Skip to content

Latest commit

 

History

History
67 lines (45 loc) · 2.01 KB

File metadata and controls

67 lines (45 loc) · 2.01 KB

Install MTPLX

MTPLX is production software for Apple Silicon Macs, distributed via pip, Homebrew, and a signed DMG.

Requirements

  • Apple Silicon Mac
  • Python 3.11+
  • macOS with MLX support
  • Enough disk for the selected model

Install

Recommended macOS install:

curl -fsSL https://raw.githubusercontent.com/youssofal/MTPLX/main/scripts/install_macos.sh | bash
mtplx help

The installer checks Homebrew Python paths directly, so it works even if a fresh Terminal tab has not put /opt/homebrew/bin on PATH yet. It installs MTPLX from PyPI into ~/.mtplx/venv and writes a durable launcher at ~/.local/bin/mtplx. On Apple Silicon Homebrew installs, it also writes /opt/homebrew/bin/mtplx when that directory is writable.

Python-only install:

python3 -m pip install -U mtplx
mtplx help

For local development:

python -m pip install -e ".[dev,server]"

Runtime Dependencies

mtplx --help, mtplx doctor, mtplx inspect, and mtplx init are designed to work even before MLX is installed. Generation and serving require MLX and a verified model.

MTPLX runs on stock PyPI MLX; no fork is required for any profile (the legacy --strict-mlx-fork-assert flag is a deprecated no-op).

Optional Thermal Tools

--max is opt-in. It is for users who need sustained throughput and accept fan noise. It is never part of the default quick start and is never used for no-fan product claims.

Check the local thermal-control state:

mtplx max --status

If ThermalForge or TG Pro is not present, MTPLX prints install instructions and continues without fan control for run, chat, and serve --max. It must not silently enable spin-loop or clock-anchor modes.

Supported public commands:

mtplx max --on       # Performance profile
mtplx max --max      # Max profile
mtplx max --off      # Silent profile
mtplx max --status   # tool/status report

MTPLX_GPU_CLOCK_ANCHOR=1 is an explicit experimental diagnostic only. Do not use it for README, release, or product benchmark claims.