Fix R2 bucket name: voicebox (not voicebox-releases)

- Update CI workflow to upload CUDA binary to R2 instead of GitHub
- Use custom domain: downloads.voicebox.sh
- Update release notes with R2 download link
- Add R2 setup documentation with GitHub secrets instructions
- Add local test script for R2 uploads
- Fixes GitHub 2GB release asset limit issue

Required GitHub secrets:
- R2_ACCESS_KEY_ID
- R2_SECRET_ACCESS_KEY
- R2_ENDPOINT

Cost: ~$0.04/month (free bandwidth with R2)

.bumpversion.cfg

@@ -1,5 +1,5 @@
 [bumpversion]
-current_version = 0.1.11
+current_version = 0.1.12
 commit = True
 tag = True
 tag_name = v{new_version}

.github/workflows/release.yml

@@ -66,24 +66,24 @@ jobs:
         run: |
           pip install -r backend/requirements-mlx.txt
 
-      - name: Install PyTorch with CUDA (Windows only)
-        if: matrix.platform == 'windows-latest'
-        run: |
-          pip install torch --index-url https://download.pytorch.org/whl/cu121 --force-reinstall --no-deps
-          pip install torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121
-
       - name: Build Python server (Linux/macOS)
         if: matrix.platform != 'windows-latest'
         run: |
           chmod +x scripts/build-server.sh
           ./scripts/build-server.sh
 
-      - name: Build Python server (Windows)
+      - name: Build CPU Python server (Windows)
         if: matrix.platform == 'windows-latest'
         shell: bash
         run: |
           cd backend
-          python build_binary.py
+
+          echo "Installing CPU-only PyTorch..."
+          pip uninstall -y torch torchvision torchaudio
+          pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu
+
+          echo "Building CPU server binary..."
+          python build_binary.py cpu
 
           # Get platform tuple
           PLATFORM=$(rustc --print host-tuple)
@@ -91,9 +91,31 @@ jobs:
           # Create binaries directory
           mkdir -p ../tauri/src-tauri/binaries
 
-          # Copy with platform suffix
+          # Copy CPU version (default for installer)
           cp dist/voicebox-server.exe ../tauri/src-tauri/binaries/voicebox-server-${PLATFORM}.exe
-          echo "Built voicebox-server-${PLATFORM}.exe"
+          echo "Built CPU server: voicebox-server-${PLATFORM}.exe (~500MB)"
+
+      - name: Build CUDA Python server (Windows)
+        if: matrix.platform == 'windows-latest'
+        shell: bash
+        run: |
+          cd backend
+
+          echo "Installing CUDA PyTorch..."
+          pip uninstall -y torch torchvision torchaudio
+          pip install torch --index-url https://download.pytorch.org/whl/cu121 --force-reinstall --no-deps
+          pip install torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121
+
+          echo "Building CUDA server binary..."
+          python build_binary.py cuda
+
+          # Get platform tuple
+          PLATFORM=$(rustc --print host-tuple)
+
+          # Copy CUDA version for separate upload
+          mkdir -p cuda-release
+          cp dist/voicebox-server-cuda.exe cuda-release/voicebox-server-cuda-${PLATFORM}.exe
+          echo "Built CUDA server: voicebox-server-cuda-${PLATFORM}.exe (~3GB)"
 
       - name: Setup Bun
         uses: oven-sh/setup-bun@v2
@@ -150,11 +172,41 @@ jobs:
             ### Installation
             - **macOS (Apple Silicon)**: Download the `aarch64.dmg` file - uses MLX for fast native inference
             - **macOS (Intel)**: Download the `x64.dmg` file - uses PyTorch
-            - **Windows**: Download the `.msi` installer
+            - **Windows**: Download the `.msi` installer - includes CPU-only inference (~500MB)
             - **Linux**: Download the `.AppImage` or `.deb` package
 
