instal the last version for apple Bite the Bullet

If you encounter problems, you can always revert back by running git checkout webui-user.sh and then you can delete the venv-torch-nightly folder. Keep in mind that the nightly PyTorch builds may have issues, especially since they are updated every day. webui.sh the web UI dependencies will be reinstalled, along with the latest nightly build of PyTorch. To use all of these new improvements, you don't need to do much just unzip this webui-user.sh file and replace the webui-user.sh file in stable-diffusion-webui. The latest nightly builds get roughly 25% better performance than 1.12.1 (the current default). It manages memory far better than any of the other cross attention optimizations available to Macs and is required for large image sizes. This is the recommended cross attention optimization to use with newer PyTorch versions. Add the command line argument -opt-sub-quad-attention to use this. Note that if you want to train embeddings or hypernetworks, you should start web UI with -no-half (e.g./webui.sh -no-half you don't have to fully override the default command lines arguments to remove -upcast-sampling as -no-half overrides it).

In that case you'll need to remove -no-half and make sure -upcast-sampling is used instead. This has been the default for macOS since e0df864 so other than upgrading web UI, no action is needed - unless you've edited webui-user.sh to override the default command line arguments and included -no-half. This significantly lowers memory usage and improves performance. Half precision support (using web UI without -no-half) with -upcast-sampling.There have been several additions and changes made recently that can improve performance on macOS: Note that these currently are not intended to be considered official builds and they will not necessary reflect exactly what this main repository has at any given point in time.Īlso, Intel versions are planned in the near future when I can track down exactly what is causing the reported PyTorch 2.0 issues with k-diffusion and UniPC samplers. The 'experimental' builds will usually have the latest features and optimizations, and other builds will be intended to provide a stable version that may include new features once they've been tested sufficiently. If you use an Apple silicon Mac and want to try the latest features and optimizations, look here:

Instal the last version for apple Bite the Bullet