does FLASH_ATTENTION versions works i don't know
(also it have addition crashes)
also most of versions are not tested at all (and as-is, no any warranty etc...)
so try other versions if it doesn't works
x64 versions not tested too (and it's build with different compiler)
example generated on XP (win32) with lcm models from civ
max resolution that i have get 512x448
last 2024.04.06 version can be build too
but it doesn't load 2GB models in q8_0 mode (looks like it checks free ram?)
later i'm build it too
SD goes to WinXP! I'm build 2024.02.25 version for win32 (cpu only no avx?) and it's works! and it's as slow as on x64 Win7 :D
It can load 2GB models in q8_0 mode or 2GB model + lora in q5_1 mode (maybe with /3GB boot flag it can? load bigger models but i not have it)
sd_2024_02_25_2_XP.zip (i put 7z in zip bcz github prohibits attach 7z in same time zip format doesn't support solid archives...)
in archive binaries for several targets
does FLASH_ATTENTION versions works i don't know (also it have addition crashes) also most of versions are not tested at all (and as-is, no any warranty etc...) so try other versions if it doesn't works x64 versions not tested too (and it's build with different compiler)
example generated on XP (win32) with lcm models from civ
max resolution that i have get 512x448
last 2024.04.06 version can be build too but it doesn't load 2GB models in q8_0 mode (looks like it checks free ram?) later i'm build it too