Lincoln Stein

> I understand the overall strategy, but I'm having trouble wrapping my head around the fix for models changing device. If I understand correctly, the solution is very simple -...

> Note: I tested this PR to see if it fixed #6375. It does not. Rats. I was rather hoping it would. I'm digging into the LoRA loading issue now.

> BTW - seamless working fine for me. Tested SD1.5 and SDXL, all permutations of axes. I wonder if there is some interaction with other settings you were using? It...

@psychedelicious @RyanJDick I have included a fix for #6375 in this PR. There was some old model cache code originally written by Stalker that traversed the garbage collector and forcibly...

The latest [commit](https://github.com/invoke-ai/InvokeAI/pull/6312/commits/5d4b747d588efe2ec16eb49b018ea69ddbe5d174) implements an optimization that circumvents the LoRA unpatching step when working with a model that is resident in CUDA VRAM. This works because the new scheme never...

> > Did you test the effect of removing the VRAM cache with a large VRAM cache size (e.g. large enough to hold all working models)? For this usage pattern,...

@RyanJDick I’ve undone the model patching changes and the removal of the VRAM cache, and what’s left is the original cpu->vram optimization, the fix to the TI patching, and the...

I'm going to merge this in and then will start working on further optimizations including the lora loading/unloading.

> Awesome. Thanks for splitting up the PRs. > > I did some quick manual regression testing - everything looked good. I tried: > > * Text-to-image, LoRA, TI >...

Ok, I'll work on a fix for this.