Running tutorial_train.py still OOM

[xwg0508]

I enable the low_vram_shift, set the batch_size=1, and set accumulate_grad_batches=4, besides I import xformers, but still OOM. My gpu memory is 10.75GB, why it isn't work?

[shravankumar147]

What's your GPU ?

[xwg0508]

RTX 2080 Ti

[qsun1]

> 24G is all you need (e.g. RTX 3090, V100)

[dennisalp]

I enable the low_vram_shift, set the batch_size=1, and set accumulate_grad_batches=4, besides I import xformers, but still OOM. My gpu memory is 10.75GB, why it isn't work?

Same problem. Couldn't run training out-of-the-box on AWS EC2 p3.2xlarge (V100, 16GB VRAM) with all memory reduction options active.

[AparnaAgrawal02]

if you have acess to 4 2080 gpus ,will it work!? also with lower preecision!?

[geroldmeisinger]

you can use the HuggingFace Diffusers ControlNet training script https://huggingface.co/docs/diffusers/training/controlnet Min 12GB on Windows, or 8GB on Linux (requires DeepSpeed, which is only available on Linux) I wrote a long article on how to train control nets https://civitai.com/articles/2078

[geroldmeisinger]

all duplicate concerning "RAM and out of memory exceptions (OOM)": https://github.com/lllyasviel/ControlNet/issues/21 https://github.com/lllyasviel/ControlNet/issues/33 https://github.com/lllyasviel/ControlNet/issues/191 https://github.com/lllyasviel/ControlNet/issues/236 https://github.com/lllyasviel/ControlNet/issues/241 https://github.com/lllyasviel/ControlNet/issues/247 https://github.com/lllyasviel/ControlNet/issues/294 https://github.com/lllyasviel/ControlNet/issues/301