llama.cpp Gemma 4 burns RAM on prompts

This looks like the classic long-context trap: the model fits, but the running conversation state does not. `-ngl` can keep weights on GPU, yet it does not make the separate context-memory budget disappear.

• –`-c 102400` is huge, and even quantized KV cache settings still scale sharply with prompt length
• –Multiple users in the thread report the same RAM growth pattern, which makes this look systemic rather than machine-specific
• –Related llama.cpp issues show KV cache behavior can land in CPU RAM in CUDA setups, which matches the symptom users are seeing
• –The practical lesson is that context length is now a deployment constraint, not just a quality knob
• –If this is a regression, the useful bug report is the exact build, backend, and cache flags, since those likely determine where the memory goes