llama.cpp Loops Plague Local LLM Runs

This looks less like a raw hardware problem and more like an overloaded long-context, tool-calling setup that is pushing the runtime into pathological behavior.

• –`CTX_SIZE=131072` is extremely aggressive and makes KV cache behavior a likely stress point, especially once the model starts looping instead of terminating cleanly
• –llama.cpp’s README emphasizes Apple Silicon support and hybrid CPU+GPU inference, but those strengths still depend on sane context, cache, and batching choices
• –q8_0 KV cache is a tradeoff, not a guarantee of stability; recent llama.cpp issues still show edge cases around KV quantization and large-context behavior
• –Remote, headless orchestration adds another failure layer: if stop conditions or tool-call plumbing are off, the model can look like it is “thinking forever” even when the backend is functioning
• –The inconsistency across Qwen and Zen4 suggests a mix of model behavior and backend configuration, not a single broken model family