New on‑device fine‑tuning tool
PMetal, a Rust‑based toolkit, now enables LLM fine‑tuning on Metal GPUs and Apple’s Neural Engine — offering a CLI/GUI path to train or adapt models locally on M‑series Macs. That tight Metal/NeuralEngine tooling flips the usual cloud workflow and lets teams iterate on edge models faster. (x.com)
PMetal (“Powdered Metal”) is published as an open-source Rust project on GitHub and is described in the repo and community posts as dual-licensed under MIT/Apache-2.0. (github.com/junpark88/pmetal-apple-finetuning) The codebase exposes custom Metal kernels and a pmetal-metal crate that implement GPU-accelerated tensor operations, and the project advertises explicit MLX framework bindings plus ANE (Neural Engine) integration for Apple Silicon acceleration. (github.com/junpark88/pmetal-apple-finetuning/tree/main/crates/pmetal-metal) The project ships both a terminal/GUI control center described with sections like Dashboard, Devices, Models, Datasets, Training, Distillation, Inference, and Jobs, and a pmetal-cli package published to the Rust ecosystem. (news.ycombinator.com/item?id=47409043; libraries.io/cargo/pmetal-cli) Community notes and the repository list model compatibility with Llama, Qwen, Mistral and Phi-family checkpoints, and the tooling advertises hardware-awareness that detects GPU family, core counts, memory bandwidth, NAX, and UltraFusion topology across M1–M5 chips (the maintainer reported dogfooding on M4 Max and M3 Ultra). (news.ycombinator.com/item?id=47409043; github.com/junpark88/pmetal-apple-finetuning) Published crates and package metadata show active releases for pmetal, pmetal-py and pmetal-cli (examples on crates index and libraries.io list pmetal-cli 0.1.2, pmetal 0.1.0 and pmetal-py 0.2.0) and the repository includes prebuilt signed binaries on its Releases page. (libraries.io/cargo/pmetal; libraries.io/cargo/pmetal-py; libraries.io/cargo/pmetal-cli) The project cites “Unsloth‑style” low-level optimizations and explicit MLX compatibility to route workloads onto Apple’s Neural Accelerators where available, aligning with Apple’s MLX research work that added Neural Accelerator access for M5-class silicon in recent macOS betas. (libraries.io/cargo/pmetal; machinelearning.apple.com)
