Ryan Dick

InvokeAI already supports: 1. Multiple IP-Adapters (in both Linear UI and Workflows). 2. Multiple images to an IP-Adapter (Workflows only). These two features are _not_ equivalent. @suonnon It sounds like...

> Which `load_state_dict()` function are you referring to here? This one: > To understand the problem better, I recommend setting a breakpoint here: https://github.com/huggingface/peft/blob/26726bf1ddee6ca75ed4e1bfd292094526707a78/src/peft/utils/save_and_load.py#L249

> @yiyixuxu as per understanding, the issue has roots in `peft`. Hence I cc'd @BenjaminBossan. I think the main issue is the kohya key conversion in diffusers - it produces...

> > I think the main issue is the kohya key conversion in diffusers - it produces keys that do not exist. > > Do not exist where? Could you...

I just did some testing of this branch. My focus was on testing that this works on an 8GB GPU, not on exhaustive feature testing, but I noticed a few...

> Was that yoda image done on an olive model per chance? Olive models don't support applying loras I believe. If it's just a standard onnx model, could you send...

I tried to reproduce this today. It turns out that there was no regression in VAE tiling behavior. There was a period of time during the switch from `tiled_decode` to...

> @RyanJDick Would you mind doing one last review of this PR? Looks like 43/44 files have changed since I last looked 😅 . I'll plan to spend a chunk...

I've totally lost track of all the model probing pathways at this point, but I see that `learned_embeds` and `pytorch_lora_weights` are still hard-coded in several places. Do those not conflict...

Patching (for LoRA or TI) is managed using context managers (applied on entry, and reversed on exit). Examples: - https://github.com/invoke-ai/InvokeAI/blob/e22248466364f6895ed55f4dacda482641a4af58/invokeai/app/invocations/compel.py#L179-L189 - https://github.com/invoke-ai/InvokeAI/blob/e22248466364f6895ed55f4dacda482641a4af58/invokeai/app/invocations/latent.py#L931-L938 Now that the model cache has the power...