Apple Silicon: RuntimeError: MPS backend out of memory

macOS imposes a per-process limit on how much unified memory can be allocated to the GPU — by default ~67% of total system RAM. Jobs that try to allocate more (e.g. loading a large model into MPS) hit this limit even though there's free RAM available.

A separate cause: the MPS allocator pools memory and can fragment, OOMing on a 10 GB allocation while the system has 30 GB free.

1. Raise the MPS memory ratio:

# Allow MPS to use up to 80% of unified memory
export PYTORCH_MPS_HIGH_WATERMARK_RATIO=0.0
# Or set an absolute cap (in GB)
export PYTORCH_MPS_MEMORY_LIMIT_MB=49152

HIGH_WATERMARK_RATIO=0.0 removes the soft cap entirely; macOS will swap if needed but the allocation will succeed.

2. Use a smaller / quantized model. PyTorch's MPS backend doesn't support 4-bit GGUF; for quantized inference on Apple Silicon, use llama.cpp or MLX-LM directly:

brew install llama.cpp
llama-cli -m model.Q4_K_M.gguf -p "Hello"

# Or MLX
pip install mlx-lm
mlx_lm.generate --model mlx-community/Llama-3.1-8B-Instruct-4bit --prompt "Hello"

3. Free unified memory by closing other apps. macOS shows actual GPU memory usage in Activity Monitor → GPU tab.

4. Restart the Python process after a previous large allocation — MPS allocator pools don't always release back to the OS until process exit.