+            ### NVIDIA GPU Acceleration (Windows)
+            Windows users with NVIDIA GPUs can enable CUDA for 4-5x faster inference:
+            1. Install the app normally (CPU version included in installer)
+            2. The app will detect your GPU and offer to download CUDA support automatically
+            3. Or manually download: [voicebox-server-cuda-x86_64-pc-windows-msvc.exe](https://downloads.voicebox.sh/cuda/__VERSION__/voicebox-server-cuda-x86_64-pc-windows-msvc.exe) (~2.4GB)
+
             The app includes automatic updates - future updates will be installed automatically.
           releaseDraft: true
           prerelease: false
           args: ${{ matrix.args }}
           includeUpdaterJson: true
+
+      - name: Upload CUDA server to Cloudflare R2 (Windows only)
+        if: matrix.platform == 'windows-latest'
+        env:
+          AWS_ACCESS_KEY_ID: ${{ secrets.R2_ACCESS_KEY_ID }}
+          AWS_SECRET_ACCESS_KEY: ${{ secrets.R2_SECRET_ACCESS_KEY }}
+          R2_ENDPOINT: ${{ secrets.R2_ENDPOINT }}
+        run: |
+          # Install AWS CLI if not available
+          pip install awscli
+
+          # Get version from tag
+          VERSION=${GITHUB_REF#refs/tags/}
+
+          # Get platform tuple
+          PLATFORM=$(rustc --print host-tuple)
+
+          # Upload to R2
+          aws s3 cp backend/cuda-release/voicebox-server-cuda-${PLATFORM}.exe \
+            s3://voicebox/cuda/${VERSION}/voicebox-server-cuda-${PLATFORM}.exe \
+            --endpoint-url $R2_ENDPOINT \
+            --acl public-read
+
+          echo "CUDA binary uploaded to: https://downloads.voicebox.sh/cuda/${VERSION}/voicebox-server-cuda-${PLATFORM}.exe"

app/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@voicebox/app",
-  "version": "0.1.11",
+  "version": "0.1.12",
   "private": true,
   "type": "module",
   "scripts": {

app/src/components/History/HistoryTable.tsx

@@ -53,6 +53,8 @@ export function HistoryTable() {
   const fileInputRef = useRef<HTMLInputElement>(null);
   const [importDialogOpen, setImportDialogOpen] = useState(false);
   const [selectedFile, setSelectedFile] = useState<File | null>(null);
+  const [deleteDialogOpen, setDeleteDialogOpen] = useState(false);
+  const [generationToDelete, setGenerationToDelete] = useState<{ id: string; name: string } | null>(null);
   const limit = 20;
   const { toast } = useToast();
 
@@ -179,6 +181,19 @@ export function HistoryTable() {
     );
   };
 
+  const handleDeleteClick = (generationId: string, profileName: string) => {
+    setGenerationToDelete({ id: generationId, name: profileName });
+    setDeleteDialogOpen(true);
+  };
+
+  const handleDeleteConfirm = () => {
+    if (generationToDelete) {
+      deleteGeneration.mutate(generationToDelete.id);
+      setDeleteDialogOpen(false);
+      setGenerationToDelete(null);
+    }
+  };
+
   const handleImportConfirm = () => {
     if (selectedFile) {
       importGeneration.mutate(selectedFile, {
@@ -320,7 +335,7 @@ export function HistoryTable() {
                           Export Package
                         </DropdownMenuItem>
                         <DropdownMenuItem
-                          onClick={() => deleteGeneration.mutate(gen.id)}
+                          onClick={() => handleDeleteClick(gen.id, gen.profile_name)}
                           disabled={deleteGeneration.isPending}
                           className="text-destructive focus:text-destructive"
                         >
@@ -351,6 +366,35 @@ export function HistoryTable() {
         </>
       )}
 
+      <Dialog open={deleteDialogOpen} onOpenChange={setDeleteDialogOpen}>
+        <DialogContent>
+          <DialogHeader>
+            <DialogTitle>Delete Generation</DialogTitle>
+            <DialogDescription>
+              Are you sure you want to delete this generation from "{generationToDelete?.name}"? This action cannot be undone.
+            </DialogDescription>
+          </DialogHeader>
+          <DialogFooter>
+            <Button
+              variant="outline"
+              onClick={() => {
+                setDeleteDialogOpen(false);
+                setGenerationToDelete(null);
+              }}
+            >
+              Cancel
+            </Button>
+            <Button
+              variant="destructive"
+              onClick={handleDeleteConfirm}
+              disabled={deleteGeneration.isPending}
+            >
+              {deleteGeneration.isPending ? 'Deleting...' : 'Delete'}
+            </Button>
+          </DialogFooter>
+        </DialogContent>
+      </Dialog>
+
       <Dialog open={importDialogOpen} onOpenChange={setImportDialogOpen}>
         <DialogContent>
           <DialogHeader>

