sdxl-turbo colab inside + questions

[eyaler]

I was able to recreate nunchaku-tech/nunchaku-sdxl-turbo/svdq-int4_r32-sdxl-turbo.safetensors using https://github.com/devgdovg/deepcompressor/tree/feat/sdxl
in this colab: https://colab.research.google.com/drive/17GA-Y2aNTjWU_37LKy68L7TI8jccyb1_?usp=sharing

It takes a couple of hours on an A100 (skipping eval).

Questions:

1. In my use case inference uses batch_size=1 and num_inference_steps=1 (with a custom sigma for img2img) . Should I change any of these before optimization?
2. I am currently trying to do sd-turbo - would appreciate any help!

[otherV]

@eyaler Any success? I'm trying to do Illustrious XL v1.0 (Based on SDXL). Using sdxl.yaml from deepcompressor-v0.3.0 branch. deepcompressor.app.diffusion.ptq only gives config.yaml and run.log files, no model.pt is saved. Any idea what I'm doing wrong?

[eyaler]

@otherV The colab worked for me for sdxl-turbo. I suggest you start from the colab and the mentioned feat/sdxl branch used. the model.pt is saved in a deep sub-directory which the colab locates using bash find