Skip to content

install.md

Latest commit

 

History

History
129 lines (96 loc) · 3.94 KB

install.md

File metadata and controls

129 lines (96 loc) · 3.94 KB
Raw

Installing ARLE

ARLE ships pre-built binaries on every v* tag. Three install paths cover the supported platforms; pick the one that matches your environment.

Support matrix

Platform Backend Binaries shipped
macOS arm64 (Apple Silicon) Metal / MLX arle
Linux x86_64 CUDA 12.x (driver required on host) arle
Other (macOS x86_64, Linux aarch64, Windows) — Build from source

CUDA binaries are linked against cudart 12.x and need a matching NVIDIA driver / CUDA runtime present on the host. Metal binaries need macOS 14+.

1. Homebrew (macOS arm64)

brew install cklxx/tap/arle
arle --doctor

Tap source: https://github.com/cklxx/homebrew-tap. The formula is bumped automatically on every v* tag from this repo's release workflow.

To upgrade:

brew update && brew upgrade arle

To uninstall:

brew uninstall arle
brew untap cklxx/tap   # optional: remove the tap entirely

2. One-line installer (macOS arm64 / Linux x86_64)

curl -fsSL https://github.com/cklxx/arle/releases/latest/download/install.sh | sh

What it does:

  1. Detects platform via uname -s / uname -m.
  2. Resolves the latest tag through GitHub's redirect.
  3. Downloads arle-<tag>-<platform>.tar.gz and SHA256SUMS.txt.
  4. Verifies the SHA256 (uses sha256sum or shasum -a 256).
  5. Extracts and install -m 0755s the binaries into $INSTALL_DIR (default ~/.local/bin).
  6. Prints a PATH hint if $INSTALL_DIR is not on PATH.

Environment overrides

Variable Default Effect
ARLE_VERSION latest Pin to a specific tag, e.g. v0.1.0.
INSTALL_DIR $HOME/.local/bin Where binaries land. Use /usr/local/bin for system-wide (needs sudo).
ARLE_NO_VERIFY unset If set, skip SHA256 verification (not recommended).

Examples:

# Pin a version, system-wide:
curl -fsSL https://github.com/cklxx/arle/releases/download/v0.1.0/install.sh \
  | INSTALL_DIR=/usr/local/bin sudo sh

# Inspect the script before running:
curl -fsSL https://github.com/cklxx/arle/releases/latest/download/install.sh -o install.sh
less install.sh
sh install.sh

To uninstall, just delete the binaries:

rm -f ~/.local/bin/arle

3. Docker (Linux + NVIDIA)

docker run --rm --gpus all -p 8000:8000 \
  -v /path/to/Qwen3.5-4B:/model:ro \
  ghcr.io/cklxx/arle:latest \
  serve --backend cuda --model-path /model --port 8000

The :latest tag tracks the newest non-prerelease release image. Tagged releases are published as ghcr.io/cklxx/arle:X.Y.Z (no v prefix - docker metadata-action strips it).

4. From source

Required for the cpu backend, CUDA/TileLang, or hacking on the runtime. See the README Quick Start for the canonical cargo build invocations per backend, and environment.md for the env-var knobs that affect the build.

CUDA source builds need TileLang. TileLang is the one build-time Python dependency (attention / GDR kernel AOT codegen), pinned in requirements-build.txt. The quickest way to set it up — it also installs the native build deps (clang/cmake/bindgen) and downloads a default model — is:

./setup.sh                     # native deps + TileLang venv + model
cargo build --release --features cuda --bin arle

If you prefer to set it up manually, see troubleshooting.md and the INFER_TILELANG_PYTHON entry in environment.md. Docker (§3) needs no compile step.

Verifying an install

arle --doctor          # human-readable
arle --doctor --json   # machine-readable, suitable for CI gates

--doctor prints the compiled backend, runtime feature flags, and a self-check of the model-loading path. If it errors out, the most common causes are documented in troubleshooting.md.