LM Studio users battle Gemma 4 memory leaks

This is a classic "abstraction leak" where user-friendly GUIs fall behind the rapid architectural shifts in the underlying GGML/llama.cpp engines.

* The memory "explosion" is likely due to how Gemma 4 handles KV cache or context windowing, which requires explicit optimization flags that LM Studio hasn't yet toggled by default for these models.

* Manual model reloads are a functional but inefficient "band-aid" fix for a problem that requires backend parameter passthrough.

* Developers of local LLM wrappers must prioritize exposing "expert" flags or implementing auto-detection for specific model families to maintain stability for non-technical users.