low-memory

[Dan-Burns]

Hi,

I've have a GPU with 24G and have been able to predict structures over 700 aa. I tried --low-memory on a 1200 aa structure but it runs out of memory. But I believe low-memory is the reason I was able to get a 900 aa prediction.

Has anybody explored how the scaling works for maximum sequence length and memory usage between normal inference and with the --low-memory flag?

Does using ESM embeddings significantly affect memory usage? I'm also interested in your thoughts on whether these embeddings improve the output - from limited experience and simply by the looks of it, it seems to.

Thanks,
Dan

[arogozhnikov]

Does using ESM embeddings significantly affect memory usage?

it doesn't affect memory at all (unless you can't run it, but with 24GB you're ok)

I'm also interested in your thoughts on whether these embeddings improve the output - from limited experience and simply by the looks of it, it seems to.

yes, they improve results; specially when mining relevant MSAs is hard

low_memory=True cleans up maybe 25% of memory (don't expect magic there) and is used by default