Epistates/pmetal
Powdered Metal — High performance LLM fine-tuning framework for Apple Silicon
What's novel
Powdered Metal — High performance LLM fine-tuning framework for Apple Silicon
Code Analysis
11 files read · 3 roundsA comprehensive Rust ML SDK for Apple Silicon providing Metal GPU kernels (FlashAttention, fused LoRA, fused MLP), LoRA/QLoRA/DoRA fine-tuning across ~15 model architectures, knowledge distillation (TAID, KL, JS), RL training (GRPO, DAPO, DPO, PPO), GGUF quantization, and a full CLI/TUI/GUI applicat
Strengths
Genuinely deep GPU kernel implementations (FlashAttention forward+backward with online softmax, fused LoRA with LoRA+ differential LR, fused SwiGLU+LoRA MLP) backed by real Metal compute shaders with auto-tuning and pipeline caching. The workspace architecture cleanly separates concerns across 20+ crates with a consistent trait-based abstraction layer, and the training algorithms (GRPO, DAPO, GSPO, TAID) are real implementations, not stubs.
Weaknesses
The trainer crate has an excessive number of `#![allow(...)]` lint suppressions (20+ at crate level), suggesting some code quality debt. Test coverage appears concentrated in a few crates (lora_trainer has tests, but many other modules like the Metal kernels, distillation losses, and GRPO/DAPO trainers lack visible test files in the tree).
Score Breakdown
Signal breakdown
Innovation
Craft
Traction
Scope
Evidence
Commits
148
Contributors
1
Files
433
Active weeks
7
Repository
Language
Rust
Stars
22
Forks
2
License
NOASSERTION