feat(windows): Native AMD ROCm GPU Acceleration (Resolves #531) (#538)

* feat(windows): add native ROCm support for AMD GPUs

Implements native ROCm architecture for Windows.

- Adds backend build pipeline for voicebox-server-rocm.exe

- Detects AMD GPUs dynamically and routes PyTorch allocations

- Adds automatic download and update logic for ROCm dependencies

- Refactors UI in GpuPage.tsx and GpuAcceleration.tsx to add AMD flows

- Fixes 'Switch to CPU' lock on Windows via Tauri backend_override state

- Resolves PyInstaller/rocm_sdk UnboundLocalError silent crashes

- Resolves Numba/NumPy 2.x incompatibilities during Qwen3-TTS load

- Resolves HF_HUB_OFFLINE Catch-22 for CustomVoice processor caching

* fix(rocm): host libs archive under the app release tag, drop offline-load regression

Align the ROCm libs download with the CUDA pattern: both the server core and
the libs archive are published under the app-version release tag, with the libs
content version encoded in the filename only. The previous code fetched libs
from a separate rocm7.2-v1 tag, which disagreed with the download test.

Also revert the unrelated Qwen CustomVoice changes that wrapped model loading in
force_offline_if_cached (not imported — a NameError on load for every platform)
and re-added a Base-model cache gate. The inference-path offline guard was
deliberately removed previously.

* feat(rocm): gate download on AMD detection and persist the backend variant

The ROCm download section now only shows when the backend reports an AMD GPU on
Windows (new supports_rocm health field, backed by the memoized
is_amd_gpu_windows detection that was previously unused), or when ROCm is already
downloaded/active.

Make the backend override honor a pinned variant: set_backend_override persists
the choice to disk so it survives an app restart, start_server reads it back,
and a cuda/rocm pin now actually selects that variant instead of always
preferring ROCm. A stale pin to a deleted backend self-heals to the default
order rather than forcing CPU. Add the web no-op stub for the new method.

* chore(rocm): drop incomplete vitest harness for the unused GpuAcceleration component

GpuAcceleration.tsx is not routed anywhere (GpuPage is the live settings view),
and the added vitest setup referenced testing-library/vitest deps that were not
in the lockfile, breaking the web typecheck. Remove the dead component's test
and its scaffolding to keep this PR scoped to the ROCm feature.

* ci(rocm): add ROCm release-artifact pipeline

Mirror the CUDA packaging path for ROCm so the runtime download has artifacts to
fetch. scripts/package_rocm.py splits the PyInstaller --rocm onedir into
voicebox-server-rocm.tar.gz (core) + rocm-libs-rocm7.2-v1.tar.gz (AMD runtime:
HIP DLLs, rocBLAS Tensile data, MIOpen kernel DBs) + rocm-libs.json, matching
the names services/rocm.py expects, both under the app-version release tag.

The new build-rocm-windows job in release.yml builds on windows-latest/cp312 and
lets build_binary.py --rocm pull the official AMD Radeon wheels.

The file classifier can't be validated against a real AMD build on CI, so it has
unit coverage (test_package_rocm.py) against a synthetic onedir layout. The
prefixes/dir markers may need a tweak after the first real build on AMD
hardware — the packager hard-fails loudly if it classifies zero ROCm files.

---------

Co-authored-by: Jamie Pine <[email protected]>

app/src/components/ServerSettings/GpuAcceleration.tsx

@@ -5,7 +5,7 @@ import { Button } from '@/components/ui/button';
 import { Card, CardContent, CardHeader, CardTitle } from '@/components/ui/card';
 import { Progress } from '@/components/ui/progress';
 import { apiClient } from '@/lib/api/client';
-import type { CudaDownloadProgress } from '@/lib/api/types';
+import type { CudaDownloadProgress, RocmDownloadProgress } from '@/lib/api/types';
 import { useServerHealth } from '@/lib/hooks/useServer';
 import { usePlatform } from '@/platform/PlatformContext';
 import { useServerStore } from '@/stores/serverStore';
@@ -21,6 +21,9 @@ export function GpuAcceleration() {
   const [restartPhase, setRestartPhase] = useState<RestartPhase>('idle');
   const [error, setError] = useState<string | null>(null);
   const [downloadProgress, setDownloadProgress] = useState<CudaDownloadProgress | null>(null);
+  const [rocmDownloadProgress, setRocmDownloadProgress] = useState<RocmDownloadProgress | null>(
+    null,
+  );
   const healthPollRef = useRef<ReturnType<typeof setInterval> | null>(null);
 
   // Query CUDA backend status
@@ -36,10 +39,26 @@ export function GpuAcceleration() {
     enabled: !!health, // Only fetch when backend is reachable
   });
 
+  // Query ROCm backend status
+  const {
+    data: rocmStatus,
+    isLoading: _rocmStatusLoading,
+    refetch: refetchRocmStatus,
+  } = useQuery({
+    queryKey: ['rocm-status', serverUrl],
+    queryFn: () => apiClient.getRocmStatus(),
+    refetchInterval: (query) => (query.state.status === 'pending' ? false : 10000),
+    retry: 1,
+    enabled: !!health, // Only fetch when backend is reachable
+  });
+
   // Derived state
   const isCurrentlyCuda = health?.backend_variant === 'cuda';
+  const isCurrentlyRocm = health?.backend_variant === 'rocm';
   const cudaAvailable = cudaStatus?.available ?? false;
   const cudaDownloading = cudaStatus?.downloading ?? false;
+  const rocmAvailable = rocmStatus?.available ?? false;
+  const rocmDownloading = rocmStatus?.downloading ?? false;
 