app/src/components/VoiceProfiles/SampleList.tsx

@@ -2,6 +2,14 @@ import { Check, Edit, Pause, Play, Plus, Trash2, Volume2, X } from 'lucide-react
 import { useEffect, useRef, useState } from 'react';
 import { Button } from '@/components/ui/button';
 import { CircleButton } from '@/components/ui/circle-button';
+import {
+  Dialog,
+  DialogContent,
+  DialogDescription,
+  DialogFooter,
+  DialogHeader,
+  DialogTitle,
+} from '@/components/ui/dialog';
 import { Slider } from '@/components/ui/slider';
 import { Textarea } from '@/components/ui/textarea';
 import { useToast } from '@/components/ui/use-toast';
@@ -140,10 +148,19 @@ export function SampleList({ profileId }: SampleListProps) {
   const [uploadOpen, setUploadOpen] = useState(false);
   const [editingSampleId, setEditingSampleId] = useState<string | null>(null);
   const [editedText, setEditedText] = useState<string>('');
+  const [deleteDialogOpen, setDeleteDialogOpen] = useState(false);
+  const [sampleToDelete, setSampleToDelete] = useState<string | null>(null);
 
-  const handleDelete = (sampleId: string) => {
+  const handleDeleteClick = (sampleId: string) => {
-    if (confirm('Are you sure you want to delete this sample?')) {
+    setSampleToDelete(sampleId);
-      deleteSample.mutate(sampleId);
+    setDeleteDialogOpen(true);
+  };
+
+  const handleDeleteConfirm = () => {
+    if (sampleToDelete) {
+      deleteSample.mutate(sampleToDelete);
+      setDeleteDialogOpen(false);
+      setSampleToDelete(null);
     }
   };
 
@@ -268,7 +285,7 @@ export function SampleList({ profileId }: SampleListProps) {
                         <CircleButton
                           icon={Trash2}
                           title="Delete sample"
-                          onClick={() => handleDelete(sample.id)}
+                          onClick={() => handleDeleteClick(sample.id)}
                           disabled={deleteSample.isPending}
                         />
                       </div>
@@ -306,6 +323,35 @@ export function SampleList({ profileId }: SampleListProps) {
       </p>
 
       <SampleUpload profileId={profileId} open={uploadOpen} onOpenChange={setUploadOpen} />
+
+      <Dialog open={deleteDialogOpen} onOpenChange={setDeleteDialogOpen}>
+        <DialogContent>
+          <DialogHeader>
+            <DialogTitle>Delete Sample</DialogTitle>
+            <DialogDescription>
+              Are you sure you want to delete this audio sample? This action cannot be undone.
+            </DialogDescription>
+          </DialogHeader>
+          <DialogFooter>
+            <Button
+              variant="outline"
+              onClick={() => {
+                setDeleteDialogOpen(false);
+                setSampleToDelete(null);
+              }}
+            >
+              Cancel
+            </Button>
+            <Button
+              variant="destructive"
+              onClick={handleDeleteConfirm}
+              disabled={deleteSample.isPending}
+            >
+              {deleteSample.isPending ? 'Deleting...' : 'Delete'}
+            </Button>
+          </DialogFooter>
+        </DialogContent>
+      </Dialog>
     </div>
   );
 }

app/src/components/ui/circle-button.tsx

@@ -6,10 +6,11 @@ export interface CircleButtonProps extends React.ButtonHTMLAttributes<HTMLButton
 }
 
 const CircleButton = React.forwardRef<HTMLButtonElement, CircleButtonProps>(
-  ({ className, icon: Icon, ...props }, ref) => {
+  ({ className, icon: Icon, type = 'button', ...props }, ref) => {
     return (
       <button
         ref={ref}
+        type={type}
         className={cn(
           'h-7 w-7 rounded-full flex items-center justify-center flex-shrink-0',
           'hover:bg-muted transition-colors',

backend/__init__.py

@@ -1,3 +1,3 @@
 # Backend package
 
-__version__ = "0.1.11"
+__version__ = "0.1.12"

backend/build_binary.py

@@ -5,6 +5,7 @@ PyInstaller build script for creating standalone Python server binary.
 import PyInstaller.__main__
 import os
 import platform
+import sys
 from pathlib import Path
 
 
@@ -13,15 +14,27 @@ def is_apple_silicon():
     return platform.system() == "Darwin" and platform.machine() == "arm64"
 
