Apple’s Xcode 17.4 introduces live Swift model debugging for on-device inference

Tech · 5 min read

Apple released Xcode 17.4 with first-party support for live inspection of Swift-based ML models running on devices and Apple Silicon Macs. Developers can now attach a debugger to a running Core ML or Swift-ML model and view layer outputs, tensor shapes, and memory usage in real time, without serializing intermediate tensors to disk.

The tool integrates with Instruments and provides a timeline view that correlates model inference events with CPU and GPU utilization, helping teams find bottlenecks introduced by custom layers or preprocessing. Apple also added presets for common on-device optimizations, like quantization-aware conversion and delegate selection for Neural Engine vs. GPU execution.

For UX and product teams, Apple updated the Simulator to emulate thermal throttling and battery conditions so designers can prototype performance-aware experiences. The release notes emphasize reduced iteration loops and sandboxed inspection to protect model IP while enabling debugging across developer and QA environments.