Skip to content

[WSL] rocm install driver points at an unshipped repo script instead of installing ROCDXG #223

Closed

Description

@rominf

Summary

On WSL, rocm install driver refuses to do anything and tells the user to run a
script that is not shipped with the CLI:

$ rocm install driver
driver install plan
  policy: wsl_rocdxg
  supported: false
  mutating: false
  reason: WSL uses the Windows host driver plus ROCDXG; run `scripts/wsl_setup_rocdxg.sh` inside WSL instead of installing Linux DKMS.
  execution_commands: <none>
  action: no driver commands will be executed

scripts/wsl_setup_rocdxg.sh exists only in a git checkout. The release bundle
contains bin/rocm, bin/rocmd, README.md, LICENSE.TXT and the platform
installer — nothing else — so anyone who installed rocm-cli the normal way is
pointed at a path they do not have.

Why it matters

ROCDXG (librocdxg) is the ROCm component that lets the runtime reach the GPU
under WSL2. Without it the GPU is visible to the distro but unusable, and the
CLI already knows this:

driver_status: wsl_rocdxg_missing
driver_detail: missing /opt/rocm/lib/librocdxg.so, ldconfig:librocdxg.so
detected_gfx_target: gfx1151

rocm serve then refuses with "no usable AMD GPU detected". So on the one
platform where a ROCm-provided library is the difference between a working and a
non-working GPU, the CLI declines to install it — while on native Linux
rocm install driver builds and executes a real plan.

rocm diagnose does not close the gap either: on WSL it declares the bare-metal
catalog out of scope and returns a docs URL, with no entry for
wsl_rocdxg_missing. That leaves the user with a diagnosis they cannot act on.

Expected

rocm install driver performs the ROCDXG installation on WSL, subject to the
same approval and dry-run semantics as the native Linux driver path, instead of
requiring the user to clone the repository.

Notes

  • Reported from a real user session; also hit on the self-hosted WSL2 CI runner.
  • The script already does the right shape of work — preconditions (/dev/dxg,
    libdxcore.so), download a versioned artifact from a GitHub release, verify
    SHA-256, install, verify the result — which is the same shape as the CLI's
    existing managed-artifact install paths.
  • Independently of the larger question, the current guidance naming an unshipped
    path is wrong and worth correcting on its own.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions