Menu
Guide MLX Metal profiling, mx.fast escalation, custom Metal kernels, and C++ extensions when profiling proves kernel-level bottlenecks.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
Based on SOC occupation classification
Optimize Python MLX inference and generation loops with warmup, batching, cache handling, synchronization, quantization, and memory checks.
Route Python-first MLX optimization work on Apple Silicon to focused audit, training, inference, Metal, or bridge workflows.
Audit Python MLX repos for lazy-eval, synchronization, compile, dtype, memory, progress, and benchmark issues.
Advise on Python-first MLX integration with Swift, C, C++, and non-native language boundaries.
Optimize Python MLX training loops with value_and_grad, accumulation, checkpointing, dtype, validation cadence, memory telemetry, and progress reporting.
| name | mlx-metal-kernels |
| description | Guide MLX Metal profiling, mx.fast escalation, custom Metal kernels, and C++ extensions when profiling proves kernel-level bottlenecks. |
Use this skill only after normal MLX-level issues have been checked or when the user explicitly asks about Metal capture, custom kernels, or native extensions.
Before any Python execution, use the target repo's .venv. Never install
Python packages globally.
../../references/profiling-and-metal.md../../references/memory-and-dtypes.md../../references/compile-and-transforms.mdmx.fast, custom Metal kernels, or C++ extensions only when the
operation is a measured bottleneck and built-in MLX ops are insufficient.State whether custom native work is justified, which evidence supports that decision, and what maintenance burden it creates.