MLX / Metal: command buffer execution failed

Apple's Metal Performance Shaders compiler hit an internal error or hardware fault during a kernel launch. Most often caused by:

• Running out of unified memory mid-inference (no clear OOM error like CUDA gives)
• macOS version too old for the MLX features the model uses
• A specific quantization format the Metal backend doesn't support yet

Free unified memory. Macs have no separate VRAM — system memory IS GPU memory. Close apps:

# Show top memory consumers
top -o MEM

Common culprits: Chrome (1-3 GB), Slack (500 MB), Spotify (300 MB), other ML scripts.

Update macOS. MLX requires macOS 13.5+ for full features; some kernels need 14+. Apple → About This Mac shows version. Update via System Settings.

Update MLX:

pip install --upgrade mlx mlx-lm

Try a different quantization. MLX has its own format (-mlx-4bit, -mlx-8bit). GGUF Q4_K_M run via llama.cpp on Metal is also a fallback path:

# llama.cpp on Metal — different code path, sometimes works when MLX doesn't
./main -m model.gguf --n-gpu-layers 999

Reduce context size if the error only happens at long contexts — KV cache is the most common pressure point.