About maximum GPU memory consumption

Hi, It is observed that when inference first starts, there is usually a very short period of high memory usage in Pytorch.
For example, in the provided demo video for docker sanity check, the maximum GPU memory consumption is 2872 MiB when the inference starts. However, when the inference is stable, it keeps comsuming only 908 MiB GPU memory. So, I want to ask whether the definitin of maximum GPU memory consumption only consider the period when inference is stable and ignore the peak memory usage when the inference just starts?

Yes! And the code is already availalbe on the evaluation page. You can find the details in the code.