Summary

Fixes #1964 — When n_gpu_layers=-1, host RAM used for model loading is not released.

Problem

When use_mmap=True (the default), the entire model file is memory-mapped into the process address space. Even after all layer weights are copied to VRAM via n_gpu_layers=-1, the mmap'd pages remain resident in the OS page cache and show up as consumed RAM in Task Manager / htop. This memory is not released until the Python process exits.

Fix

In Llama.__init__, when all three conditions are met:

• n_gpu_layers == -1 (all layers offloaded)
• use_mmap == True
• llama_supports_gpu_offload() returns True

…mmap is automatically disabled. With mmap off, llama.cpp reads weights via a temporary buffer that is freed after GPU upload, so host RAM drops back down after loading.

A verbose stderr message is printed when this auto-disable kicks in. Users can override with use_mmap=True explicitly.

Changes

• llama_cpp/llama.py — auto-disable mmap logic in Llama.__init__
• llama_cpp/server/settings.py — updated use_mmap field description to document the behavior
• tests/test_mmap_gpu_offload.py — 6 unit tests covering the auto-disable logic

Testing

All 6 tests pass. The tests use mocked native libraries so they don't require a compiled llama.cpp binary.

Note: Full integration testing (verifying actual RAM reduction with a real model + GPU) is not feasible in CI since it requires a GPU and a model file. The fix has been validated at the logic level; end-to-end confirmation would need a machine with a GPU.