RockBuilder

RockBuilder is a configuration-based build system that simplifies the process of integrating and building one or multiple AI applications on top of AMD’s ROCm SDK.

TheRock provides the official ROCm images, while RockBuilder supports building additional applications on top of the ROCm SDK that are not tightly integrated with TheRock. RockBuilder is mainly targeted at developers and users who want to add support for building or using these additional applications.

Note: RockBuilder currently supports all stable ROCm versions from TheRock where applicable.

RockBuilder can use:

• An existing ROCm SDK installation.

• A new ROCm SDK, either:

  • Built from source, or

  • Installed from Python wheels.

RockBuilder supports both Linux and Windows for building the applications.

Note: Windows support is still less tested compared to Linux. You may encounter issues that haven’t been fully validated yet.

Download RockBuilder

git clone https://github.com/roCm/rockbuilder

ROCm SDK and Target GPU Configuration

Initialize and activate the Python virtual environment with all Python dependencies required by RockBuilder:

• Linux (Ubuntu 24.04):

  cd rockbuilder
  source ./init_rcb_env.sh
  

• Windows (x86_64, Visual Studio Command Prompt):

  cd rockbuilder
  init_rcb_env.bat
  

The init_rcb_env script checks if the Python virtual environment is active. If not, it will initialize and activate one in .venv-directory and install all the Python packages required by RockBuilder.

After the Python virtual environment is activated, you can start RockBuilder:

python rockbuilder.py

When you run RockBuilder for the first time, it checks whether a rockbuilder.cfg configuration file exists to decide if it needs to build or install the ROCm SDK before proceeding to build applications:

• If the file exists, RockBuilder automatically selects a stable, tested version of the ROCm SDK as the base environment, and you can proceed directly to building your applications.

• If the file does not exist, you will be prompted to choose how to install the ROCm SDK and specify your target GPUs:

  • New ROCm SDK Build: Lists all supported AMD GPUs.

    

  • ROCm SDK from Python Wheels: Lists only GPUs with available wheel packages (some GPUs may share the same package).

    

Once confirmed (by pressing Enter), your selections will be saved to rockbuilder.cfg.

Note: Building the ROCm SDK allows for greater customization, but the process can take anywhere from one to several hours, depending on your system.

Build an Application Set

In many cases, multiple applications need to be built to achieve full functionality. RockBuilder handles this by listing all related applications and their corresponding versions in an .apps file.

Example usage to build PyTorch nightly and its dependencies:

./rockbuilder.py apps/pytorch_nightly.apps

This will download, configure, and build all applications that are specified in the pytorch_nightly.apps file.

[apps]
app_list=
    deps_common
    pytorch_aotriton_nightly
    triton_pytorch_nightly
    pytorch_nightly
    pytorch_vision_nightly
    pytorch_torchcodec_nightly
    pytorch_audio_nightly

Applications are built and installed in the listed order above. Each application will be installed into the currently active Python virtual environment. Any additional libraries or executables built by CMake will be installed to the configured ROCm SDK.

Built Python wheels will be copied to the packages/wheels directory in RockBuilder.

Build Applications One By One

Instead of building a set of applications, you can also build them one by one in the correct dependency order.

./rockbuilder.py apps/deps_common.cfg
./rockbuilder.py apps/pytorch_aotriton_nightly.cfg
./rockbuilder.py apps/triton_pytorch_nightly.cfg
./rockbuilder.py apps/pytorch_nightly.cfg
./rockbuilder.py apps/pytorch_vision_nightly.cfg
./rockbuilder.py apps/pytorch_torchcodec_nightly.cfg
./rockbuilder.py apps/pytorch_audio_nightly.cfg

Each of these .cfg files provides an application-specific configuration. These files define the application name, version, source repository, and the commands required to configure, build, and install the application.

Configuration file format is specified in CONFIG.md.

