cuDNN only selects sm80 kernels on blackwell devices

AI & Data Science Deep Learning (Training & Inference) cuDNN
1

Cross-linking this xla issue, since this is probably a cuDNN bug. The issue seems to be that both PyTorch and Jax/XLA (and thus probably cuDNN) only select sm80 conv kernels on a 5060ti even though sm100 kernels should be available.

4

Hi @cbtcbt2 ,
I am checking on this with the Engineering, and will share the update.

Thanks

5

Any updates on this? I encountered the same problem when executing conv3d on 5090 using pytorch API.