   // Clean up health poll on unmount
   useEffect(() => {
@@ -51,7 +70,7 @@ export function GpuAcceleration() {
     };
   }, []);
 
-  // SSE progress tracking during download
+  // SSE progress tracking during CUDA download
   useEffect(() => {
     if (!cudaDownloading || !serverUrl) {
       return;
@@ -88,6 +107,43 @@ export function GpuAcceleration() {
     };
   }, [cudaDownloading, serverUrl, refetchCudaStatus]);
 
+  // SSE progress tracking during ROCm download
+  useEffect(() => {
+    if (!rocmDownloading || !serverUrl) {
+      return;
+    }
+
+    const eventSource = new EventSource(`${serverUrl}/backend/rocm-progress`);
+
+    eventSource.onmessage = (event) => {
+      try {
+        const data = JSON.parse(event.data) as RocmDownloadProgress;
+        setRocmDownloadProgress(data);
+
+        if (data.status === 'complete') {
+          eventSource.close();
+          setRocmDownloadProgress(null);
+          refetchRocmStatus();
+        } else if (data.status === 'error') {
+          eventSource.close();
+          setError(data.error || 'Download failed');
+          setRocmDownloadProgress(null);
+          refetchRocmStatus();
+        }
+      } catch (e) {
+        console.error('Error parsing ROCm progress event:', e);
+      }
+    };
+
+    eventSource.onerror = () => {
+      eventSource.close();
+    };
+
+    return () => {
+      eventSource.close();
+    };
+  }, [rocmDownloading, serverUrl, refetchRocmStatus]);
+
   // Start aggressive health polling during restart
   const startHealthPolling = useCallback(() => {
     if (healthPollRef.current) return;
@@ -113,7 +169,7 @@ export function GpuAcceleration() {
     }, 1000);
   }, [queryClient]);
 
-  const handleDownload = async () => {
+  const handleDownloadCuda = async () => {
     setError(null);
     try {
       await apiClient.downloadCudaBackend();
@@ -128,6 +184,21 @@ export function GpuAcceleration() {
     }
   };
 
+  const handleDownloadRocm = async () => {
+    setError(null);
+    try {
+      await apiClient.downloadRocmBackend();
+      refetchRocmStatus();
+    } catch (e: unknown) {
+      const msg = e instanceof Error ? e.message : 'Failed to start download';
+      if (msg.includes('already downloaded')) {
+        refetchRocmStatus();
+      } else {
+        setError(msg);
+      }
+    }
+  };
+
   const handleRestart = async () => {
     setError(null);
     setRestartPhase('stopping');
@@ -154,18 +225,17 @@ export function GpuAcceleration() {
     }
   };
 
-  const handleSwitchToCpu = async () => {
-    // To switch to CPU: delete the CUDA binary, then restart.
-    // start_server always prefers CUDA if present, so we must remove it first.
+  const handleSwitchToCpuFromCuda = async () => {
     setError(null);
     setRestartPhase('stopping');
 
     try {
-      await apiClient.deleteCudaBackend();
+      // Tell Rust launcher to skip GPU binary detection on next start.
+      // We cannot delete an active .exe on Windows, so we override instead.
+      await platform.lifecycle.setBackendOverride('cpu');
       setRestartPhase('waiting');
       startHealthPolling();
       await platform.lifecycle.restartServer();
-      // Invoke resolved — server is likely ready
       if (healthPollRef.current) {
         clearInterval(healthPollRef.current);
         healthPollRef.current = null;
@@ -184,7 +254,36 @@ export function GpuAcceleration() {
     }
   };
 
-  const handleDelete = async () => {
+  const handleSwitchToCpuFromRocm = async () => {
+    setError(null);
+    setRestartPhase('stopping');
+
+    try {
+      // Tell Rust launcher to skip GPU binary detection on next start.
+      // We cannot delete an active .exe on Windows, so we override instead.
+      await platform.lifecycle.setBackendOverride('cpu');
+      setRestartPhase('waiting');
+      startHealthPolling();
+      await platform.lifecycle.restartServer();
+      if (healthPollRef.current) {
+        clearInterval(healthPollRef.current);
+        healthPollRef.current = null;
+      }
+      setRestartPhase('ready');
+      queryClient.invalidateQueries();
+      setTimeout(() => setRestartPhase('idle'), 2000);
+    } catch (e: unknown) {
+      setRestartPhase('idle');
+      if (healthPollRef.current) {
+        clearInterval(healthPollRef.current);
+        healthPollRef.current = null;
+      }
+      setError(e instanceof Error ? e.message : 'Failed to switch to CPU');
+      refetchRocmStatus();
+    }
+  };
+
+  const handleDeleteCuda = async () => {
     setError(null);
     try {
       await apiClient.deleteCudaBackend();
@@ -194,6 +293,16 @@ export function GpuAcceleration() {
     }
   };
 
+  const handleDeleteRocm = async () => {
+    setError(null);
+    try {
+      await apiClient.deleteRocmBackend();
+      refetchRocmStatus();
+    } catch (e: unknown) {
+      setError(e instanceof Error ? e.message : 'Failed to delete ROCm backend');
+    }
+  };
+
   const formatBytes = (bytes: number): string => {
     if (bytes === 0) return '0 B';
     const k = 1024;
@@ -205,7 +314,7 @@ export function GpuAcceleration() {
   // Don't render until health data is available
   if (!health) return null;
 
