llama.cpp fixes Gemma 4 VRAM bloat

This is the kind of unglamorous runtime fix that determines whether a model feels usable or broken in practice.

• –The win is not about raw model quality; it’s about memory accounting, which is often the difference between “runs” and “OOMs”
• –Community reports suggest the fix landed in a recent llama.cpp update and may also be reflected in packaged apps like LM Studio, so the exact behavior depends on which backend build you’re on
• –The improvement matters most for local inference and agent workflows, where KV cache size quickly becomes the bottleneck at higher context lengths
• –It also underlines how much open-model UX depends on inference-stack maintenance, not just model releases
• –For developers, the practical takeaway is to update the runtime before blaming the model or resizing hardware