WSL2(Ubuntu 26.04) + Docker で Radeon GPU を認識させるThis article explains how to configure WSL2 (Ubuntu 26.04) and Docker to…

Running large language models locally on AMD Radeon GPUs through WSL2 has long been more complicated than the equivalent NVIDIA workflow, and this article walks through the steps needed to get Docker inside WSL2 on Ubuntu 26.04 to recognize a Radeon card. For developers who want to experiment with local inference without buying NVIDIA hardware, or who already own a capable Radeon GPU, this configuration matters because it removes one of the main friction points that has historically pushed the local LLM community toward CUDA.

The core challenge stems from how GPU access works across the Windows and Linux boundary. WSL2 runs a real Linux kernel inside a lightweight virtual machine, and GPU compute is exposed through a paravirtualized device rather than a directly attached PCIe card. On the NVIDIA side, this has been handled for some time via the CUDA on WSL driver and the NVIDIA Container Toolkit. AMD's equivalent path relies on ROCm, AMD's open compute stack, together with driver support that surfaces the GPU to the WSL2 guest. The article's premise is that this combination has matured to the point where a documented, repeatable procedure is feasible on a recent Ubuntu release.

The general shape of the setup involves several layers that must line up. First, a current AMD Adrenalin driver on the Windows host is required, since the host driver is what ultimately arbitrates access to the physical GPU. Next, WSL2 itself must be recent enough to expose the necessary device nodes, typically appearing as /dev/dxg for the DirectX-based passthrough and, for ROCm workloads, the kernel fusion driver device such as /dev/kfd along with /dev/dri render nodes. Inside Ubuntu, the ROCm user-space packages provide the runtime libraries and tools like rocminfo, which is commonly used to confirm that the GPU is visible before layering Docker on top.

Docker adds another step because containers by default do not inherit host device access. The article appears to cover passing the relevant devices into the container, which in practice means granting access to /dev/kfd and /dev/dri and adding the container user to the appropriate video and render groups. Some ROCm container images and run commands also rely on the SYS_PTRACE capability or a loosened security profile, and matching the ROCm version inside the container to the host stack is important to avoid subtle incompatibilities. Getting these permissions right is often where setups fail, so verifying with rocminfo or a small PyTorch check inside the container is a sensible sanity test.

It is worth noting the version specifics. Ubuntu 26.04 would be the release following the 24.04 and 25.10 lines, and using a very new distribution can be a double-edged situation. Newer kernels and packages may improve WSL2 GPU support, but ROCm's official support matrix tends to certify a specific list of Ubuntu versions and GPU models at any given time. Readers should confirm that their particular Radeon card and their ROCm release are on AMD's supported list, since consumer RDNA GPUs have variable levels of official coverage, and some cards require environment variables such as HSA_OVERRIDE_GFX_VERSION to run at all. Because of this, results are likely to vary between hardware configurations, and the exact commands may need adjustment as ROCm and WSL2 evolve.

For context, this fits into a broader trend of AMD trying to close the software gap with NVIDIA in the AI space. ROCm now offers builds of PyTorch and increasingly supports popular inference tooling, and projects like llama.cpp support AMD acceleration through both ROCm and the more portable Vulkan backend. The Vulkan route is sometimes simpler for casual local LLM use because it avoids the full ROCm installation, though ROCm generally provides better performance and broader framework compatibility for serious workloads. Ollama and text-generation-webui are other common front ends that developers may layer on once GPU access is confirmed.

Overall, the article serves a practical niche: it documents a workflow that is technically supported but not yet as turnkey as the NVIDIA path. Anyone following it should treat the procedure as a snapshot tied to specific driver, kernel, and ROCm versions, and should expect to consult AMD's current documentation for their exact card. For those willing to work through the layered permissions and version matching, it offers a viable route to running local LLMs on Radeon hardware inside a familiar Docker and WSL2 environment.