Running LoRA in 24g vram but 30 seconds for single image

[Suprhimp]

Hi, Thanks for the cool models first.

I just want to ask about the vram memory spend durring generation.

I tested with a10 gpu 24 gb vram

but when I checked the memory, it uses only 20 gb of vram

and also the running time is quite slow for me should I run stage c, and stage b all the time?

In the example, generate 4 images spend only 26 seconds

Should I set some flag like --highvram or something? or what gpu should I use for the fast generate as example?

[wodsoe]

gdf sampler is very slow.

[SimonAndMilky]

how did you get it running in the first place? I was trying to train a Lora following the examples ( after setting the env variables for SLURMs all to 1, and fixed the - file error for webdataset), the code stuck at the progress bar after loading the models and images... I am using A100, and I can see around 18GiB of RAM is used with no 0% GPU usage.... are you following https://github.com/Stability-AI/StableCascade/blob/master/train/readme.md as well?

[Suprhimp]

I didn't start train yet, I just checked the output of the example :)

[wen020]

@SimonAndMilky +1，Have you solved this problem?

[wen020]

I didn't start train yet, I just checked the output of the example :)

Hi, I use code(https://github.com/Stability-AI/StableCascade/blob/master/inference/lora.ipynb) to generate images by lora, but the following error will be reported.

I understand that this is caused by the model not being downloaded correctly. The code I used to download the model is as follows. Is there any problem with this operation?

download models :bash download_models.sh essential small-small bfloat16

[Suprhimp]

@wen020 Hi , I just run download models all, not the only essential