-  // If the system already has native GPU (MPS, etc.), only show info - no CUDA needed
+  // If the system already has native GPU (MPS, ROCm active, etc.), only show info - no download needed
   const hasNativeGpu =
     health.gpu_available &&
     !isCurrentlyCuda &&
@@ -241,8 +350,6 @@ export function GpuAcceleration() {
           )}
         </div>
 
-        {/* Native GPU detected - no CUDA download needed */}
-
         {/* Currently running CUDA - show switch back to CPU */}
         {isCurrentlyCuda && platform.metadata.isTauri && (
           <>
@@ -261,7 +368,12 @@ export function GpuAcceleration() {
                   Running with CUDA GPU acceleration. Switch back to CPU if needed (you can
                   re-download later).
                 </p>
-                <Button onClick={handleSwitchToCpu} variant="outline" className="w-full" size="sm">
+                <Button
+                  onClick={handleSwitchToCpuFromCuda}
+                  variant="outline"
+                  className="w-full"
+                  size="sm"
+                >
                   <RotateCw className="h-4 w-4 mr-2" />
                   Switch to CPU Backend
                 </Button>
@@ -276,39 +388,207 @@ export function GpuAcceleration() {
           </>
         )}
 
-        {/* CUDA download/manage section - show when no native GPU and not currently running CUDA */}
-        {!hasNativeGpu && !isCurrentlyCuda && (
+        {/* Currently running ROCm - show switch back to CPU */}
+        {isCurrentlyRocm && platform.metadata.isTauri && (
           <>
-            {/* Download progress (manual download or auto-update) */}
-            {cudaDownloading && downloadProgress && (
-              <div className="space-y-2">
-                <div className="flex items-center justify-between text-sm">
-                  <div className="flex items-center gap-2">
-                    <Loader2 className="h-4 w-4 animate-spin" />
-                    <span>
-                      {downloadProgress.filename ||
-                        (cudaAvailable
-                          ? 'Updating CUDA backend...'
-                          : 'Downloading CUDA backend...')}
-                    </span>
-                  </div>
-                  {downloadProgress.total > 0 && (
-                    <span className="text-muted-foreground">
-                      {downloadProgress.progress.toFixed(1)}%
-                    </span>
-                  )}
-                </div>
-                {downloadProgress.total > 0 && (
-                  <>
-                    <Progress value={downloadProgress.progress} className="h-2" />
-                    <div className="text-xs text-muted-foreground">
-                      {formatBytes(downloadProgress.current)} /{' '}
-                      {formatBytes(downloadProgress.total)}
-                    </div>
-                  </>
-                )}
+            {restartPhase !== 'idle' ? (
+              <div className="flex items-center gap-2 p-3 rounded-lg bg-primary/5 border">
+                <Loader2 className="h-4 w-4 animate-spin" />
+                <span className="text-sm">
+                  {restartPhase === 'stopping' && 'Stopping server...'}
+                  {restartPhase === 'waiting' && 'Restarting server...'}
+                  {restartPhase === 'ready' && 'Server restarted successfully!'}
+                </span>
+              </div>
+            ) : (
+              <div className="space-y-3">
+                <p className="text-sm text-muted-foreground">
+                  Running with ROCm GPU acceleration for AMD. Switch back to CPU if needed (you can
+                  re-download later).
+                </p>
+                <Button
+                  onClick={handleSwitchToCpuFromRocm}
+                  variant="outline"
+                  className="w-full"
+                  size="sm"
+                >
+                  <RotateCw className="h-4 w-4 mr-2" />
+                  Switch to CPU Backend
+                </Button>
               </div>
             )}
+            {error && (
+              <div className="flex items-center gap-2 text-sm text-destructive">
+                <AlertCircle className="h-4 w-4 shrink-0" />
+                <span>{error}</span>
+              </div>
+            )}
+          </>
+        )}
+
+        {/* Backend download/manage sections - show when no native GPU and not currently running GPU */}
+        {!hasNativeGpu && !isCurrentlyCuda && !isCurrentlyRocm && (
+          <>
+            {/* CUDA Section */}
+            <div className="space-y-4">
+              <div className="text-sm font-medium">NVIDIA (CUDA)</div>
+
+              {/* CUDA Download progress */}
+              {cudaDownloading && downloadProgress && (
+                <div className="space-y-2">
+                  <div className="flex items-center justify-between text-sm">
+                    <div className="flex items-center gap-2">
+                      <Loader2 className="h-4 w-4 animate-spin" />
+                      <span>
+                        {downloadProgress.filename ||
+                          (cudaAvailable
+                            ? 'Updating CUDA backend...'
+                            : 'Downloading CUDA backend...')}
+                      </span>
+                    </div>
+                    {downloadProgress.total > 0 && (
+                      <span className="text-muted-foreground">
+                        {downloadProgress.progress.toFixed(1)}%
+                      </span>
+                    )}
+                  </div>
+                  {downloadProgress.total > 0 && (
+                    <>
+                      <Progress value={downloadProgress.progress} className="h-2" />
+                      <div className="text-xs text-muted-foreground">
+                        {formatBytes(downloadProgress.current)} /{' '}
+                        {formatBytes(downloadProgress.total)}
+                      </div>
+                    </>
+                  )}
+                </div>
+              )}
+
+              {/* CUDA Actions */}
+              {restartPhase === 'idle' && !cudaDownloading && (
+                <div className="space-y-2">
+                  {!cudaAvailable && (
+                    <div className="space-y-3">
+                      <p className="text-sm text-muted-foreground">
+                        Download the CUDA backend (~2.4 GB) for NVIDIA GPU acceleration. Requires an
+                        NVIDIA GPU with CUDA support.
+                      </p>
+                      <Button onClick={handleDownloadCuda} className="w-full" size="sm">
+                        <Download className="h-4 w-4 mr-2" />
+                        Download CUDA Backend
+                      </Button>
+                    </div>
+                  )}
+
+                  {cudaAvailable && platform.metadata.isTauri && (
