Low processing performance

[cyberbrain666666]

I have a problem with the automatic1111 software. I have a RTX 2070 super card which has 8 GB RAM Windows 10 64 computer has 32 GB RAM. Between Feb 3 and Feb 11, 2024 CET there was some strange update in the automatic1111 program. Which made the program, which used to perform 3 iterations every 1 s, now perform 1 iteration every 3 seconds in my case. The program started generating images 10 times slower. Installing older drivers (December 2022), with which the program worked fine, did not change anything. I did this because I thought that was the problem. And it turned out that the program generates even slower. With the current version 1.80, the program runs even slower doing 1 iteration every 5 seconds.

[cyberbrain666666]

All I have to do is take one arbitrary lore it slows down 4-5 times and I add a second, even the smallest available lore it slows down unimaginably. I am currently generating without lore and surprisingly it can, but it doesn't do everything. How to fix it ?

[ostap667inbox]

Enable Process Manager and observe the VRAM status during generation. You are describing a situation where VRAM runs out of VRAM and additional system RAM starts to be used. At this time, the generation process slows down many times.

In the NVIDIA Panel settings, you can disable using RAM when VRAM is running low for python.exe. This will not solve the problem, you will just get an 'Out Of Memory' error instead of the generation process hanging.

The very cause of this situation is a bug in VRAM management in A1111 (and in Forge too). In some cases after another image generation when you connect a new LoRA you get a bug in the form of VRAM overflow. The bug remains until you restart the WebUI, also sometimes selecting a different checkpoint helps.

[cyberbrain666666]

It is like this all the time, nothing helps. Previously, this problem did not exist for me. It has nothing to do with memory. Because it does not depend at all on the size of the LoRa And if I give one more extra LoRa even as small as it is, it slows down many times over. In the past, I used to use 30 LoRas each Np. I combined them into one. This happened after the program was updated between February 3 and February 11, 2024.

[ostap667inbox]

Re-read more carefully what I wrote above. I didn't say that it somehow depends on the size or number of LoRa.

[cyberbrain666666]

But this problem did not occur before, or the software was deliberately broken....
But then all the presets were put in one file and there was a limit to the number of characters, not the number of presets, and I'm forced to save everything I create in my file on the flash drive