AI with AMD ROCm on Ubuntu: your questions answered

Canonical and AMD™ teamed up last fall to package AMD ROCm™ AI/ML and HPC libraries right into the Ubuntu archive. Our shared goal is to ensure that Ubuntu offers a seamless, out-of-the-box experience for high-performance AI and HPC on AMD hardware. 

Just last month, we launched our first version of ROCm in Ubuntu 26.04 LTS!  This marks the start of a great future. The reaction so far has been great and supportive, but I’d like to cover some questions asked by the community.

What is ROCm?

ROCm is the AMD software stack for AI/ML and HPC acceleration on AMD hardware. It enables high-speed local inference and model training using AMD GPUs, APUs, and CPUs. With these libraries now native to Ubuntu, users can get the most out of their hardware, as well as popular tools like llama.cpp, Lemonade, ComfyUI, and many more with a simple installation.

How do I use ROCm on Ubuntu?

Prior to the inclusion of ROCm directly in the Ubuntu archives, users needed to download an installation script from AMD and run it locally on their machine (amdgpu-install). This script would then download the required components, compile them as necessary for your local machine, and install them as system libraries. It was fairly straightforward, but had a few drawbacks – especially in production deployments:  

1. To upgrade versions or apply security patches, the user needs to uninstall manually, download a new install script, and reinstall the full suite of libraries. It is up to the user to identify that there’s a new version available and whether or not it is compatible with other applications they’re using.
2. ROCm needs to be installed on the host or in the container separately from applications that may want to use it: it can’t be an auto-managed dependency. This is a manual process separate from installing the actual applications you want to use.

As of Ubuntu 26.04 LTS, having ROCm in the Ubuntu archives makes things much more streamlined: 

In many cases, applications can automatically install the specific ROCm libraries they need, and all their dependencies, by simply declaring them as a dependency of their own installation. In these scenarios, no additional user action is required.

For users:  to install the full suite of ROCm libraries on a host or in a container:  

sudo apt install rocm

For developers: to install the subset of libraries and headers needed for ROCm-enabled application development:

sudo apt install rocm-dev

Once installed, your regular system updates (sudo apt upgrade) automatically manage all ROCm updates and security patches. Our internal CI/CD processes verify version compatibility before release, ensuring a stable update path for your environment.

Which versions are supported? 

To ensure maximum stability for this LTS release, we integrated ROCm 7.1.0, which was the most recent version available when we started the initial packaging efforts. Preparing such a wide range of new packages for their initial inclusion in Ubuntu archive takes considerable work and time. While ROCm 7.1.0 is not the latest upstream version, this release establishes a foundation for the future.  Now that the first version is complete, future updates will get easier and faster – allowing us to keep pace in this rapidly evolving environment. By aligning our packaging efforts with the Debian community and upstreaming core components, we have created a blueprint to close the version gap quickly.

Our goals for the future

With ROCm 7.1 now released in Ubuntu 26.04 LTS, we are focusing on delivering newer ROCm versions through Stable Release Updates (SRU). This process allows us to provide support for the latest AMD hardware and new ROCm features without requiring users to wait for a completely new Ubuntu release.

We are currently working on packaging ROCm 7.2, as well as expanding the range of hardware support we build and test for. The main reason for starting here and not jumping directly to the 7.13+ branches is to work out the in-place upgrade complexities for ROCm and ensure that we have everything ready to issue updates cleanly and smoothly.   Because these updates are automatically rolled out when users update their system with apt upgrade (or with Landscape), this merits caution. 

One of the big challenges we face any time we try to update versions of a package in-place is if there are changes to the ABI (“Application Binary Interface”, or essentially the interfaces that programs use to access and use the functionality of the library.) These changes risk breaking applications built for older versions of ROCm. 

