SDXL: index 21 is out of bounds for dimension 0 with size 21

[zixianwang2022]

Hi,

Run command:

python3 main.py --dataset "coco-1024" --dataset-path coco2014 --profile stable-diffusion-xl-pytorch --model-path /work1/zixian/ziw081/CM/models/SDXL/official_pytorch/fp16/stable_diffusion_fp16/ --dtype fp16 --device cuda --time 30 --scenario Offline --threads 2

When I am running with --threads>1, we are seeing

ERROR:main:thread: failed on contentid=['/work1/zixian/ziw081/inference/text_to_image/coco2014/validation/data/COCO_val2014_000000386589.jpg'], index 21 is out of bounds for dimension 0 with size 21

I saw similar event here: #1519 (comment) , but there's no actual solutions posted regarding this error.

We believed that it was a direct cause of threading.Thread of the source code.

[arjunsuresh]

Hi @zixianwang2022, yes improving the code for multi-threading can be a useful change but that's unlikely to improve the performance.

[zixianwang2022]

Hi @zixianwang2022, yes improving the code for multi-threading can be a useful change but that's unlikely to improve the performance.

Hello @arjunsuresh ,

When we have --threads>1, we are seeing a significant improvement in terms of performance. However, the run isn't valid as it is not passing the accuracy test for the above error. Are you aware of any changes that can avoid getting the error?

[arjunsuresh]

@zixianwang2022 unfortunately no. Since the performance was expected using 1 thread we haven't checked the reference implementation since.