Test the Applications Build

RockBuilder includes simple example applications to verify that the PyTorch build was successful. If you are running the tests from a new terminal window, you’ll need to activate the Python virtual environment first. If it’s already active, you can skip this step:

source ./init_rcb_env.sh 

Then, run the example application itself with the following command:

python examples/torch_gpu_hello_world.py 

If successful, the application should print output similar to the following in the terminal:

Pytorch version: 2.8.0
ROCM HIP version: 7.1.25441-b9b1250
cuda device count: 2
default cuda device name: AMD Radeon Pro W7900 Dual Slot
device type: cuda
Tensor training running on cuda: True
Running simple model training test
    tensor([0., 1., 2.], device='cuda:0')
Hello World, test executed succesfully

Example output from the test applications on Windows using the AMD Radeon W7900 GPU:

(.venv) D:\rockbuilder\examples>python torch_gpu_hello_world.py
Pytorch version: 2.7.0
ROCM HIP version: 6.5.25222-1f8e4aaca
cuda device count: 1
default cuda device name: AMD Radeon PRO W7900 Dual Slot
device type: cuda
Tensor training running on cuda: True
Running simple model training test
    tensor([0., 1., 2.], device='cuda:0')
Hello World, test executed succesfully

(.venv) D:\rockbuilder\examples>python torch_vision_hello_world.py
pytorch version: 2.7.0
pytorch vision version: 0.22.0

(.venv) D:\rockbuilder\examples>python torch_audio_hello_world.py
pytorch version: 2.7.0
pytorch audio version: 2.7.0

You can also test the flash-attention support in PyTorch with the following example application:

python examples/torch_attention_check.py

Other RockBuilder Usage Examples

RockBuilder also supports optional build arguments as follows:

Checkout Only the Source Code

This command checks out the source code for the PyTorch 2.8–related applications without building them. The source code will be checked out to the src_apps directory.

python rockbuilder.py --checkout apps/pytorch_28_amd.apps

Checkout Source Code to a Custom Directory

This command checks out the source code for each project to the custom_src_location directory instead of the default src_apps directory.

python rockbuilder.py --checkout --src-base-dir custom_src_location apps/pytorch_28_amd.apps

Build and Install Python Wheel to a Custom Directory

This command builds and installs only PyTorch Audio and copies the produced PyTorch Audio wheel to the test directory instead of the default packages/wheels directory.

Note: PyTorch Audio requires PyTorch to be built and installed first.

python rockbuilder.py apps/pytorch_audio.cfg --output-dir test

Checkout the Source Code of a Single Application to a Custom Directory

This command checks out the source code of a single application to the src_prj/py_audio directory.

python rockbuilder.py --checkout apps/pytorch_audio.cfg --src-dir src_prj/py_audio

Checkout a Custom Version

This command checks out the source code of PyTorch Audio version 2.6.0 instead of the version specified in the pytorch_audio.cfg file.

python rockbuilder.py --checkout pytorch_audio --pytorch_audio-version=v2.6.0

Execute Only the Install Phase

This command executes only the install phase for a previously built PyTorch Audio.

Note: PyTorch Audio requires PyTorch to be built and installed first.

python rockbuilder.py --install apps/pytorch_audio.cfg

Add a New Application to RockBuilder

RockBuilder uses two types of configuration files stored under the applications directory.

Application Set Configuration File

apps/core.apps is an example of an application set configuration file, listing applications that RockBuilder can build:

[apps]
app_list=
    pytorch
    pytorch_vision
    pytorch_audio

Application Configuration File

apps/pytorch.cfg is an example of an application configuration file, defining the actions that RockBuilder executes for a specific project, including:

• inut
• checkout
• clean
• pre-configure
• configure
• post-configure
• build
• install
• post-install

By default the RockBuilder executes init, checkout, pre-configure, configure, post-configure, build, install, and post-install phases for the application. You can override this by specifying the desired command phase. For example:

python rockbuilder.py --checkout apps/pytorch.cfg