Feature Flags

PMetal uses Cargo feature flags to control which crates and capabilities are compiled.

Feature Matrix

Feature	Default	Crate	Description
core	Yes	pmetal-core	Foundation types, configs, traits
gguf	Yes	pmetal-gguf	GGUF format support
metal	Yes	pmetal-metal	Metal GPU kernels
hub	Yes	pmetal-hub	HuggingFace Hub integration
distributed	Yes	pmetal-distributed	Multi-Mac discovery, all-reduce, pipeline harnesses
cli	Yes	pmetal	Command-line binary and shared CLI dependencies
dashboard	Yes	—	Ratatui terminal TUI
trainer	Yes	pmetal-trainer	SFT, LoRA/QLoRA, pretraining, GRPO, RLKD, distillation orchestration
lora	Yes	pmetal-lora	LoRA/QLoRA adapters
merge	Yes	pmetal-merge	Model merging strategies and save pipeline
ane	Yes	pmetal-metal, pmetal-trainer, pmetal-models	Apple Neural Engine integration
mlx	Yes	pmetal-mlx	MLX backend
models	Yes	pmetal-models	LLM architectures
data	Yes	pmetal-data	Dataset loading, tokenization, chat templates
distill	Yes*	pmetal-distill	Knowledge distillation (*via trainer)
lora-metal-fused	No	—	~2× LoRA training speedup via fused Metal kernels
native-only	No	pmetal-bridge	Bridge-only native runtime without the full MLX feature set
vocoder	No	pmetal-vocoder	BigVGAN neural vocoder
mhc	No	pmetal-mhc	Manifold-Constrained Hyper-Connections
serve	No	pmetal-serve	OpenAI-compatible inference server
mcp	No	pmetal-mcp	MCP server tools
full	No	—	All library feature groups; add serve/mcp separately for those CLI surfaces

Usage

# Default features
cargo install pmetal

# With specific features
cargo install pmetal --features "serve,mcp"

# All library feature groups
cargo install pmetal --features full

# Library feature groups plus optional CLI surfaces
cargo install pmetal --features "full,serve,mcp"

# Minimal CLI build
cargo build --release --no-default-features --features cli

# TUI without default trainer/ANE/distributed stack
cargo build --release --no-default-features --features dashboard

As a Library Dependency

[dependencies]
pmetal = { version = "0.5", default-features = false, features = ["core", "models"] }

# Or specific crates
pmetal-models = "0.5"
pmetal-trainer = "0.5"
pmetal-lora = "0.5"

See Also

• Installation — Build options
• Advanced SDK Usage — Crate-level API