| name | mlx-performance-audit |
| description | Audit Python MLX repos for lazy-eval, synchronization, compile, dtype, memory, progress, and benchmark issues. |
MLX Performance Audit
Use this skill for repo-wide MLX performance reviews and unknown bottlenecks.
Python Environment
Before any Python execution, use the target repo's .venv. Never install
Python packages globally.
Required References
../../references/mlx-core-concepts.md../../references/eval-and-synchronization.md../../references/compile-and-transforms.md../../references/memory-and-dtypes.md../../references/reporting-format.md
Workflow
- Inspect repo state with
git --no-pager status --short.
- Locate dependency files, MLX imports, training loops, inference loops,
dataloaders, benchmark scripts, and progress-reporting patterns.
- Run
../../scripts/mlx_audit.py from this plugin repo when
the target repo is local and scanning is useful.
- Keep candidate findings separate from verified findings.
- Recommend measurement before changes: warmup, synchronization, repeated runs,
memory telemetry, correctness checks, and representative input sizes.
Report Shape
Use ../../templates/optimization-report.md as the report structure. Every
finding needs file, line, evidence, candidate impact, confidence, verification,
and residual risk.
