This short guide covers creating a basic development environment, installing needed dependencies, and best practices for building and running Triton on different Intel platforms.
The primary software dependency is oneAPI. The best way to get oneAPI for PyTorch / Triton development is to use the Intel PyTorch Dependency Bundle. Scroll down to "Step 2: Install Intel Support Packages" and choose an appropriate option for your system.
- Option 2A-2C: Install using System Package Manager is recommended for system-wide installs of the currently released packages with automatic updates.
- Option 2D: Offline Install is recommended for more control over package versions, or to have multiple installations of different versions of the oneAPI dependency bundle on the same machine.
Intel GPU drivers and development packages are required. Several release streams are available. Ubuntu 22.04 release stream information is linked below. Other release stream information is available at the same link.
The LTS, Production, and Rolling Stable release streams are supported. Some performance optimizations may not be available with the LTS driver package. For development, we recommend rolling stable.
The Rolling Stable release stream is supported. Other release streams may not work for all test cases.
- Building Triton with
compile-triton.shcan be memory intensive due to compiling several large dependencies from source. Use the environment variableMAX_JOBS=8on machines with < 64GB of memory. - Pre-built wheels will be fetched during invocation of the
test-triton.shscript if the GitHub CLI toolghis installed. The default version on Ubuntu 22.04 is too old to support some flags we use, so install the latest from GitHub APT repo using these instructions.