 
-def build_server():
+def build_server(variant="cpu"):
-    """Build Python server as standalone binary."""
+    """Build Python server as standalone binary.
+
+    Args:
+        variant: 'cpu' for CPU-only build (~500MB) or 'cuda' for CUDA build (~3GB)
+    """
     backend_dir = Path(__file__).parent
 
+    if variant not in ['cpu', 'cuda']:
+        raise ValueError(f"Invalid variant: {variant}. Must be 'cpu' or 'cuda'")
+
+    # Set binary name based on variant
+    binary_name = f'voicebox-server-{variant}' if variant == 'cuda' else 'voicebox-server'
+
+    print(f"Building {variant.upper()} variant: {binary_name}")
+
     # PyInstaller arguments
     args = [
         'server.py',  # Use server.py as entry point instead of main.py
         '--onefile',
-        '--name', 'voicebox-server',
+        '--name', binary_name,
     ]
 
     # Add local qwen_tts path if specified (for editable installs)
@@ -100,9 +113,14 @@ def build_server():
     
     # Run PyInstaller
     PyInstaller.__main__.run(args)
-    
+
-    print(f"Binary built in {backend_dir / 'dist' / 'voicebox-server'}")
+    print(f"\n{'='*60}")
+    print(f"Build complete: {variant.upper()} variant")
+    print(f"Binary: {backend_dir / 'dist' / binary_name}")
+    print(f"{'='*60}\n")
 
 
 if __name__ == '__main__':
-    build_server()
+    # Accept variant as command line argument
+    variant = sys.argv[1] if len(sys.argv) > 1 else 'cpu'
+    build_server(variant)

backend/build_both.bat

@@ -0,0 +1,30 @@
+@echo off
+REM Build both CPU and CUDA server binaries for Windows
+
+echo ============================================================
+echo Building BOTH server binaries (CPU + CUDA)
+echo This will take a while...
+echo ============================================================
+
+call build_cpu.bat
+if errorlevel 1 (
+    echo CPU build failed!
+    exit /b 1
+)
+
+echo.
+echo.
+
+call build_cuda.bat
+if errorlevel 1 (
+    echo CUDA build failed!
+    exit /b 1
+)
+
+echo.
+echo ============================================================
+echo Both binaries built successfully!
+echo ============================================================
+echo CPU binary:  dist\voicebox-server.exe (~500MB)
+echo CUDA binary: dist\voicebox-server-cuda.exe (~3GB)
+echo ============================================================

backend/build_cpu.bat

@@ -0,0 +1,28 @@
+@echo off
+REM Build CPU-only server binary for Windows
+REM This creates a ~500MB binary without CUDA support
+
+echo ============================================================
+echo Building CPU-only server binary
+echo ============================================================
+
+echo.
+echo Step 1: Installing CPU-only PyTorch...
+pip uninstall -y torch torchvision torchaudio
+pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu
+
+echo.
+echo Step 2: Building binary with PyInstaller...
+python build_binary.py cpu
+
+echo.
+echo Step 3: Restoring CUDA PyTorch for development...
+pip uninstall -y torch torchvision torchaudio
+pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121
+
+echo.
+echo ============================================================
+echo CPU binary built successfully!
+echo Location: dist\voicebox-server.exe
+echo Size: ~500MB
+echo ============================================================

backend/build_cpu.sh

@@ -0,0 +1,30 @@
+#!/bin/bash
+# Build CPU-only server binary
+# This creates a ~500MB binary without CUDA support
+
+set -e
+
+echo "============================================================"
+echo "Building CPU-only server binary"
+echo "============================================================"
+
+echo ""
+echo "Step 1: Installing CPU-only PyTorch..."
+pip uninstall -y torch torchvision torchaudio || true
+pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu
+
+echo ""
+echo "Step 2: Building binary with PyInstaller..."
+python build_binary.py cpu
+
+echo ""
+echo "Step 3: Restoring CUDA PyTorch for development..."
+pip uninstall -y torch torchvision torchaudio || true
+pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121
+
+echo ""
+echo "============================================================"
+echo "CPU binary built successfully!"
+echo "Location: dist/voicebox-server"
+echo "Size: ~500MB"
+echo "============================================================"

backend/build_cuda.bat

@@ -0,0 +1,22 @@
+@echo off
+REM Build CUDA server binary for Windows
+REM This creates a ~3GB binary with CUDA support
+
+echo ============================================================
+echo Building CUDA server binary
+echo ============================================================
+
+echo.
+echo Step 1: Ensuring CUDA PyTorch is installed...
+pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121 --upgrade
+
+echo.
+echo Step 2: Building binary with PyInstaller...
+python build_binary.py cuda
+
+echo.
+echo ============================================================
+echo CUDA binary built successfully!
+echo Location: dist\voicebox-server-cuda.exe
+echo Size: ~3GB
+echo ============================================================

