I'm wondering if anyone else has come across this. I've tested the same model on llama.cpp and vLLM with similar settings and quantizations. The performance and concurrency in vLLM are much noticeably better, but sometimes the model feels less reliable. Some things I've noticed: * More mistakes with formatting and tool calls * Forgetting context suddenly * Sometimes acting like messages didn't exist * Lower quality code even with similar parameters I'm not trying to start a compari...