Regarding this issue, how can I upgrade or replace the cuDNN version built into my current PyTorch installation?

1

Significant Memory Regression in F.conv3d with bfloat16 Inputs in PyTorch 2.9.0 (#166643) This release provides work around this issue. If you are impacted please install nvidia-cudnn package version 9.15+ from pypi. (#166480) (#167111) .

2

After installing the PyTorch binary you can update cuDNN via pip install nvidia-cudnn-cu12==VERSION where VERSION corresponds to the release you want to update to. If you are using our PyTorch binaries built with CUDA 13 use nvidia-cudnn-cu13 instead.