Wan 2.2 LoRAs - type i32 unsupported

[alex-ie]

I have tried to run per docs https://github.com/leejet/stable-diffusion.cpp/blob/master/docs/wan.md, downloaded model from the link in docs https://huggingface.co/QuantStack/Wan2.2-T2V-A14B-GGUF/tree/main and it works on its own (command without LoRAs).

Docs also mention:

Wan2.2 T2V 14B with Lora
-p "a lovely cat<lora:wan2.2_t2v_lightx2v_4steps_lora_v1.1_low_noise:1><lora:|high_noise|wan2.2_t2v_lightx2v_4steps_lora_v1.1_high_noise:1>"

but do not give links for those LoRAs. I have tried to use : https://huggingface.co/Comfy-Org/Wan-2.2/resolve/main/split_files/loras/wan2.2_t2v_lightx2v_4steps_lora_v1.1_high_noise.safetensors and got in terminal:

what(): type i32 unsupported for integer quantization:

Which LoRAs to use? TIA

P.S. interestingly koboldcpp --analyze wan2.2_t2v_lightx2v_4steps_lora_v1.1_high_noise.safetensors shows tensors of types F32 and I64 only.

[Medamineelkhattabi]

The Comfy-Org files are the right ones. They're byte-identical (same SHA-256 on HF) to lightx2v's Seko-V1.1 files, and they're the names the docs example uses. The crash is a loader bug that's already fixed, so update sd.cpp.

The 400 I64 tensors are the per-layer .alpha scalars (value 8, rank 64). The loader narrows I64 to I32 and loads LoRA alphas as F32, but older builds had no I32 to F32 case in convert_tensor, so it fell into the dequantize branch and threw that message. That's #2261, fixed by #2263, first shipped in master-1043-3be2378 (current code). sd-cli --version shows which build you have.

Put both files in one folder (the extension is optional):

--lora-model-dir ..\models\loras -p "a lovely cat<lora:wan2.2_t2v_lightx2v_4steps_lora_v1.1_low_noise:1><lora:|high_noise|wan2.2_t2v_lightx2v_4steps_lora_v1.1_high_noise:1>"

|high_noise| routes that LoRA to --high-noise-diffusion-model. Without --lora-model-dir, sd-cli doesn't load <lora:...> tags at all.

If you can't update (or you run it through koboldcpp, whose bundled copy in 1.124.2 doesn't have this fix), convert the alphas to F32 instead of deleting them. With no alpha the scale falls back to 1.0 instead of 8/64, so the LoRA would be 8x too strong:

import sys
from safetensors.numpy import load_file, save_file
t = load_file(sys.argv[1])
save_file({k: v.astype("float32") if v.dtype.kind in "iu" else v for k, v in t.items()}, sys.argv[2])

Run it on both files: python fix.py in.safetensors out.safetensors.

[alex-ie]

Thanks! Though I am curios:

The crash is a loader bug that's already fixed, so update sd.cpp.

But fix only a week ago whereas Wan 2.2 and even those LoRA files are a year old. Why bug fix only very recently? Can you please explain? Either few use Wan 2.2 or the bug got introduced recently or else.

[wbruna]

Why bug fix only very recently?

Note how old the bug was when it got fixed.

I'd say it's because people who contribute to ComfyUI usually aren't terribly interested in contributing to stable-diffusion.cpp; and they are also used to code customization, so you have a lot of random format variations out there.

[alex-ie]

@wbruna , I do not understand your comment "Note how old the bug was". #2261 says "open 1w ago" - that is one week ago, correct?

[wbruna]

Yes. One reason to have been fixed only recently is that it was only opened recently.