+                    <div className="space-y-3">
+                      <p className="text-sm text-muted-foreground">
+                        CUDA backend is downloaded and ready. Restart the server to enable GPU
+                        acceleration.
+                      </p>
+                      <Button onClick={handleRestart} className="w-full" size="sm">
+                        <RotateCw className="h-4 w-4 mr-2" />
+                        Switch to CUDA Backend
+                      </Button>
+                    </div>
+                  )}
+
+                  {cudaAvailable && (
+                    <Button
+                      onClick={handleDeleteCuda}
+                      variant="ghost"
+                      className="w-full text-muted-foreground hover:text-destructive"
+                      size="sm"
+                    >
+                      <Trash2 className="h-4 w-4 mr-2" />
+                      Remove CUDA Backend
+                    </Button>
+                  )}
+                </div>
+              )}
+            </div>
+
+            {/* Divider */}
+            <div className="border-t" />
+
+            {/* ROCm Section */}
+            <div className="space-y-4">
+              <div className="text-sm font-medium">AMD (ROCm)</div>
+
+              {/* ROCm Download progress */}
+              {rocmDownloading && rocmDownloadProgress && (
+                <div className="space-y-2">
+                  <div className="flex items-center justify-between text-sm">
+                    <div className="flex items-center gap-2">
+                      <Loader2 className="h-4 w-4 animate-spin" />
+                      <span>
+                        {rocmDownloadProgress.filename ||
+                          (rocmAvailable
+                            ? 'Updating ROCm backend...'
+                            : 'Downloading ROCm backend...')}
+                      </span>
+                    </div>
+                    {rocmDownloadProgress.total > 0 && (
+                      <span className="text-muted-foreground">
+                        {rocmDownloadProgress.progress.toFixed(1)}%
+                      </span>
+                    )}
+                  </div>
+                  {rocmDownloadProgress.total > 0 && (
+                    <>
+                      <Progress value={rocmDownloadProgress.progress} className="h-2" />
+                      <div className="text-xs text-muted-foreground">
+                        {formatBytes(rocmDownloadProgress.current)} /{' '}
+                        {formatBytes(rocmDownloadProgress.total)}
+                      </div>
+                    </>
+                  )}
+                </div>
+              )}
+
+              {/* ROCm Actions */}
+              {restartPhase === 'idle' && !rocmDownloading && (
+                <div className="space-y-2">
+                  {!rocmAvailable && (
+                    <div className="space-y-3">
+                      <p className="text-sm text-muted-foreground">
+                        Download the ROCm backend (~2-3 GB) for AMD GPU acceleration. Requires an
+                        AMD Radeon GPU with ROCm support.
+                      </p>
+                      <Button onClick={handleDownloadRocm} className="w-full" size="sm">
+                        <Download className="h-4 w-4 mr-2" />
+                        Download AMD ROCm Backend
+                      </Button>
+                    </div>
+                  )}
+
+                  {rocmAvailable && platform.metadata.isTauri && (
+                    <div className="space-y-3">
+                      <p className="text-sm text-muted-foreground">
+                        ROCm backend is downloaded and ready. Restart the server to enable AMD GPU
+                        acceleration.
+                      </p>
+                      <Button onClick={handleRestart} className="w-full" size="sm">
+                        <RotateCw className="h-4 w-4 mr-2" />
+                        Switch to ROCm Backend
+                      </Button>
+                    </div>
+                  )}
+
+                  {rocmAvailable && (
+                    <Button
+                      onClick={handleDeleteRocm}
+                      variant="ghost"
+                      className="w-full text-muted-foreground hover:text-destructive"
+                      size="sm"
+                    >
+                      <Trash2 className="h-4 w-4 mr-2" />
+                      Remove ROCm Backend
+                    </Button>
+                  )}
+                </div>
+              )}
+            </div>
 
             {/* Restart in progress */}
             {restartPhase !== 'idle' && (
@@ -329,52 +609,6 @@ export function GpuAcceleration() {
                 <span>{error}</span>
               </div>
             )}
-
-            {/* Actions */}
-            {restartPhase === 'idle' && !cudaDownloading && (
-              <div className="space-y-2">
-                {/* Not downloaded yet - show download button */}
-                {!cudaAvailable && (
-                  <div className="space-y-3">
-                    <p className="text-sm text-muted-foreground">
-                      Download the CUDA backend (~2.4 GB) for NVIDIA GPU acceleration. Requires an
-                      NVIDIA GPU with CUDA support.
-                    </p>
-                    <Button onClick={handleDownload} className="w-full" size="sm">
-                      <Download className="h-4 w-4 mr-2" />
-                      Download CUDA Backend
-                    </Button>
-                  </div>
-                )}
-
-                {/* Downloaded but not active - show switch button */}
-                {cudaAvailable && platform.metadata.isTauri && (
-                  <div className="space-y-3">
-                    <p className="text-sm text-muted-foreground">
-                      CUDA backend is downloaded and ready. Restart the server to enable GPU
-                      acceleration.
-                    </p>
-                    <Button onClick={handleRestart} className="w-full" size="sm">
-                      <RotateCw className="h-4 w-4 mr-2" />
-                      Switch to CUDA Backend
-                    </Button>
-                  </div>
-                )}
-
-                {/* Delete option when downloaded (and not active) */}
-                {cudaAvailable && (
-                  <Button
-                    onClick={handleDelete}
-                    variant="ghost"
-                    className="w-full text-muted-foreground "
-                    size="sm"
-                  >
-                    <Trash2 className="h-4 w-4 mr-2" />
-                    Remove CUDA Backend
-                  </Button>
-                )}
-              </div>
-            )}
           </>
         )}
       </CardContent>

