llama.cpp server freezes on larger Qwen quants

This smells like a loader-path ceiling disguised as a quant problem. `llama.cpp` can span RAM + VRAM, but `--no-mmap` and Vulkan offload make the practical headroom much tighter than the file-size math suggests.

• –bartowski lists IQ2_XXS at 106.57GB and IQ2_M at 133.28GB, so the jump is big enough to cross a staging threshold even on a 160GB nominal pool once OS, driver, and allocator overhead are counted.
• –`--no-mmap` removes the lazy paging path, so the host-RAM ceiling matters a lot more than the combined RAM + VRAM arithmetic implies.
• –Context-size and KV-cache tweaks only affect runtime cache, not the weight load that is already on the edge.
• –AMD/Vulkan multi-GPU offload has recent allocation-failure reports on second-GPU paths, so the MI60 pair is a plausible weak spot here.