[Feature] On-the-fly LoRA Conversion is Slow

[tkdtkdwx]

我使用的腾讯云上的L20服务器，nunchaku版本是0.3.0，cuda12.8 pytorch2.7.0 python 3.11.12, ComfyUI版本是0.3.40，[ComfyUI-nunchaku]插件版本也是0.3.0，现在的问题是，我正常使用nunchaku跑FLUX文生图，fill抹除重绘，readux这些都正常，就是无法正常使用lora，加载lora运行时，后台会自动开始进行转换，运行要等好久，虽然转换后可以起效果，但是如果我把Lora加载节点忽略重开，或者替换别的lora再换回来，也需要重新转换，

[lmxyy]

You can use https://github.com/mit-han-lab/nunchaku/blob/main/nunchaku/lora/flux/compose.py to convert the LoRA and dump it. Also, to load nunchaku LoRA in ComfyUI, you need to use #446

[lzj1214]

You can use https://github.com/mit-han-lab/nunchaku/blob/main/nunchaku/lora/flux/compose.py to convert the LoRA and dump it. Also, to load nunchaku LoRA in ComfyUI, you need to use #446

#446 only support single lora not multiple nunchaku format loras

[lmxyy]

Use https://github.com/mit-han-lab/nunchaku/blob/main/nunchaku/lora/flux/compose.py to fuse multiple LoRA into a single nunchaku one, first.

This strategy is only used if you do not want on-the-fly conversion.

[github-actions]

This issue has been automatically closed due to 30-day inactivity. Please feel free to reopen it with /reopen if needed.

[xing-shuai]

@lmxyy Hi, lora file saved by function to_nunchaku[ import from lora.flux.nunchaku_converter] is so big, almost double size of the original size. I'm using nunchaku 0.3.2_dev0. Any way to reduce the size? Appreciate! by the way, I'm using kontext model in my code.

[lmxyy]

@xing-shuai What is the rank of your LoRA? Nunchaku's LoRA needs to concatenate our low-rank branch, and sometimes QKV's LoRA are duplicated.

[xing-shuai]

[@xing-shuai]（https://github.com/xing-shuai） 你的 LoRA 的等级是多少？双节双节的 LoRA 需要连接我们的低秩分支，有时 QKV 的 LoRA 是重复的。

32 or 16

[lmxyy]

Duplicated. Merged into nunchaku-ai/ComfyUI-nunchaku#383