add-afm-model

To add HuggingFace MLX models to AFM, you invoke the onboarding workflow with a model ID or URL. The system analyzes the model's config.json to determine MLX status, architecture, and registry compatibility before applying necessary Swift backend patches.

MLX model architecture compatibility checking works by analyzing the model's config.json to determine its MLX status and architecture. A structured decision workflow then classifies the model as already supported, requiring a port from Python, or needing a new architecture entirely.

Yes, you can onboard vision-language models (VLMs) into AFM. The onboarding workflow includes specific considerations for VLM variants and provides guidance for registering them in the appropriate VLM registries alongside the standard LLMTypeRegistry.

Yes, the AFM onboarding workflow supports Mixture of Experts (MoE) models. The structured decision process includes specific MoE considerations to classify the model and apply the correct backend patches for successful integration.

Patching the Swift backend involves classifying the model's architecture requirements and then applying structured code patches. After patching, you must register the model in LLMTypeRegistry and validate the integration using a dedicated test workflow.

You need to port Python models to Swift for AFM integration when the model architecture is not already supported natively. The workflow identifies this requirement during config analysis, guiding you through the Swift backend patching and LLMTypeRegistry registration.