What cpu doing while offloading all layers in GPU

[FNsi]

We already know, with a dumb gpu, the speed is lower than the CPU mode, even -t let the cpu usage 100%.

That's the real question.

(that might be changed cause I didn't update llama.cop a week, but I did not seen any related in recent commits.)

[bullno1]

Last time I checked, spin waiting.

It's a known issue.

It makes sense for CPU or hybrid CPU&GPU though.

[FNsi]

I am thinking might it be possible like mmap both vram and ram utilise them...