convert_lora.py fails on Lycoris LoKr

[Enyakk]

The inclusion of Lycoris support via the following PRs worked great and greatly expands the possible training options. Lycoris LoKr also works perfectly when training, sampling and generating videos via musubi repository.

• add lycoris support #33
• Update lycoris for hv_generate_video.py #56

However convert_lora.py does not recognize the lokr-specific keys produced. This results in an empty .safetensors when trying to convert a training LoKr .safetensors.

This produces warnings as such:
WARNING:__main__:unexpected key: lora_unet_double_blocks_0_img_attn_proj.lokr_w1 in default LoRA format WARNING:__main__:unexpected key: lora_unet_double_blocks_0_img_attn_proj.lokr_w2 in default LoRA format WARNING:__main__:unexpected key: lora_unet_double_blocks_0_img_attn_qkv.lokr_w1 in default LoRA format WARNING:__main__:unexpected key: lora_unet_double_blocks_0_img_attn_qkv.lokr_w2 in default LoRA format WARNING:__main__:unexpected key: lora_unet_double_blocks_0_img_mlp_fc1.lokr_w1 in default LoRA format WARNING:__main__:unexpected key: lora_unet_double_blocks_0_img_mlp_fc1.lokr_w2 in default LoRA format WARNING:__main__:unexpected key: lora_unet_double_blocks_0_img_mlp_fc2.lokr_w1 in default LoRA format WARNING:__main__:unexpected key: lora_unet_double_blocks_0_img_mlp_fc2.lokr_w2 in default LoRA format

To train Lycoris LoKr the following lines have been used:

network_module="lycoris.kohya"
network_dim = 100000
network_alpha = 1
network_args = [ "preset=attn-mlp", "algo=lokr", "factor=140", ]

Attachements:

• convert_lora.py log
• training config + minimal LokR safetensors (ZIP)

convert_lora.log
test_lokr.zip

[sdbds]

To be honest, I think ComfYUI should use the default Lycoris Key, because this is the official format of Lycoris.

[Enyakk]

Well, I could raise the issue on the ComfyUI github if that makes sense.