It appears that Leave Applications in Memory is not working with GPU tasks.
Any plans to introduce the feature also for GPU applications?

It appears that Leave Applications in Memory is not working with GPU tasks.
Any plans to introduce the feature also for GPU applications?

No - it was deliberately taken out.

CPU (main system) memory is routinely swapped out to a pagefile by the operating system if things get tight - so LAIM has minimal effect on system performance.

But GPU memory has no swapfile system - so anything left in memory, is in memory. That bit of BOINC was written when 512MB GPUs were common, and many projects (including Einstein, and SETI with cuda 2.3 and above) can only fit one task in that little VRAM. Leaving an app behind in VRAM would prevent any other GPU task running.

But maybe it is time to re-enable it.
From my point of view it is not transparent enough which application it is going to be left in memory, CPU or GPU.
My suggestion would be that in future BOINC versions hould be written Leave CPU application in memory.
Also since it was working in the past, maybe a new option should be inserted smth like:
"Leave GPU applications in memory (experimental)"
Nowadays, most graphic cards have at least 512 MB, and I think I read somewhere that the new OpenCL specifications should allow the GPU to access Shared System Memory.