backend/profiles.py

@@ -22,7 +22,7 @@ from .database import (
 )
 from .utils.audio import validate_reference_audio, load_audio, save_audio
 from .utils.images import validate_image, process_avatar
-from .utils.cache import _get_cache_dir
+from .utils.cache import _get_cache_dir, clear_profile_cache
 from .tts import get_tts_model
 from . import config
 
@@ -121,6 +121,10 @@ async def add_profile_sample(
     db.commit()
     db.refresh(db_sample)
     
+    # Invalidate combined audio cache for this profile
+    # Since a new sample was added, any cached combined audio is now stale
+    clear_profile_cache(profile_id)
+    
     return ProfileSampleResponse.model_validate(db_sample)
 
 
@@ -242,6 +246,9 @@ async def delete_profile(
     if profile_dir.exists():
         shutil.rmtree(profile_dir)
     
+    # Clean up combined audio cache files for this profile
+    clear_profile_cache(profile_id)
+    
     return True
 
 
@@ -263,6 +270,9 @@ async def delete_profile_sample(
     if not sample:
         return False
     
+    # Store profile_id before deleting
+    profile_id = sample.profile_id
+    
     # Delete audio file
     audio_path = Path(sample.audio_path)
     if audio_path.exists():
@@ -272,6 +282,10 @@ async def delete_profile_sample(
     db.delete(sample)
     db.commit()
     
+    # Invalidate combined audio cache for this profile
+    # Since the sample set changed, any cached combined audio is now stale
+    clear_profile_cache(profile_id)
+    
     return True
 
 
@@ -295,10 +309,17 @@ async def update_profile_sample(
     if not sample:
         return None
     
+    # Store profile_id before updating
+    profile_id = sample.profile_id
+    
     sample.reference_text = reference_text
     db.commit()
     db.refresh(sample)
     
+    # Invalidate combined audio cache for this profile
+    # Since the reference text changed, cache keys and combined text are now stale
+    clear_profile_cache(profile_id)
+    
     return ProfileSampleResponse.model_validate(sample)
 
 

