#cudnn — Public Fediverse posts

Converting CUDA to ROCm wasn't easy, but relatively straightforward due to how AMD created their API. It was a 1:1 replacement.

But converting CUDNN to MIOpen is a pain. Some APIs switch the order in which parameters are passed, or take a different number of parameters. And some APIs (e.g., filters) don't even exist so I'll need to understand what this old code is doing before I can attempt to fix what is going on. #darknet #yolo #cudnn #miopen #cpp

Wenn man alte #GPUs hat, in meinem Falle NVIDIA Tesla K40, also #Kepler-Architektur, kompiliert man sich dusselig. Man muss einen alten Treiber benutzen, hat eine nicht mehr unterstütze compute capability, kein #CUDNN und muss die Libraries, primär pytorch, immer für exakt die #Python-Version kompiliert haben, die das fragliche Environment benötigt. Also verlangt jedes Projekt, das man zum Laufen bringen will, seine eigene, spezifisch kompilierte, Version von #pytorch.

I'm trying to install #CUDA and #cuDNN on #PopOS. #Tensorflow only seems to work with CUDA 12.2 and cuDNN 8.8. As these are older, archived releases, is there a better way to install the drivers without interfering with existing #Nvidia software? Would using conda help?
I currently use Poetry and am not sure if adding conda to the mix would improve the situation.

This man has suffered enough.. poor guy..😔

#nvidia #cudnn #tensorflow #fun #humor