By convention, minor and patch version changes (the x and y in v.x.y version numbers) shouldn’t contain any ABI changes that break systems, and should be fully backwards compatible with other releases in the same major version (the first number of the version).  However, with the pace of AI software development, sometimes breaking changes are introduced into minor version changes, meaning we have to find creative ways to solve the challenges that may create. After all, our value is in providing solutions that are rock solid and that you can depend on, and which don’t break installations that were running just fine on older software.

We are currently evaluating this with ROCm 7.13 and later, which includes major restructuring of the libraries in preparation for ROCm 8 (“The Rock”). This will be a significant uplift, and we will choose the release path that provides the best user experience, flexibility, and reliability for the future.

What’s next?

ROCm 7.1 in Ubuntu 26.04 LTS is just the start. We have big plans, and having this foundation available has already resulted in a growing ecosystem of AI applications that support AMD hardware out of the box.

• Short-term:  in-place upgrade to ROCm 7.2.x (already underway.)
• Longer term:  Looking at ROCm 7.13+, ROCm 8, and beyond.

Our intention is to provide an upgrade path in Ubuntu 26.04 LTS and future releases as far as we possibly can. With ROCm 7.10 and beyond (“The Rock”) being a major change with a complete restructuring of the entire suite of libraries, it is possible there will be changes preventing us from going that far with in-place upgrades – but that has yet to be determined.

The wider AI community and ecosystem with AMD hardware

From the very start, we announced our joint intention with AMD to put ROCm into Ubuntu, and also to help upstream that work to Debian itself.  The packaging work we’re doing is structured to make that possible, and we’re proud to support the wider community with these efforts. The AMD team and Debian AI community are leading those efforts already.  

Now that ROCm is easily available in Ubuntu, we’re starting to see the community adopt these packages and release user applications that take advantage of them. Some of this work is being done by Canonical and AMD engineers, and more by interested community members.

One notable example is Lemonade Server (https://lemonade-server.ai/).  This is an open-source project that provides a hardware-accelerated back-end for AI applications using standards-compliant APIs for front-end applications to interact with. It is an essential tool when connecting ROCm to standard front-end APIs.

Our engineers have packaged Lemonade Server and Lemonade Desktop (a front-end you can use as a chat interface and more) in both snap packages and deb packages. They can be installed with either:

Snap packaging (containerized – to eliminate version or library conflicts with other applications):

snap install lemonade-server

snap install lemonade-desktop

Deb packaging (to install directly in your host environment or container):

sudo apt install lemonade-server

sudo apt install lemonade-desktop

Lemonade is similar to Ollama in many ways, but with out-of-the-box support for AMD hardware – including not only CPU and GPU support, but also NPU support – making it easy to get the most out of your system.  I’m excited about this personally because it’s really the first time we can light up the NPU silicon in AMD Strix Point and Strix Halo systems (Ryzen™ AI Max and Max+ processors), and take full advantage of up to 128GB of shared memory for huge AI models.

Once Lemonade Server is running, it handles all the complex interactions with ROCm and your hardware automatically. This allows you to use a wide range of applications that do not need native ROCm support. By simply pointing these tools to the Lemonade API port, you can perform local AI inference on your AMD hardware using:

• ComfyUI (Image Generation)
• OpenWebUI (Chat Interface)
• OpenClaw (Agentic Tasks)
• OpenCode

Because Lemonade Server provides a standard OpenAI-compatible backend, almost any application designed for standard AI APIs can now leverage the full power of your AMD system without additional configuration.

AMD and Canonical engineers will be at the upcoming Ubuntu Summit 26.04 to talk about Lemonade and ROCm in Ubuntu. Make sure to tune in or watch the replay!

This is just the beginning

Now that we have the ROCm base as an integral part of the Ubuntu ecosystem, we have enabled the full community to build on top of that.  We’ll be working to ensure that AMD hardware is fully supported in our AI applications and roadmap, and can’t wait to see what the community does with it as well.  Let me know how you’re using ROCm in Ubuntu, and maybe I’ll feature you in an upcoming post!

If you’ve got a project you need help with, or simply want to learn more about how we’re growing this important ecosystem, don’t hesitate to contact us.