Thanks! These things you're mentioning like "You may be able to offload some layers to GPU...", "You can keep the KV cache on GPU..." configured as part of the llama.cpp? I wouldn't know what to prompt with or how to evaluate "correctness" (outside of literally feeding your comment into claude and seeing what happens).

Aside: what is your tooling setup? Which harness you're using (if any), what's running the inference and where, what runs in WSL vs Windows, etc.

I struggle to even ask the right questions about the workflow and environment.

In my case, I was also running an ASR model and a TTS model so it was a bit much for my RTX 3090. I opted to offset like 5 layers to the cpu while adding a GPU-only speculative decoding with their 0.8B model.

Working well so far.