What problem does it solve? When a Vision-Language Model fails inside the check_models harness, it is hard to tell whether the bug lives in the harness or in upstream mlx-vlm. This Skill provides a disciplined workflow for building minimal native reproductions using python -m mlx_vlm.generate or the load/apply_chat_template/stream_generate Python path, so failures can be attributed correctly and reported upstream. ## Core Features & Use Cases - Native CLI and Python repro patterns: Canonical commands for text-only and image inference, including revision pinning, trust-remote-code, and deterministic settings (temperature 0.0, bounded max-tokens). - Harness-aware starting points: Reuses existing artifacts such as issue drafts, diagnostics reproduction blocks, and results.jsonl metadata instead of inventing new commands. - Failure routing and thinking-output checks: A symptom-to-cause table (unsupported model_type, missing weights, processor errors, media shape errors) plus rules for judging thinking-block output. - Use Case: A model crashes in check_models with a chat-template error. You pull the native CLI command from the issue draft, confirm flags with --help, run one model per process, and confirm the failure is upstream before filing an issue. ## Quick Start Ask the assistant to reproduce a failing model natively with mlx-vlm using the conda mlx-vlm environment, starting from the diagnostics or issue-draft command for that model.
