Backend-Sensitive Numerical Differences and Discrete Decisions in Core MLThis is the full text of my paper, republished from Zenodo, with formatting adapted for the web. Robert Thomas Collins · Independent Researcher Abstract Core ML compute-unit choices can change individ2d ago · 43 min readselfenrichment.hashnode.dev
dgx-spark-inference - Keeping good habits around local inferenceLocal inference has a habit of becoming folklore. You launch a model from a shell history fragment you can never quite find again. A context setting lives in a note somewhere. A faster quantization geJul 1 · 8 min readselfenrichment.hashnode.dev
BF16 Edge Cases When Porting CUDA-Oriented TTS Models to MPS and MLX-AudioWhen porting CUDA-oriented TTS models to Apple Silicon, one subtle but important issue is BF16 behavior on Metal. This has shown up in DiaTTS work on MPS, and the same class of issue appears relevant Jun 3 · 8 min readselfenrichment.hashnode.dev
Revisiting MetalGraph: What I Learned About MLX, Replay, and Sidecar PerformanceTo set the stage, this post covers follow-up work that made MetalGraph feel less like it is doomed to be only a pure investigation tool and more like something that could be useful in practice. MetalGMay 11 · 16 min readselfenrichment.hashnode.dev
The Honest Limits Reached With MetalGraphMetal Graph is an early-stage reusable explicit graph execution API over Apple Metal. MetalGraph started with a simple question: can I (or rather, chatGPT and I) build something graph-like around ApplMay 10 · 14 min readselfenrichment.hashnode.dev
