Safetensors?

[KintCark]

Can we use safetensor models on this or only diffusers ?

[vitoplantamura]

hi,

HF Diffusers support loading safetensor files, so I don't think I quite understand your question :-)

More info here: https://github.com/vitoplantamura/OnnxStream#how-to-convert-and-run-a-custom-stable-diffusion-15-model-with-onnxstream-by-gaelicthunder

Vito

[KintCark]

I'm saying like if I wanted to use safetensor models from civitai would otherwise work I installed onnxstream on android but when I looked in the sdxl-turbo folder it's not a safetensor but a Diffuser file like it has the folders like text_encoder vae_decoder tokenizer etc..... but if I put an sdxl safetensor file which is one single file will it work?

[vitoplantamura]

no, the safetensor file needs to be first converted to the format compatible with OnnxStream.

This is not a simple procedure. The link I posted before explains how to do it for SD1.5. The procedure for SDXL is similar but it is still not trivial.

Vito

[noah003]

no, the safetensor file needs to be first converted to the format compatible with OnnxStream. This is not a simple procedure. The link I posted before explains how to do it for SD1.5. The procedure for SDXL is similar but it is still not trivial. Vito

It's too difficult to convert the .safetensors model...

[AeroX2]

I've taken a quick stab at this and made this: https://github.com/AeroX2/safetensor2onnx2txt

It currently only supports SDXL Turbo models but could probably be easily modified to support other models as a bonus it should also support Lora's although your milage may vary as I've only tested it one SDXL Turbo Lora

And it still needs a bunch more debugging as some models do not work at all and I have no idea why and I still need to support the tiling mode of OnnxStream (Actually looks like this works with just the unet conversion, no need to convert the vae)

[vitoplantamura]

really interesting!

I took a look at the code: the idea of ​​a single project that manages the entire conversion process is excellent! I plan to try it ASAP!

Thanks, Vito

[vitoplantamura]

hi AeroX2,

I finally found the time to try your project calmly and it works flawlessly.

I tried with a LoRA model for SDXL, modifying the code a bit (such as the size of the latents input to the UNET model).

It took 70 minutes on an Ubuntu VM on my 2018 laptop, with 16GB RAM + 16GB swap (16GB RAM wasn't enough, the simplification of the model was killed). It would probably take a lot less time if I reserved 32GB of RAM directly.

I hope to find some time in the future to fork the project and implement all cases, which, as you also wrote, should be a simple thing.

Thank you, Vito