CUDA out of memory despite having a lot of free GPU memory

[Mh-Magdy]

torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate 1.37 GiB. GPU .
The GPU memory is totally free and i still get this error each run?!

[jingheya]

What is the resolution of your images? And how about the total GPU memory of your device?

[Mh-Magdy]

Test Image Resolution: 612x408
My GPU memory: 6144MiB

[jingheya]

A little bit odd (⊙_⊙). I tried an image with a resolution of 685x609, but it only required 2947MiB of memory.
Have you enabled half precision?

[Mh-Magdy]

I got the same error even with the infer.sh script on the example images with half precision enabled by default inside infer.sh.

I tested Lotus on google colab [T4 gpu] and ran the infer.sh script and got the same error.

[jingheya]

Thanks for pointing out the problem, let me check the code.

[jingheya]

Hi, you can add torch.cuda.empty_cache() after each inference. Please refer to this line.

[Mh-Magdy]

Thank you for your prompt response. I'll try the fix and let you know the update.
Thank once again for your great work.

[haodong2000]

Issue closed, please re-open it upon any problems 🤗