How to run on a single GPU for debugging?

wchengming

Hello, thank you for sharing this great project! I only have access to one GPU (12 GB memory), and I would like to run the code in a debug mode (it does not need to reproduce full performance, just to make sure the pipeline works). Could you please provide guidance on: How to configure configs/multiple.yaml or configs/configs_ABCD.json for single-GPU training? Which parameters (e.g., batch size, sequence length, image resolution, number of workers) should I reduce in order to make it fit into 12 GB memory? Is there a recommended minimal setup or “debug config” that can at least run end-to-end on one GPU? Thank you very much for your help!

idejie

You can try the config for single-gpu: https://github.com/idejie/AR/blob/main/configs/single.yaml but for 12 GB memory, you can try reduce the batch size by yourself.

wchengming

Thanks for the quick reply! I’ll switch to configs/single.yaml and reduce the batch size to fit 12 GB—appreciate the guidance.