MTPLX

https://github.com/youssofal/mtplx
Admin
MTPLX is a native MTP speculative decoding runtime for Apple Silicon that achieves ~2.24× faster LLM...
Tokens:268,159
Snippets:2,144
Trust Score:8.2
License:Apache-2.0
Update:3 weeks ago
Show doc for...
Tokens:
Raw
### Launch MTPLX Live Dashboard with mtplx start, dashboard, or serve

Source: https://github.com/youssofal/mtplx/blob/main/docs/dashboard.md

Quickstart commands for launching the MTPLX Live Dashboard. Use `mtplx start` to pick the dashboard during setup, `mtplx dashboard` if a server is already running, or `mtplx serve --port 8000` followed by `mtplx dashboard` for full control. The dashboard opens at http://127.0.0.1:8000/dashboard.

```bash
# Easiest: pick "Live Dashboard" at step 3 of `mtplx start`.
mtplx start

# Or, if a server is already running:
mtplx dashboard

# Or, with full control:
mtplx serve --port 8000  # any client (Web UI, OpenAI SDK, hippo) drives load
mtplx dashboard          # opens http://127.0.0.1:8000/dashboard in your browser
```

--------------------------------

### Install mtplx CLI with Homebrew and start it

Source: https://github.com/youssofal/mtplx/blob/main/README.md

Use this to install the mtplx command-line tool via Homebrew and then start it. This is the primary CLI installation method.

```bash
brew install youssofal/mtplx/mtplx
mtplx start
```

--------------------------------

### Setup and verify development environment

Source: https://github.com/youssofal/mtplx/blob/main/CONTRIBUTING.md

Commands to install development dependencies, run essential tests, build the package, and perform a smoke test in a fresh virtual environment.

```bash
python -m pip install -e ".[dev,server]"
python -m pytest tests/test_no_mlx_imports.py tests/test_public_cli.py tests/test_runtime_kpis.py
python -m build
scripts/fresh_venv_smoke.sh
```

--------------------------------

### Install MTPLX with Homebrew and run basic commands

Source: https://github.com/youssofal/mtplx/blob/main/docs/quickstart.md

Install MTPLX via Homebrew on macOS (the recommended path) and run basic commands: help, a summary doctor check, and start. On M3/M4/M5, the app and CLI recommend specific quantized models based on memory (e.g., Ternary Bonsai 2 27B from 16 GB, Qwen 3.8 Flash-Next Optimized Speed from 256 GB), while M1/M2 keep the FP16 policy; an explicit model selection always wins.

```bash
brew install youssofal/mtplx/mtplx

mtplx help
mtplx doctor --summary
mtplx start

```

--------------------------------

### Run tests and build/launch app bundle

Source: https://github.com/youssofal/mtplx/blob/main/apps/MTPLXApp/README.md

Run the test suite and then build and launch the app bundle with verification. The run script creates a small .app bundle under dist/ and launches that bundle instead of running the SwiftUI executable as a raw command-line process.

```sh
swift test
```

```sh
./script/build_and_run.sh --verify
```

--------------------------------

### Start MTPLX server, CLI, and quickstart with options

Source: https://github.com/youssofal/mtplx/blob/main/docs/quickstart.md

Run MTPLX after a verified model is available. `--no-mtp` switches generation to target-only AR; for MTP-equipped models the MTP runtime stays loaded, enabling `/mtp off`, `/mtp on`, and `/mtp status` in terminal chat without reloading. Native AR-only models install an unloaded AR route instead.

```bash
mtplx start
mtplx start cli
mtplx start cli --no-mtp
mtplx quickstart --port 8000 --no-stats-footer

```

--------------------------------

### Install MTPLX and manage thermal profiles

Source: https://github.com/youssofal/mtplx/blob/main/INSTALL.md

These commands cover the recommended macOS install, a Python-only install, a local development install, checking thermal-control state, and the public `mtplx max` subcommands. The installer avoids writing into Homebrew directories and works even if `/opt/homebrew/bin` is not on PATH. The thermal `--max` mode is opt-in and never part of the default quick start.

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

```bash
python3 -m pip install -U mtplx
mtplx help
```

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

```bash
mtplx max --status
```

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

--------------------------------

### Run diagnostic check with mtplx doctor --json

Source: https://github.com/youssofal/mtplx/blob/main/TROUBLESHOOTING.md

Start troubleshooting by running the doctor command to get a JSON health report. Missing MLX should be reported as an actionable runtime dependency issue, not a traceback.

```bash
mtplx doctor --json
```

--------------------------------

### Start MTPLX server bound to all interfaces with API key file

Source: https://github.com/youssofal/mtplx/blob/main/docs/server.md

Use this command to make MTPLX reachable from other devices or VMs. Requires an API key; if the key file doesn't exist, it is created with a fresh key and printed once. Startup prints a Network OpenAI API Base URL with the Mac's LAN address.

```bash
mtplx serve --host 0.0.0.0 --port 8000 --api-key-file ~/.mtplx/api-key
```

--------------------------------

### Build dashboard from source with bun

Source: https://github.com/youssofal/mtplx/blob/main/docs/dashboard.md

Run these commands in the dashboard directory to install dependencies and build the static assets. The output goes to ../mtplx/dashboard/_static/ and is bundled into the wheel via pyproject.toml package-data.

```bash
cd dashboard
bun install
bun run build      # outputs into ../mtplx/dashboard/_static/

```
Tokens:
Raw