Core ML Conversion

How to turn a model you've already trained in PyTorch, TensorFlow, or Keras into a Core ML model with Apple's Python coremltools package, and how to prove the converted model gives the same answers as the original.

When to Use

Use this skill when you're:

• Converting a trained model to .mlpackage for the first time
• Choosing between tracing and exporting a PyTorch model
• Picking a minimum deployment target, and seeing what each one enables
• Converting a transformer whose KV cache should live in model state
• Converting a model that was already compressed before conversion
• Hitting coremltools import errors, unsupported operations, or output that differs from the original

Example Prompts

Questions you can ask your assistant that will draw from this skill:

• "How do I convert my PyTorch model to Core ML?"
• "Should I use torch.jit.trace or torch.export?"
• "My converted model gives different results from PyTorch. How do I find out why?"
• "What changes if I set the deployment target to iOS 18?"
• "How do I convert a transformer with a stateful KV cache?"
• "coremltools says an op isn't supported. What are my options?"

What This Skill Provides

• The unified converter – Supported source formats, and tracing versus exporting a PyTorch model
• Deployment targets – What each minimum_deployment_target unlocks, such as fused attention at iOS 18
• Stateful conversion – Declaring a KV cache as model state, with the constraints that apply
• Already-compressed sources – Converting a model that was compressed during training
• Parity validation – Comparing converted output with the original before you ship
• Common failures – Import errors, unsupported operations, and output divergence

Documentation Scope

This page documents the coreml-conversion skill (in the axiom-ai suite), which your assistant loads when you convert a trained model to Core ML. Language-model-scale models headed for the iOS 27 Core AI runtime use a different toolchain, covered by Core AI.

Related

• ios-ml – The hub that routes Core ML work to each stage
• coreml-compression – Shrinking the converted model
• coreml-deployment – Running the converted model in your app
• coreml-ref – An index of the coremltools conversion and optimization APIs

Resources

WWDC: 2024-10159

Docs: /coreml