mit-han-lab

/

svdquant-models

like 69

Follow

MIT HAN Lab 153

We are still cleaning the converting script and will release the guide soon. Do you need the Hyper-SD LoRA? If you need the Hyper-SD LoRA, we can first convert that for you first.

Hi, first, thanks for your effort. Now I'm trying to create the quantized svdq int4 model from a custom flux transformer which hyper SD LoRA merged using comand below:

python -m deepcompressor.app.diffusion.ptq configs/model/flux.1-custom.yaml configs/svdquant/int4.yaml --save-model /root/autodl-tmp/flux.1-custom-svdquant-int4

https://github.com/mit-han-lab/deepcompressor/issues/24

I'm wondering how long it would take to finish the job. I'm using an H800 to do the job. @Lmxyy

We are still cleaning the converting script and will release the guide soon. Do you need the Hyper-SD LoRA? If you need the Hyper-SD LoRA, we can first convert that for you first.

Please do provide, it will be allow High quality images in low steps with speeds.

HyperSD和Alimama的Turbo FLUX Alpha在FLUX加速上的表现都非常好，现在我已经在FP16格式上从HyperSD转为使用Turbo Flux了，因为横纹和噪音都少了很多。

We are still cleaning the converting script and will release the guide soon. Do you need the Hyper-SD LoRA? If you need the Hyper-SD LoRA, we can first convert that for you first.

I also look forward to the early release of the LoRA conversion script.

@Lmxyy I'm here to suggest the same thing. Combining SVDQuant with Hyper-SD would be powerful.

@bvrv I just curious about how to combine svdquant and hyper-sd，I have convert both lora into "nunchaku" format！did you successful combine those？

@daniel19950522 it doesn’t seem to be working well yet: https://github.com/mit-han-lab/nunchaku/issues/108

Hi @bvrv , I have checked this LoRA and found that it also has LoRA on non-quantized layers. I will fix this corner case this weekend. Other LoRA should work well.

This issue is caused by the large LoRA's rank. You can change this line to

using LoraRanks = std::integer_sequence<int, 0, 32, 48, 64, 80, 96, 128, 160, 192, 224, 256, 288, 320, 384, 448, 512>;

and rebuild the package. The compilation may take a while. I've already fixed this issue in v0.1.4, which will release on Thursday or Friday. I will let you know then.