wslc preview: Missing --device flag and /dev/kfd CDI mapping breaks AMD ROCm GPU passthrough

Windows Version

Microsoft Windows [Version 10.0.26200.8655]

WSL Version

WSL-Version: 2.9.3.0

Are you using WSL 1 or WSL 2?

• WSL 2
• WSL 1

Kernel Version

wslc backend VM kernel

Distro Version

alpine:latest (probe) and yanwk/comfyui-boot:rocm7

Other Software

Hardware: AMD Radeon 9070XT

Framework: AMD ROCm 7.2

Workload: ComfyUI / PyTorch (ROCm build)

Repro Steps

Attempt to run a ROCm-based container via the new wslc public preview using an AMD GPU.

Run the command using the --gpus all flag to map the hardware:

wslc run -d --name=comfyui-amd --gpus all -e HSA_OVERRIDE_GFX_VERSION=12.0.0 -p 8188:8188 yanwk/comfyui-boot:rocm7

Check container logs (wslc logs comfyui-amd) and observe the PyTorch crash: RuntimeError: No HIP GPUs are available.

To verify the hardware mount in the container namespace, run a basic probe:

wslc run --rm alpine ls -la /dev

Observe that /dev/dxg may be present (if DirectML is mapped), but /dev/kfd (Kernel Fusion Driver) and /dev/dri are completely missing from the container's device namespace.

Attempt to bypass this by manually volume mounting the devices (-v /dev/kfd:/dev/kfd), which fails due to standard Linux cgroup permissions.

Expected Behavior

The --gpus all CDI specification should map raw AMD compute devices (/dev/kfd and /dev/dri) into the container namespace alongside /dev/dxg so ROCm workloads can run natively.

Alternatively, wslc should support the standard Docker/Podman --device and --group-add flags to allow users to manually pass through AMD hardware with the correct cgroup permissions.

Actual Behavior

The wslc container namespace fundamentally lacks access to /dev/kfd and /dev/dri. Because --gpus all seems strictly tied to NVIDIA or DirectML mapping, and because the CLI omits the --device flag entirely, AMD users are currently locked out of running native ROCm ML workloads in WSL Containers.

Diagnostic Logs

Output of wslc run --rm alpine ls -la /dev proving the AMD compute devices are missing from the container namespace:

total 4
drwxr-xr-x    5 root     root           340 Jul  3 14:28 .
drwxr-xr-x    1 root     root          4096 Jul  3 14:28 ..
lrwxrwxrwx    1 root     root            11 Jul  3 14:28 core -> /proc/kcore
lrwxrwxrwx    1 root     root            13 Jul  3 14:28 fd -> /proc/self/fd
crw-rw-rw-    1 root     root        1,   7 Jul  3 14:28 full
drwxrwxrwt    2 root     root            40 Jul  3 14:28 mqueue
crw-rw-rw-    1 root     root        1,   3 Jul  3 14:28 null
lrwxrwxrwx    1 root     root             8 Jul  3 14:28 ptmx -> pts/ptmx
drwxr-xr-x    2 root     root             0 Jul  3 14:28 pts
crw-rw-rw-    1 root     root        1,   8 Jul  3 14:28 random
drwxrwxrwt    2 root     root            40 Jul  3 14:28 shm
lrwxrwxrwx    1 root     root            15 Jul  3 14:28 stderr -> /proc/self/fd/2
lrwxrwxrwx    1 root     root            15 Jul  3 14:28 stdin -> /proc/self/fd/0
lrwxrwxrwx    1 root     root            15 Jul  3 14:28 stdout -> /proc/self/fd/1
crw-rw-rw-    1 root     root        5,   0 Jul  3 14:28 tty
crw-rw-rw-    1 root     root        1,   9 Jul  3 14:28 urandom
crw-rw-rw-    1 root     root        1,   5 Jul  3 14:28 zero