Model Conversion to Core ML Format for Apple Devices
Core ML — Apple's native ML framework. Models in .mlpackage/.mlmodel format run via Neural Engine (ANE), GPU, or CPU — iOS selects automatically. Result: optimal performance on Apple Silicon with minimal power consumption.
Conversion Tools
coremltools (primary tool):
import coremltools as ct
model = ct.convert(pytorch_model, inputs=[ct.TensorType(shape=input_shape)])
model.save("model.mlpackage")
Supports: PyTorch, TensorFlow/Keras, ONNX as intermediate format.
Core ML Tools Optimization:
- 8-bit quantization:
ct.optimize.coreml.linear_quantize_weights(model, mode='linear_symmetric') - 4-bit palettization for size
- Pruning support
Conversion Nuances
Unsupported ops: not all PyTorch operations have Core ML equivalents. Custom operations require implementation via ct.CompositeOp or replacement with supported operations.
Numerics: float32 → float16 by default on some hardware. Precision checked via numeric tolerance tests.
Shape flexibility: static vs. dynamic shapes. ANE works better with static shapes.
LLM Conversion
Apple CoreML Tools 8+ supports transformer conversion with optimizations. mlx-lm (Apple MLX framework) — more efficient path for LLM on Apple Silicon.
Testing After Conversion
Comparison of PyTorch vs Core ML outputs on test set. Max absolute error < 1e-4 for float16.