OOM appears using DeepFloyd_IF (I use 32G video memory)

[qixuanwang-233]

As in the title, OOM appears when I use DeepFloyd IF through diffuser (I use 32GB video memory).

I used the diffuser example directly:

[Bigfield77]

Stage 3 is pretty useless, it's just a generic upscaler by stability.

does it still oom if you bypass it?

[qixuanwang-233]

Stage 3 is pretty useless, it's just a generic upscaler by stability.

does it still oom if you bypass it?

Thank you for your response, stage 3 is all I need and I solved this problem by using xFormers.

[Bigfield77]

Good for you!
Btw, how do you use xformer with deepfloyd?

*update: found it enable_xformers_memory_efficient_attention() on the pipeline

[Snimm]

@qixuanwang-233 Can you share me your exact dependencies/ requirements.txt ? I cant enable xformer I get

NotImplementedError: Memory efficient attention with `xformers` is currently not supported when `self.added_kv_proj_dim` is defined.

I cannot upgrade my pytorch because in the requirements.txt of deep floyd say it needs to be less than 2.0.0

Can you please share your requirements.txt / the dependencies you are using?

[yangsencai]

@qixuanwang-233 Can you share me your exact dependencies/ requirements.txt ? I cant enable xformer I get你能告诉我你的dependencies/ requirements.txt文件吗？我不能启用xformer

NotImplementedError: Memory efficient attention with `xformers` is currently not supported when `self.added_kv_proj_dim` is defined.

I cannot upgrade my pytorch because in the requirements.txt of deep floyd say it needs to be less than 2.0.0我无法升级我的pytorch，因为在deep floyd的requirements.txt中说它需要小于2.0.0

Can you please share your requirements.txt / the dependencies you are using?你能分享你的requirements.txt /你正在使用的依赖吗？

I used this dependencies which can avoid OOM and deep-floyd-if also run well：

tqdm==4.61.2
numpy==1.24.2
torch==2.4.1
torchvision==0.14.1
omegaconf==2.3.0
matplotlib==3.7.1
Pillow==9.4.0
huggingface_hub==0.26.5
transformers==4.46.3
accelerate==1.0.1
diffusers==0.31.0
tokenizers==0.20.3
sentencepiece==0.1.99
ftfy==6.1.3
beautifulsoup4==4.11.2
protobuf==3.19.6
gradio==4.44.1