Hello, I noticed that Pytorch 1.9 still uses cuDNN 8.0.5, same as 1.8 even though cuDNN 8.2.1 has been released. Is there a reason for this?
I read in the cuDNN release notes that there were some performance regressions for some legacy APIs since 8.0.5.
Before updating a cudnn version (or any other 3rd party library) we are usually running it for a couple of weeks in the nightlies to get visibility into potential issues.
In the 1.9.0 release we were working on fixing the library split issue, the missing cutlass/xmma kernels etc.
While cudnn8.2.1 is already released, the risk was too high to fix the outstanding bugs and update cudnn without the visibility of the nightlies, which is why it’s delayed.
Gotcha, thanks for the update! I’ll keep my eye on the nightlies