LykosAI/StabilityMatrix · error · InvalidOperationException

Installed torch is not a usable Windows ROCm build

Error message

Installed torch is not a usable Windows ROCm build (detected version: {torchInfo.Version}).

What it means

EnsureWindowsSdkDevelAsync checks the installed torch version via IsUsableWindowsNativeTorchBuild, which requires the version string to indicate a ROCm build (contains 'rocm') or report a hip version. If the installed torch is a CPU or CUDA build (or has an unparseable version), the method throws InvalidOperationException with the detected version so the user knows the build is unusable for Windows ROCm.

Solutions

  1. Uninstall the existing torch (pip uninstall torch) and reinstall the Windows ROCm build from the AMD ROCm package index
  2. Run the helper's InstallWindowsNativeTorchAsync flow, which installs the correct ROCm-indexed wheel
  3. Check torch.__version__ in the venv - it should contain 'rocm' (e.g. 2.x.x.devYYYYMMDD+rocm...); if not, the wrong wheel is installed
  4. Ensure any pip constraints/extra-index-url entries are not pulling torch from PyPI instead of the ROCm index

Example fix

// before
// venv has torch 2.4.0+cpu installed
await rocmHelper.EnsureWindowsSdkDevelAsync(venvRunner, ...);
// after
await venvRunner.PipUninstall("torch");
await rocmHelper.InstallWindowsNativeTorchAsync(venvRunner, ...); // installs ROCm wheel
await rocmHelper.EnsureWindowsSdkDevelAsync(venvRunner, ...);
Defensive patterns

Strategy: validation

Validate before calling

var torchInfo = await venvRunner.PipShow("torch");
bool usable = torchInfo?.Version is string v &&
    (v.Contains("rocm", StringComparison.OrdinalIgnoreCase));
if (!usable) { /* reinstall from the ROCm index before proceeding */ }

Type guard

bool IsRocmBuild(string? version) =>
    !string.IsNullOrWhiteSpace(version) &&
    version.Contains("rocm", StringComparison.OrdinalIgnoreCase);

Try / catch

try { await rocmHelper.EnsureWindowsSdkDevelAsync(venvRunner, ...); }
catch (InvalidOperationException ex) when (ex.Message.Contains("not a usable Windows ROCm build"))
{
    await venvRunner.PipUninstall("torch");
    await rocmHelper.InstallWindowsNativeTorchAsync(venvRunner, ...);
}

Prevention

When it happens

Trigger: Calling EnsureWindowsSdkDevelAsync in a venv where torch is installed but is a CPU-only or CUDA Windows wheel (version string lacks 'rocm' and no hip metadata) - e.g. torch was installed from PyPI defaults instead of the ROCm index, or an earlier CUDA-based install was never replaced.

Common situations: User pre-installed torch with plain 'pip install torch' (CUDA/CPU wheel) before running ROCm setup; switched GPU vendor flows and reused an old environment; nightly ROCm wheel renamed so version no longer contains 'rocm'.

Understand the failure class

Background: "is not a compatible type" / "cannot merge" errors: when a value's type doesn't match what the library requires — this error's family across 65 libraries.

Related errors

AI-assisted analysis of LykosAI/StabilityMatrix@af93d6ef57 (2026-09-12). Data as JSON: /api/errors/ecb113953b9e2388. Report an issue: GitHub.

Appendix: source

Thrown at StabilityMatrix.Core/Services/Rocm/RocmPackageHelper.cs:149

        CancellationToken cancellationToken = default
    )
    {
        var state = machineState.Value;
        var multiArchPythonPackageIndexUrl = WindowsRocmSupport.GetMultiArchPythonPackageIndexUrl(
            state.RuntimeGfxArch
        );

        var torchInfo = await venvRunner.PipShow("torch").ConfigureAwait(false);
        if (torchInfo is null)
        {
            throw new InvalidOperationException(
                "torch is not installed in this environment. Install the Windows ROCm torch build first."
            );
        }

        if (!IsUsableWindowsNativeTorchBuild(torchInfo.Version, null))
        {
            throw new InvalidOperationException(
                $"Installed torch is not a usable Windows ROCm build (detected version: {torchInfo.Version})."
            );
        }

        var nightlyBuildDateToken = TryGetNightlyBuildDateToken(torchInfo.Version);
        var installedRocmSdkDevel = await venvRunner.PipShow(RocmSdkDevelPackageName).ConfigureAwait(false);
        if (
            !string.IsNullOrWhiteSpace(nightlyBuildDateToken)
            && HasNightlyBuildDateToken(installedRocmSdkDevel?.Version, nightlyBuildDateToken)
        )
        {
            return;
        }

        var indexResult = await venvRunner
            .PipIndex(RocmSdkDevelPackageName, multiArchPythonPackageIndexUrl)
            .ConfigureAwait(false);

View on GitHub (pinned to af93d6ef57)