3.5x speed upgrade with Nvidia .dll replacement working for InvokeAI Standalone 3.2.0 and a 2x speed increase for standalone 2.3.2 and the regular install of 2.3.4.
I've replaced the CUDA files for invokeAI 2.3.0 and it makes it 3.5x faster with my Nvidia 4090. I'm running InvokeAI 2.3.0 at 28 it/s just by replacing the CUDA .dll files alone. However, I cant seem to get the latest to be as fast by replacing the .dll files but at least it's still 2x faster on InvokeAI 2.3.2 and also the normal install of the latest 2.3.4 with .dll replacement. Not sure why I cant get it going as fast as 2.3.0 at 3.5x the speed. For in-painting I'm still using Invoke 2.3.0 with the upgraded .dll's and loving the 3.5x speed. Any idea why the versions after 2.3.0 are not performing as fast as InvokeAI 2.3.0 with the latest CUDA files? anyone know how to get more speed out of Invoke 2.3.2-2.3.4?
I Downloaded the Nvidia zip file here and used the .dll files found in the bin folder:
https://developer.download.nvidia.com/compute/redist/cudnn/v8.7.0/local_installers/11.8/
Then I replace the old .dll files with the new ones into the following: InvokeAI\venv\Lib\site-packages\torch\lib