FP32 causes 8GB vram gpu out of memory error, RunDiffusion/Juggernaut-X-v10 FP16 is not supported

[xhoxye]

# SDXL
sdxl_name = RunDiffusion/Juggernaut-X-v10 #FP16 is not supported
# sdxl_name = SG161222/RealVisXL_V4.0
# sdxl_name = stabilityai/stable-diffusion-xl-base-1.0

FP32 can be downloaded, however it causes 8GB vram gpu out of memory error

[Duemellon]

I see the syntax you used to change models but I'm not sure how to apply that to my own. Of course all my models are located in the A1111 folder somewhere else. What is the syntax? Is there a problem concern with the file extension?

[xhoxye]

@Duemellon
It's not a .safetensors file name, it's the huggingface directory name
https://huggingface.co/SG161222/RealVisXL_V4.0

[Duemellon]

So you can't point this to a local download of a file at this time? It can only do a download after you point to the HF link?

[xhoxye]

If you want to load the local .safetensors file, you will need to modify more code

[Manul07]

no undesend

[Duemellon]

If you want to load the local .safetensors file, you will need to modify more code

yes, that's what would be needed. I have a terabyte or more of locally installed safetensor fiels that I could use instead of pointing to a a HF directory. I'm not familiar with PY at all to make these changes but that is a different topic than this thread at this point.

[xhoxye]

models_dir = os.path.join(os.getcwd(), 'models/checkpoints')

sdxl_name = 'RealVisXL_V4.0'

model_path = os.path.join(models_dir, sdxl_name + '.safetensors')

if os.path.isfile(model_path):
    temp_pipeline = StableDiffusionXLImg2ImgPipeline.from_single_file(model_path)
    
    tokenizer = temp_pipeline.tokenizer
    tokenizer_2 = temp_pipeline.tokenizer_2
    text_encoder = temp_pipeline.text_encoder
    text_encoder_2 = temp_pipeline.text_encoder_2
    text_encoder_2 = CLIPTextModel(config=text_encoder_2.config)
    vae = temp_pipeline.vae
    unet = temp_pipeline.unet
else:
    raise FileNotFoundError(f"Model file {model_path} not found.")  

[xhoxye]

[xhoxye]

I'll submit a PR later

[xhoxye]

"torch_dtype=torch.float32" It run incorrectly

[xhoxye]

#81