search

Stability-AI / StableCascade

Official Code for Stable Cascade
MIT License
6.44k stars 518 forks source link

Is there anyway to make the gpu vram usage smaller to train controlnet? #88

Open universewill opened 4 months ago

universewill commented 4 months ago

I used an A100 40g vram gpu to train controlnet with 1 batch size and 512 image size, and get torch.cuda.OutOfMemoryError, and i switched to H100 with 80gb ram, still cuda out of memory.

How much gpu ram needed to train controlnet? Is there anyway to make the gpu vram usage smaller?

asutermo commented 4 months ago

Are you using 1B or 3.6B? That will make a difference (though still pretty hefty).

universewill commented 4 months ago

Are you using 1B or 3.6B? That will make a difference (though still pretty hefty).

changed to 1B model， still cuda out of memory ...