How to tell for sure that some layers have been offloaded to CPU/RAM?

[remon-nashid]

... and how many of them.

With a model larger than VRAM, and when specifying -ngl 99, llama-server logs show that all layers are being offloaded to GPU despite they are clearly not.

Currently I check RAM use for clear increase to indicate that some layers are offloaded to RAM, but I don't believe it to be the smartest way.

Any clues? Any additional indicative log messages I'm missing? Ultimately, a display similar to ollama's GPU/CPU percentages would be ideal.

Thank you.