app/src/components/ServerTab/GpuPage.tsx

@@ -5,7 +5,7 @@ import { useTranslation } from 'react-i18next';
 import { Button } from '@/components/ui/button';
 import { Progress } from '@/components/ui/progress';
 import { apiClient } from '@/lib/api/client';
-import type { CudaDownloadProgress, HealthResponse } from '@/lib/api/types';
+import type { CudaDownloadProgress, RocmDownloadProgress, HealthResponse } from '@/lib/api/types';
 import { useServerHealth } from '@/lib/hooks/useServer';
 import { usePlatform } from '@/platform/PlatformContext';
 import { useServerStore } from '@/stores/serverStore';
@@ -50,7 +50,10 @@ function GpuInfoCard({ health }: { health: HealthResponse }) {
     : null;
   const gpuBackend = hasGpu ? health.gpu_type!.replace(/\s*\(.+\)$/, '') : null;
   const isApple = gpuBackend === 'MPS' || gpuBackend === 'Metal';
-  const showBackendVariant = health.backend_variant && health.backend_variant !== 'cpu';
+  const showBackendVariant = 
+    health.backend_variant && 
+    health.backend_variant !== 'cpu' &&
+    health.backend_variant.toLowerCase() !== gpuBackend?.toLowerCase();
 
   return (
     <div className="rounded-lg border border-border/60 p-4">
@@ -115,10 +118,14 @@ export function GpuPage() {
 
   const [restartPhase, setRestartPhase] = useState<RestartPhase>('idle');
   const [error, setError] = useState<string | null>(null);
+  const [cudaStreaming, setCudaStreaming] = useState(false);
+  const [rocmStreaming, setRocmStreaming] = useState(false);
   const [downloadProgress, setDownloadProgress] = useState<CudaDownloadProgress | null>(null);
+  const [rocmDownloadProgress, setRocmDownloadProgress] = useState<RocmDownloadProgress | null>(
+    null,
+  );
   const healthPollRef = useRef<ReturnType<typeof setInterval> | null>(null);
-  // Hold the latest `t` in a ref so the CUDA progress SSE effect below doesn't
-  // tear down and reconnect the EventSource every time the language changes.
+
   const tRef = useRef(t);
   useEffect(() => {
     tRef.current = t;
@@ -136,9 +143,27 @@ export function GpuPage() {
     enabled: !!health,
   });
 
+  const {
+    data: rocmStatus,
+    isLoading: _rocmStatusLoading,
+    refetch: refetchRocmStatus,
+  } = useQuery({
+    queryKey: ['rocm-status', serverUrl],
+    queryFn: () => apiClient.getRocmStatus(),
+    refetchInterval: (query) => (query.state.status === 'pending' ? false : 10000),
+    retry: 1,
+    enabled: !!health,
+  });
+
   const isCurrentlyCuda = health?.backend_variant === 'cuda';
+  const isCurrentlyRocm = health?.backend_variant === 'rocm';
   const cudaAvailable = cudaStatus?.available ?? false;
   const cudaDownloading = cudaStatus?.downloading ?? false;
+  const rocmAvailable = rocmStatus?.available ?? false;
+  const rocmDownloading = rocmStatus?.downloading ?? false;
+  // The ROCm backend only applies to AMD GPUs on Windows. Show the section when
+  // the backend detects applicable hardware, or it is already downloaded/active.
+  const supportsRocm = (health?.supports_rocm ?? false) || rocmAvailable || isCurrentlyRocm;
 
   useEffect(() => {
     return () => {
@@ -150,7 +175,7 @@ export function GpuPage() {
   }, []);
 
   useEffect(() => {
-    if (!cudaDownloading || !serverUrl) return;
+    if ((!cudaDownloading && !cudaStreaming) || !serverUrl) return;
 
     const eventSource = new EventSource(`${serverUrl}/backend/cuda-progress`);
 
