## 🌟 Summary `ultralytics-inference` `0.0.16` is a reliability-focused release that mainly fixes a **CoreML crash on macOS** 🍎, while also improving test coverage and making default model input sizing more consistent across YOLO models 📦. ## 📊 Key Changes - 🐞 **Fixed a CoreML crash on macOS** - The biggest change in this release is a fix for a CoreML loading and warmup failure tied to an internal input rename like `graph_input_cast_0`. - The project now uses **CoreML NeuralNetwork format** instead of the problematic MLProgram path in this case, which avoids the input-name mismatch entirely. - ✅ **Added regression tests for CoreML** - New tests were added to make sure this CoreML crash does not come back. - Tests also cover a previously known warmup issue where a specific `GatherElements` warning can appear during dummy-input warmup. - An end-to-end CoreML integration test was added for stronger real-world validation. - 🍏 **Expanded macOS CI coverage for CoreML** - CI now includes **CoreML-specific checks on `macos-latest`**, so CoreML behavior is tested automatically during development. - This should help catch Apple-platform issues earlier before release. - 📐 **Improved default image-size handling** - Default input sizes are now centralized instead of being scattered as hardcoded values. - Standard YOLO models use a default of **640×640**, while **OBB** models use **1024×1024** when model metadata is missing. - The model now resolves image size more consistently from: 1. user config, 2. model metadata, 3. task-specific defaults. - 🛡️ **Safer CoreML platform handling** - CoreML is now more carefully gated for supported macOS versions. - If CoreML is requested on unsupported macOS versions, it falls back more safely instead of failing unexpectedly. - 🔧 **CI maintenance improvements** - Added Rust toolchain verification steps to CI. - Reordered formatting checks to fail faster. - Updated a GitHub Actions dependency for coverage tooling. - 🚀 **Version bump to `0.0.16`** - Documentation and package metadata were updated for the new release. ## 🎯 Purpose & Impact - 🍎 **Better CoreML reliability for Mac users** - If you run `ultralytics-inference` with CoreML on macOS, this is the most important improvement in the release. - It should reduce crashes during model load or warmup, especially for YOLO models exported to ONNX and run through CoreML. - 🔁 **Lower risk of regressions** - The new unit and integration tests make it much less likely that the same CoreML bugs will silently return in future updates. - 🤖 **More predictable model behavior** - Centralized default image sizes mean models are less likely to behave oddly when metadata is incomplete or missing. - This is especially useful for users working with different YOLO task types, including OBB models. - 🧰 **Improved developer confidence** - Better CI checks on macOS and toolchain validation should make builds more reliable and debugging easier for contributors. - 👥 **Broad user impact** - **Mac/CoreML users** will see the biggest benefit from this release. - **All users** benefit indirectly from cleaner defaults, better testing, and stronger CI stability. ## ⭐ Most Important Takeaway The headline change in `v0.0.16` is a **fix for a macOS CoreML crash** by @onuralpszr 🎉. For users deploying YOLO models with CoreML, this release should make inference startup and warmup **more stable, more predictable, and easier to trust**. ## What's Changed * refactor: ♻️ remove magic numbers default image sizes for YOLO models by @onuralpszr in https://github.com/ultralytics/inference/pull/170 * Bump taiki-e/install-action from v2.77.7 to v2.78.1 in /.github/workflows by @UltralyticsAssistant in https://github.com/ultralytics/inference/pull/171 * ci: 👷 add toolchain checks and re-order formatting step in CI workflow by @onuralpszr in https://github.com/ultralytics/inference/pull/172 * `ultralytics-inference 0.0.16` fix: 🐞 resolve CoreML graph_input_cast_0 crash and add regression tests by @onuralpszr in https://github.com/ultralytics/inference/pull/173 **Full Changelog**: https://github.com/ultralytics/inference/compare/v0.0.15...v0.0.16