--medvram triggers OOM after a few generations.

[KatonRyu]

I'm asking the question here rather than submitting a bug report, since I'm relatively certain the problem is in the settings on my end. When I run the webui with --medvram enabled on my 8GB RTX 2070, generation is smooth.

However, after a few generations, sometimes around five, sometimes only two, a CUDA OutOfMemory error is thrown. I presume this has something to do with the 'parts' of the model being swapped in VRAM (excuse the terminology, I'm not that at home in this yet), but not enough VRAM being available to perform it.

I have system memory share enabled, however, so I would expect that it would simply fall back to using system RAM, as it would when I have medvram disabled, instead of crashing hard.

I don't necessarily need medvram to generate, but it speeds up the process tremendously compared to using my system RAM, so I would like to be able to keep it on if at all possible. Is medvram at all compatible with the system memory share? I've seen sysmemfallback recommended to be disabled, but given that I generate at 1024x1024 (since the model I use just doesn't work well at 512x512) I just can't get away with turning it off, since doing so guarantees a crash.

Hazarding a guess, I would say part of the problem is that I'm still using my PC while generation is running, diverting my limited VRAM to other applications and leading to the OOM when it spikes at the end of generation, but again, I'd expect that the system memory share would handle that.

It's entirely possible that I'm simply asking too much, in which case I'll just stick to keeping medvram off and using system RAM, but if anyone has any insights I'd be glad to hear them. Thanks!