@@ -162,11 +187,13 @@ export function GpuPage() {
         if (data.status === 'complete') {
           eventSource.close();
           setDownloadProgress(null);
+          setCudaStreaming(false);
           refetchCudaStatus();
         } else if (data.status === 'error') {
           eventSource.close();
           setError(data.error || tRef.current('settings.gpu.errors.downloadFailed'));
           setDownloadProgress(null);
+          setCudaStreaming(false);
           refetchCudaStatus();
         }
       } catch (e) {
@@ -176,12 +203,50 @@ export function GpuPage() {
 
     eventSource.onerror = () => {
       eventSource.close();
+      setCudaStreaming(false);
     };
 
     return () => {
       eventSource.close();
     };
-  }, [cudaDownloading, serverUrl, refetchCudaStatus]);
+  }, [cudaDownloading, cudaStreaming, serverUrl, refetchCudaStatus]);
+
+  useEffect(() => {
+    if ((!rocmDownloading && !rocmStreaming) || !serverUrl) return;
+
+    const eventSource = new EventSource(`${serverUrl}/backend/rocm-progress`);
+
+    eventSource.onmessage = (event) => {
+      try {
+        const data = JSON.parse(event.data) as RocmDownloadProgress;
+        setRocmDownloadProgress(data);
+
+        if (data.status === 'complete') {
+          eventSource.close();
+          setRocmDownloadProgress(null);
+          setRocmStreaming(false);
+          refetchRocmStatus();
+        } else if (data.status === 'error') {
+          eventSource.close();
+          setError(data.error || tRef.current('settings.gpu.errors.downloadFailed'));
+          setRocmDownloadProgress(null);
+          setRocmStreaming(false);
+          refetchRocmStatus();
+        }
+      } catch (e) {
+        console.error('Error parsing ROCm progress event:', e);
+      }
+    };
+
+    eventSource.onerror = () => {
+      eventSource.close();
+      setRocmStreaming(false);
+    };
+
+    return () => {
+      eventSource.close();
+    };
+  }, [rocmDownloading, rocmStreaming, serverUrl, refetchRocmStatus]);
 
   const clearHealthPolling = useCallback(() => {
     if (healthPollRef.current) {
@@ -224,10 +289,11 @@ export function GpuPage() {
     [platform, startHealthPolling, clearHealthPolling],
   );
 
-  const handleDownload = async () => {
+  const handleDownloadCuda = async () => {
     setError(null);
     try {
       await apiClient.downloadCudaBackend();
+      setCudaStreaming(true);
       refetchCudaStatus();
     } catch (e: unknown) {
       const msg = e instanceof Error ? e.message : t('settings.gpu.errors.downloadStart');
@@ -239,28 +305,64 @@ export function GpuPage() {
     }
   };
 
-  const handleRestart = async () => {
+  const handleDownloadRocm = async () => {
     setError(null);
     try {
-      await restartServerWithPolling(t('settings.gpu.errors.restartFailed'));
+      await apiClient.downloadRocmBackend();
+      setRocmStreaming(true);
+      refetchRocmStatus();
     } catch (e: unknown) {
-      setError(e instanceof Error ? e.message : t('settings.gpu.errors.restartFailed'));
+      const msg = e instanceof Error ? e.message : t('settings.gpu.errors.downloadStart');
+      if (msg.includes('already downloaded')) {
+        refetchRocmStatus();
+      } else {
+        setError(msg);
+      }
     }
   };
 
+
   const handleSwitchToCpu = async () => {
     setError(null);
     setRestartPhase('stopping');
     try {
-      await apiClient.deleteCudaBackend();
+      await platform.lifecycle.setBackendOverride('cpu');
       await restartServerWithPolling(t('settings.gpu.errors.switchCpu'));
     } catch (e: unknown) {
+      setRestartPhase('idle');
       setError(e instanceof Error ? e.message : t('settings.gpu.errors.switchCpu'));
       refetchCudaStatus();
+      refetchRocmStatus();
+    }
+  };
+
+  const handleSwitchToCuda = async () => {
+    setError(null);
+    setRestartPhase('stopping');
+    try {
+      await platform.lifecycle.setBackendOverride('cuda');
+      await restartServerWithPolling(t('settings.gpu.errors.restartFailed'));
+    } catch (e: unknown) {
+      setRestartPhase('idle');
+      setError(e instanceof Error ? e.message : t('settings.gpu.errors.restartFailed'));
+      refetchCudaStatus();
     }
   };
 
-  const handleDelete = async () => {
+  const handleSwitchToRocm = async () => {
+    setError(null);
+    setRestartPhase('stopping');
+    try {
+      await platform.lifecycle.setBackendOverride('rocm');
+      await restartServerWithPolling(t('settings.gpu.errors.restartFailed'));
+    } catch (e: unknown) {
+      setRestartPhase('idle');
+      setError(e instanceof Error ? e.message : t('settings.gpu.errors.restartFailed'));
+      refetchRocmStatus();
+    }
+  };
+
+  const handleDeleteCuda = async () => {
     setError(null);
     try {
       await apiClient.deleteCudaBackend();
@@ -270,6 +372,16 @@ export function GpuPage() {
     }
   };
 
+  const handleDeleteRocm = async () => {
+    setError(null);
+    try {
+      await apiClient.deleteRocmBackend();
+      refetchRocmStatus();
+    } catch (e: unknown) {
+      setError(e instanceof Error ? e.message : t('settings.gpu.errors.deleteRocm'));
+    }
+  };
+
   const formatBytes = (bytes: number): string => {
     if (bytes === 0) return '0 B';
     const k = 1024;
@@ -283,6 +395,7 @@ export function GpuPage() {
   const hasNativeGpu =
     health.gpu_available &&
     !isCurrentlyCuda &&
+    !isCurrentlyRocm &&
     health.gpu_type &&
     !health.gpu_type.includes('CUDA');
 
@@ -290,33 +403,188 @@ export function GpuPage() {
     <div className="space-y-8 max-w-2xl">
       <GpuInfoCard health={health} />
 
