Is --spec-draft-ngl working for Qwen3.x MTP?

[kstoykov]

I'm using Qwen MTP models. They require a little bit more VRAM compared to non-MTP version which is expected because of the draft model (weight + kv cache). I see that there is flag --spec-draft-ngl which must specify how many of draft model's layers to be offloaded to the GPU. I tried 'auto', 'all', '0', '999' but all options resulted in same VRAM consumption and performance. Is this expected? I would expected value 0 to move entire draft model to CPU.

[Diablo-D3]

MTP has no layers to move to CPU. It doesn't use a second model, it uses the same model with different less complex heads. The extra heads only take up about half a gig.