[bug]: Memory Management - freeing memory when not generating? about invokeai

[mcbexx]

Ah, thank you, this seems to have fixed it.

It wasn't explicitly enabled in my current config file (it just contained the custom paths to my output dir and the db path), but I can see I have a .bak file from April 2nd that had it enabled. Never had this particular problem before, but I assume that with the introduction of 4.x it has been set to true by default and the memory handling itself changed internally?

[psychedelicious]

Have you tried disabling lazy_offload? It's in the table here: https://invoke-ai.github.io/InvokeAI/features/CONFIGURATION/#invokeai.app.services.config.config_default.InvokeAIAppConfig

[psychedelicious]

It had always been enabled by default, but we were handling the config file a bit awkwardly until v4 - we wrote defaults to the config file. This caused a number of headaches.

When we did the v4 upgrade and updated the config file setup, we only copied config file settings that were not set to the default value. In this case, we skipped lazy_offload for your config because it was set to the default value, so it disappeared from your config file.

As for why you only recently started running into this problem, I'm not sure... The model loading code was totally rewritten for v4, it's possible the lazy offload wasn't actually working correctly before.

Anyways glad that's working now.