PMetal: Rust toolkit for fine‑tuning LLMs on Apple Silicon
A new full‑stack Rust toolkit called PMetal surfaced for LLM fine-tuning on Apple Silicon, leveraging Metal GPU kernels and the Neural Engine with CLI/TUI/GUI tooling shared. That gives engineers a native, systems-level path for local model ops and could change how we prototype on-device model updates without heavy server infrastructure.
PMetal’s primary repositories are published under the names [Epistates/pmetal] (github.com) and a visible fork [junpark88/pmetal-apple-finetuning] (github.com), and the project includes both Apache‑2.0 and MIT license files in the repo. (github.com) The codebase is implemented in Rust, with dedicated crates for command-line tooling and bindings shown in the repo tree [crates/pmetal-cli] (github.com) and a Metal kernel crate at [crates/pmetal-metal] that contains the custom GPU shader implementations. (github.com) The README documents an ANE pipeline that invokes private AppleNeuralEngine.framework APIs and MIL 1.3 kernels with zero‑copy IOSurface transfers and describes a “dynamic weight pipeline” comprising nine compiled kernels. (github.com) Examples in the repository demonstrate LoRA fine‑tuning flags (e.g., --lora‑r values 4, 8, 16, 32) and sample commands for training and inference against models such as Qwen in the provided examples directory. (github.com) Community notes and the initial HN announcement list support for a range of open‑weight families including Llama, Qwen, Mistral and Phi, and state the project is hardware‑aware (detecting GPU family, core counts, memory bandwidth, NAX and UltraFusion topology). (news.ycombinator.com) Distribution artifacts include published crates/binaries (libraries.io and crates listings reference pmetal and pmetal‑cli) and a Releases page in the repo for signed builds, with libraries.io showing a pmetal 0.1.0 entry and pmetal‑cli metadata. (libraries.io)