-      {!hasNativeGpu && !isCurrentlyCuda && (
-        <SettingSection
-          title={t('settings.gpu.cuda.title')}
-          description={t('settings.gpu.cuda.description')}
-        >
-          {cudaDownloading && downloadProgress && (
-            <SettingRow title={t('settings.gpu.cuda.downloading')}>
-              <div className="space-y-1.5">
-                <Progress value={downloadProgress.progress} className="h-2" />
-                <div className="flex items-center justify-between text-xs text-muted-foreground">
-                  <span>
-                    {downloadProgress.filename ||
-                      (cudaAvailable
-                        ? t('settings.gpu.cuda.updating')
-                        : t('settings.gpu.cuda.downloadingShort'))}
-                  </span>
-                  <span>
-                    {downloadProgress.total > 0
-                      ? `${formatBytes(downloadProgress.current)} / ${formatBytes(downloadProgress.total)}`
-                      : `${downloadProgress.progress.toFixed(1)}%`}
-                  </span>
+      {!hasNativeGpu && !isCurrentlyCuda && !isCurrentlyRocm && (
+        <>
+          <SettingSection
+            title={t('settings.gpu.cuda.title')}
+            description={t('settings.gpu.cuda.description')}
+          >
+            {cudaDownloading && downloadProgress && (
+              <SettingRow title={t('settings.gpu.cuda.downloading')}>
+                <div className="space-y-1.5">
+                  <Progress value={downloadProgress.progress} className="h-2" />
+                  <div className="flex items-center justify-between text-xs text-muted-foreground">
+                    <span>
+                      {downloadProgress.filename ||
+                        (cudaAvailable
+                          ? t('settings.gpu.cuda.updating')
+                          : t('settings.gpu.cuda.downloadingShort'))}
+                    </span>
+                    <span>
+                      {downloadProgress.total > 0
+                        ? `${formatBytes(downloadProgress.current)} / ${formatBytes(downloadProgress.total)}`
+                        : `${downloadProgress.progress.toFixed(1)}%`}
+                    </span>
+                  </div>
                 </div>
-              </div>
-            </SettingRow>
-          )}
+              </SettingRow>
+            )}
 
