What? SDXL can run only less than 4GB VRAM!?

[ipinz]

using --lowvram sdxl can run with only 4GB VRAM, anyone?
Slow progress but still acceptable, estimated 80 secs to completed.

[EvianMcKeown]

Does 1024 x 1024 also work?

[ipinz]

works too. But for loading model at first time at least it needs 12GB free RAM (not VRAM)

[EvianMcKeown]

I'm really surprised at how little vram sdxl uses actually!

• 1024 x 1280:
  

[bjornlarssen]

Ouch. Sounds like bad news for my 8GB M2 Mac Mini… (haven't tried yet, it has an old version of A1111 running)

[AyYildizPilot]

EvianMcKeown How did you run it?, it didn't work even at 512x512 settings on my laptop.
i5-11400H
24GB of RAM
RTX3050 4GB VRAM

sdxl style is not loading, cuda and pytorch are giving error.

[MalfreCryvertia]

i cant, it stops loading the model for me in the first place, i got 32gb ram and 4gb nvidia1050 gpu/vram
sometimes it loads then i always get OOM even when generating 128x128 for a test

when it is not OOM, always the following error
While copying the parameter named "model.diffusion_model.output_blocks.5.1.transformer_blocks.0.norm3.weight", whose dimensions in the model are torch.Size([640]) and whose dimensions in the checkpoint are torch.Size([640]), an exception occurred : ('Cannot copy out of meta tensor; no data!',).

my webui ARGS settings
set COMMANDLINE_ARGS=--lowvram --xformers --port 9997
set PYTORCH_CUDA_ALLOC_CONF=garbage_collection_threshold:0.5,max_split_size_mb:22