jrp2014

You can run and debug native mlx-vlm inference, draft reproducible upstream GitHub issues, contribute fixes matching upstream test and lint expectations, benchmark MLX performance with rigorous timing protocols, inspect local Hugging Face cache model discovery, and manage Hugging Face Hub repos, datasets, jobs, and endpoints.

They target ML engineers and contributors working with mlx-vlm vision-language models on Apple Silicon who need to verify harness findings, minimize reproductions, prepare upstream PRs with perf evidence, or manage Hugging Face Hub assets like models, datasets, spaces, and inference endpoints.

The mlx-vlm skills assume a conda plus pip environment and explicitly never use uv. They rely on an editable upstream mlx-vlm checkout, mypy/ty/pyrefly for typing checks, ruff matching upstream black/isort/autoflake hooks, and the hf command line for Hugging Face Hub operations.

Use the benchmarking-mlx-vlm skill: run median-of-N measurements with warmup iterations, call mx.eval or mx.synchronize before stopping timers, follow the peak-memory protocol, and apply A/B discipline across MLX versions when comparing local builds or preparing numbers for maintainers.

No. The upstream-mlx-vlm-issues skill only creates or improves reproducible, maintainer-ready issue drafts from check_models failures or native repros, preferring existing diagnostics artifacts. It will not open a GitHub issue unless the user explicitly asks.