (if any) patches are required for 1.13 since aarch64 support has improved in PyTorch master.
You might also want to try these PyTorch 1.14 wheels for JetPack 5.1: https://developer.download.nvidia.com/compute/redist/jp/v51/pytorch/
This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.