-          {restartPhase !== 'idle' && (
+            {restartPhase !== 'idle' && (
+              <SettingRow
+                title={
+                  restartPhase === 'ready'
+                    ? t('settings.gpu.restart.ready')
+                    : restartPhase === 'waiting'
+                      ? t('settings.gpu.restart.waiting')
+                      : t('settings.gpu.restart.stopping')
+                }
+                action={<Loader2 className="h-4 w-4 animate-spin text-muted-foreground" />}
+              />
+            )}
+
+            {error && (
+              <SettingRow title={t('common.error')}>
+                <div className="flex items-center gap-2 text-sm text-destructive">
+                  <AlertCircle className="h-4 w-4 shrink-0" />
+                  <span>{error}</span>
+                </div>
+              </SettingRow>
+            )}
+
+            {restartPhase === 'idle' && !cudaDownloading && (
+              <>
+                {!cudaAvailable && !isCurrentlyCuda && (
+                  <SettingRow
+                    title={t('settings.gpu.download.title')}
+                    description={t('settings.gpu.download.description')}
+                    action={
+                      <Button onClick={handleDownloadCuda} size="sm">
+                        <Download className="h-3.5 w-3.5 mr-1.5" />
+                        {t('settings.gpu.download.button')}
+                      </Button>
+                    }
+                  />
+                )}
+
+                {cudaAvailable && !isCurrentlyCuda && platform.metadata.isTauri && (
+                  <SettingRow
+                    title={t('settings.gpu.switchToCuda.title')}
+                    description={t('settings.gpu.switchToCuda.description')}
+                    action={
+                      <Button onClick={handleSwitchToCuda} size="sm">
+                        <RotateCw className="h-3.5 w-3.5 mr-1.5" />
+                        {t('settings.gpu.switchToCuda.button')}
+                      </Button>
+                    }
+                  />
+                )}
+
+                {cudaAvailable && !isCurrentlyCuda && (
+                  <SettingRow
+                    title={t('settings.gpu.remove.title')}
+                    description={t('settings.gpu.remove.description')}
+                    action={
+                      <Button
+                        onClick={handleDeleteCuda}
+                        variant="ghost"
+                        size="sm"
+                        className="text-muted-foreground hover:text-destructive"
+                      >
+                        <Trash2 className="h-3.5 w-3.5 mr-1.5" />
+                        {t('settings.gpu.remove.button')}
+                      </Button>
+                    }
+                  />
+                )}
+              </>
+            )}
+          </SettingSection>
+
+          {supportsRocm && (
+          <SettingSection
+            title={t('settings.gpu.rocm.title')}
+            description={t('settings.gpu.rocm.description')}
+          >
+            {rocmDownloading && rocmDownloadProgress && (
+              <SettingRow title={t('settings.gpu.rocm.downloading')}>
+                <div className="space-y-1.5">
+                  <Progress value={rocmDownloadProgress.progress} className="h-2" />
+                  <div className="flex items-center justify-between text-xs text-muted-foreground">
+                    <span>
+                      {rocmDownloadProgress.filename ||
+                        (rocmAvailable
+                          ? t('settings.gpu.rocm.updating')
+                          : t('settings.gpu.rocm.downloadingShort'))}
+                    </span>
+                    <span>
+                      {rocmDownloadProgress.total > 0
+                        ? `${formatBytes(rocmDownloadProgress.current)} / ${formatBytes(rocmDownloadProgress.total)}`
+                        : `${rocmDownloadProgress.progress.toFixed(1)}%`}
+                    </span>
+                  </div>
+                </div>
+              </SettingRow>
+            )}
+
+            {restartPhase === 'idle' && !rocmDownloading && (
+              <>
+                {!rocmAvailable && !isCurrentlyRocm && (
+                  <SettingRow
+                    title={t('settings.gpu.downloadRocm.title')}
+                    description={t('settings.gpu.downloadRocm.description')}
+                    action={
+                      <Button onClick={handleDownloadRocm} size="sm">
+                        <Download className="h-3.5 w-3.5 mr-1.5" />
+                        {t('settings.gpu.downloadRocm.button')}
+                      </Button>
+                    }
+                  />
+                )}
+
+                {rocmAvailable && !isCurrentlyRocm && platform.metadata.isTauri && (
+                  <SettingRow
+                    title={t('settings.gpu.switchToRocm.title')}
+                    description={t('settings.gpu.switchToRocm.description')}
+                    action={
+                      <Button onClick={handleSwitchToRocm} size="sm">
+                        <RotateCw className="h-3.5 w-3.5 mr-1.5" />
+                        {t('settings.gpu.switchToRocm.button')}
+                      </Button>
+                    }
+                  />
+                )}
+
+                {rocmAvailable && !isCurrentlyRocm && (
+                  <SettingRow
+                    title={t('settings.gpu.removeRocm.title')}
+                    description={t('settings.gpu.removeRocm.description')}
+                    action={
+                      <Button
+                        onClick={handleDeleteRocm}
+                        variant="ghost"
+                        size="sm"
+                        className="text-muted-foreground hover:text-destructive"
+                      >
+                        <Trash2 className="h-3.5 w-3.5 mr-1.5" />
+                        {t('settings.gpu.removeRocm.button')}
+                      </Button>
+                    }
+                  />
+                )}
+              </>
+            )}
+          </SettingSection>
+          )}
+        </>
+      )}
+
+      {(isCurrentlyCuda || isCurrentlyRocm) && platform.metadata.isTauri && (
+        <SettingSection
+          title={isCurrentlyCuda ? t('settings.gpu.cuda.activeTitle') : t('settings.gpu.rocm.activeTitle')}
+          description={t('settings.gpu.activeBackend.description')}
+        >
+          {restartPhase !== 'idle' ? (
             <SettingRow
               title={
                 restartPhase === 'ready'
@@ -327,8 +595,18 @@ export function GpuPage() {
               }
               action={<Loader2 className="h-4 w-4 animate-spin text-muted-foreground" />}
             />
+          ) : (
+            <SettingRow
+              title={t('settings.gpu.switchToCpu.title')}
+              description={t('settings.gpu.switchToCpu.description')}
+              action={
+                <Button onClick={handleSwitchToCpu} variant="outline" size="sm">
+                  <RotateCw className="h-3.5 w-3.5 mr-1.5" />
+                  {t('settings.gpu.switchToCpu.button')}
+                </Button>
+              }
+            />
           )}
-
           {error && (
             <SettingRow title={t('common.error')}>
               <div className="flex items-center gap-2 text-sm text-destructive">
@@ -337,67 +615,6 @@ export function GpuPage() {
               </div>
             </SettingRow>
           )}
-
-          {restartPhase === 'idle' && !cudaDownloading && (
-            <>
-              {!cudaAvailable && !isCurrentlyCuda && (
-                <SettingRow
-                  title={t('settings.gpu.download.title')}
-                  description={t('settings.gpu.download.description')}
-                  action={
-                    <Button onClick={handleDownload} size="sm">
-                      <Download className="h-3.5 w-3.5 mr-1.5" />
-                      {t('settings.gpu.download.button')}
-                    </Button>
-                  }
-                />
-              )}
-
-              {cudaAvailable && !isCurrentlyCuda && platform.metadata.isTauri && (
-                <SettingRow
-                  title={t('settings.gpu.switchToCuda.title')}
-                  description={t('settings.gpu.switchToCuda.description')}
-                  action={
-                    <Button onClick={handleRestart} size="sm">
-                      <RotateCw className="h-3.5 w-3.5 mr-1.5" />
-                      {t('settings.gpu.switchToCuda.button')}
-                    </Button>
-                  }
-                />
-              )}
-
-              {isCurrentlyCuda && platform.metadata.isTauri && (
-                <SettingRow
-                  title={t('settings.gpu.switchToCpu.title')}
-                  description={t('settings.gpu.switchToCpu.description')}
-                  action={
-                    <Button onClick={handleSwitchToCpu} variant="outline" size="sm">
-                      <RotateCw className="h-3.5 w-3.5 mr-1.5" />
-                      {t('settings.gpu.switchToCpu.button')}
-                    </Button>
-                  }
-                />
-              )}
-
-              {cudaAvailable && !isCurrentlyCuda && (
-                <SettingRow
-                  title={t('settings.gpu.remove.title')}
-                  description={t('settings.gpu.remove.description')}
-                  action={
-                    <Button
-                      onClick={handleDelete}
-                      variant="ghost"
-                      size="sm"
-                      className="text-muted-foreground "
-                    >
-                      <Trash2 className="h-3.5 w-3.5 mr-1.5" />
-                      {t('settings.gpu.remove.button')}
-                    </Button>
-                  }
-                />
-              )}
-            </>
-          )}
         </SettingSection>
       )}