llama.cpp oom issue

In three linesUser reports memory leak in llama.cpp with Qwen3.6-27B-MTP-GGUF after 20-40 minutes of active use. Process gradually consumes more system RAM despite various configuration attempts (--no-mmap, --cache-ram 0, without MTP). Issue persists across multiple builds and Docker images.

Summary generated by Claude — human-verified