Segmentation fault (core dumped) when sdxl-turbo inference with torch 2.2.1+cu118

[rickyxie2004]

Describe the bug

When sdxl-turbo inferencing, I encountered Segmentation fault (core dumped) after loading the model. And this will happen with torch 2.2.1+cu118(xformers0.0.25+cu118), and it will not happen with torch 2.0.1+cu118(xformers0.0.20+cu118). However, I need to run under torch 2.2.1, so could anyone help me to solve this issue? Thanx!

Reproduction

code:
pipe = AutoPipelineForText2Image.from_pretrained( cache_dir, torch_dtype=torch.float16, variant="fp16")

pipe.to("cuda")

prompt = "A cinematic shot of a baby racoon wearing an intricate italian priest robe."

image = pipe(prompt=prompt, num_inference_steps=1, guidance_scale=0.0).images[0]

Logs

Loading pipeline components...: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 7/7 [00:01<00:00,  4.47it/s]
  0%|                                                                                                                                                                           | 0/1 [00:00<?, ?it/s]Segmentation fault (core dumped)

System Info

python==3.10
torch==2.2.1+cu118
diffusers==0.26.3
accelerate==1.0.1

Who can help?

@yiyixuxu @sayakpaul @DN6

[sayakpaul]

Easiest is to update your PyTorch version and see if this error goes away.

Maybe this is some setup related bug? I.e., the 2.2.1 installation isn't compatible with the CUDA setup you have locally?

[rickyxie2004]

Easiest is to update your PyTorch version and see if this error goes away.

Maybe this is some setup related bug? I.e., the 2.2.1 installation isn't compatible with the CUDA setup you have locally?

Thanks for your reply! I find out this might be some problems with xformers==0.0.25, when I downgrade xformers to 0.0.20, the problem solved. But I still wonder the reason.

[sayakpaul]

Well, when using PyTorch 2.0 or greater and unless explicitly specified, we always use SDPA (i.e., F.scaled_dot_product_attention()), so, I don't that is an issue here. Am I missing something else?

[github-actions]

This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.

Please note that issues that do not follow the contributing guidelines are likely to be ignored.