performance overhead

[231929]

Hello! Regarding the code snippet we have, I was wondering how many 3090 GPUs were used to run it. I tried running it with two 3090 GPUs and a batch size of 1, but it barely worked and was quite struggling. Could you please let me know the GPU configuration that was used for this code? Thank you!

[Ayews]

I apologize for the late response. During training, we used only a single 3090 GPU and achieved suitable results. Your batch size is too small; generally, it should be at least 4 or 8, and if possible, 16. I suggest you try a larger batch size and aim to fully utilize the available GPU memory.