SyzygyDocs

Installation

Install mach from PyPI, add optional speculative decoding, and verify the setup.

Install from PyPI

Requires an Apple Silicon Mac, macOS, and Python 3.12+.

pip install mach-engine

This installs the engine and its console scripts:

CommandRole
mach-serveOpenAI- and Anthropic-compatible HTTP server
mach-generateOne-shot CLI generation
mach-checkPreflight: is a checkpoint servable, and does it fit in memory?
mach-modelsList locally cached checkpoints
mach-archsPrint the supported-architecture registry

mlx is the only framework dependency; mlx-lm is not required.

Optional: speculative decoding

Speculative decoding (~163 tok/s vs ~117 tok/s single-stream) needs the optional dflash extra and a local draft checkpoint:

pip install "mach-engine[dflash]"

Pass the draft with mach-serve --draft-dir. Without the extra or a usable draft, the server runs the exact decode path with no configuration needed. See Serving.

Install from source

git clone https://github.com/SyzygyResearch/mach-engine
cd mach-engine
pip install -e "."

Verify

mach-check SyzygyResearch/Mach-1-Additive-35B

mach-check reports, without downloading weights, whether the pack is supported, which serving paths it can run, and whether it fits your machine's memory (fits / tight / wont_fit), then prints the exact serve command.

Model downloads are cached under ~/.cache/mach (override with --cache-dir on any command).

Next steps

  • Serving: start the server and connect a client
  • CLI: generation and inspection commands

On this page