![]() Sdp - scaled dot product - Looks like the best option for me, but very RAM-hungry. And it's practically impossible to run post 1.3.0 without it. Sub-quadratic - Our go-to choice in the previous version, but unfortunately DOESN'T WORK with token merging. Safe optionĭoggettX - Essentially the split-attention as we know it. ![]() V1 - Original v1 - The least memory-hungry version of the standard split-attention. arguments into the webui-user.bat, and instead you choose them from the dropdown list right here. ![]() Essentially, right now it's unnecessary to add any of -opt. ![]() Optimizationsīy all means read and take into the consideration everything written below, but the update 1.3.0 adds an important tab into the options: "Optimizations". So I should probably add the part 0 for the new stuff in 1.3.0 to be aware of. Prefaceīefore I even start, I think I should make it clear that the guide has been written before the large 1.3.0 update that considerably changed a lot going on under the hood. ![]() The article will get expanded as I amass more knowledge. Nevertheless, this post has been made from the perspective of AMD RX 580 (8GB) owner. Considering that DirectML implementation is more of a translation layer rather than a low-level rewrite of the original code, some features of the original SD webui are bound to not function properly, and different AMD cards may also need a different approaches. After about 2 months of being a SD DirectML power user and an active person in the discussions here I finally made my mind to compile the knowledge I've gathered after all that time. ![]()
0 Comments
Leave a Reply. |
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |