>>103084349
its probably blockswap, so much time is lost because you have to keep moving the latents around, and also torch compile does add a bit to the first step, thats doing the maths to optimize vram, and both are vram saving options so yeah. 6090 when