Old LoRA format giving noisy results

[nivedwho]

So I have an old LoRA trained using similar code as this repo - lucataco/cog-sdxl-lcm. But when I load the lora.safetensors file using the latest diffusers version with peft, the results I get are noisy. Results are fine if I'm on older diffusers version.

Can someone help me understand why this happens, and a potential solution to this?

[tolgacangoz]

Could you share a reproducible code snippet so that one can also try it directly?