backend/test_cuda_compression.py

@@ -0,0 +1,137 @@
+#!/usr/bin/env python3
+"""
+Test CUDA binary compression to verify it fits under GitHub's 2GB release asset limit.
+
+Usage:
+    python test_cuda_compression.py [path/to/voicebox-server-cuda.exe]
+
+If no path provided, looks for the binary in ./dist/
+"""
+
+import os
+import sys
+import subprocess
+from pathlib import Path
+
+
+def format_size(bytes_size):
+    """Format bytes into human-readable size."""
+    for unit in ['B', 'KB', 'MB', 'GB']:
+        if bytes_size < 1024.0:
+            return f"{bytes_size:.2f} {unit}"
+        bytes_size /= 1024.0
+    return f"{bytes_size:.2f} TB"
+
+
+def get_file_size(filepath):
+    """Get file size in bytes."""
+    return os.path.getsize(filepath)
+
+
+def compress_with_7z(input_file, output_file):
+    """Compress file using 7z with maximum compression."""
+    print(f"\nCompressing with 7z (maximum compression)...")
+    print(f"This may take several minutes for a ~2.5GB file...\n")
+
+    cmd = [
+        '7z', 'a',
+        '-t7z',           # 7z format
+        '-m0=lzma2',      # LZMA2 compression
+        '-mx=9',          # Maximum compression
+        '-mfb=64',        # Fast bytes
+        '-md=32m',        # Dictionary size
+        '-ms=on',         # Solid archive
+        output_file,
+        input_file
+    ]
+
+    try:
+        subprocess.run(cmd, check=True, capture_output=True, text=True)
+        return True
+    except subprocess.CalledProcessError as e:
+        print(f"Error during compression: {e}")
+        print(f"stderr: {e.stderr}")
+        return False
+    except FileNotFoundError:
+        print("ERROR: 7z not found. Please install 7-Zip:")
+        print("  Windows: https://www.7-zip.org/download.html")
+        print("  macOS: brew install p7zip")
+        print("  Linux: apt-get install p7zip-full")
+        return False
+
+
+def main():
+    # Find CUDA binary
+    if len(sys.argv) > 1:
+        cuda_binary = Path(sys.argv[1])
+    else:
+        # Look in dist directory
+        dist_dir = Path(__file__).parent / 'dist'
+        candidates = list(dist_dir.glob('voicebox-server-cuda*.exe'))
+
+        if not candidates:
+            print("ERROR: CUDA binary not found in ./dist/")
+            print("Please provide the path as an argument:")
+            print("  python test_cuda_compression.py path/to/voicebox-server-cuda.exe")
+            sys.exit(1)
+
+        cuda_binary = candidates[0]
+
+    if not cuda_binary.exists():
+        print(f"ERROR: File not found: {cuda_binary}")
+        sys.exit(1)
+
+    print("=" * 70)
+    print("CUDA Binary Compression Test")
+    print("=" * 70)
+
+    # Get original size
+    original_size = get_file_size(cuda_binary)
+    print(f"\nOriginal file: {cuda_binary.name}")
+    print(f"Original size: {format_size(original_size)} ({original_size:,} bytes)")
+
+    # Check if already over 2GB
+    github_limit = 2 * 1024 * 1024 * 1024  # 2GB in bytes
+    print(f"GitHub limit:  {format_size(github_limit)} ({github_limit:,} bytes)")
+
+    if original_size > github_limit:
+        print(f"\n[WARNING] Original file exceeds GitHub limit by {format_size(original_size - github_limit)}")
+    else:
+        print(f"\n[OK] Original file is under GitHub limit")
+
+    # Compress
+    output_file = cuda_binary.parent / f"{cuda_binary.stem}.7z"
+    if output_file.exists():
+        print(f"\nRemoving existing compressed file: {output_file.name}")
+        output_file.unlink()
+
+    success = compress_with_7z(cuda_binary, output_file)
+
+    if not success:
+        sys.exit(1)
+
+    # Check compressed size
+    compressed_size = get_file_size(output_file)
+    compression_ratio = (1 - compressed_size / original_size) * 100
+
+    print("\n" + "=" * 70)
+    print("Compression Results")
+    print("=" * 70)
+    print(f"\nCompressed file: {output_file.name}")
+    print(f"Compressed size: {format_size(compressed_size)} ({compressed_size:,} bytes)")
+    print(f"Compression ratio: {compression_ratio:.1f}%")
+    print(f"Space saved: {format_size(original_size - compressed_size)}")
+
+    if compressed_size <= github_limit:
+        print(f"\n[SUCCESS] Compressed file fits under GitHub's 2GB limit!")
+        print(f"   Margin: {format_size(github_limit - compressed_size)} remaining")
+    else:
+        print(f"\n[FAILED] Compressed file still exceeds GitHub limit")
+        print(f"   Over by: {format_size(compressed_size - github_limit)}")
+        print(f"\n   Alternative: Host on external storage (S3, Azure Blob, etc.)")
+
+    print("\n" + "=" * 70)
+
+
+if __name__ == '__main__':
+    main()

backend/test_r2_upload.sh

@@ -0,0 +1,86 @@
+#!/bin/bash
+# Test R2 upload locally before running in CI
+
+set -e
+
+echo "============================================================"
+echo "Cloudflare R2 Upload Test"
+echo "============================================================"
+
+# Check for required environment variables
+if [ -z "$AWS_ACCESS_KEY_ID" ] || [ -z "$AWS_SECRET_ACCESS_KEY" ] || [ -z "$R2_ENDPOINT" ]; then
+    echo "ERROR: Missing required environment variables"
+    echo ""
+    echo "Please set:"
+    echo "  export AWS_ACCESS_KEY_ID='your-r2-access-key-id'"
+    echo "  export AWS_SECRET_ACCESS_KEY='your-r2-secret-access-key'"
+    echo "  export R2_ENDPOINT='https://your-account-id.r2.cloudflarestorage.com'"
+    echo ""
+    exit 1
+fi
+
+# Check for AWS CLI
+if ! command -v aws &> /dev/null; then
+    echo "Installing AWS CLI..."
+    pip install awscli
+fi
+
+# Find CUDA binary
+CUDA_BINARY=$(ls dist/voicebox-server-cuda*.exe 2>/dev/null | head -1)
+
+if [ -z "$CUDA_BINARY" ]; then
+    echo "ERROR: CUDA binary not found in dist/"
+    echo "Run: bash build_cuda.bat"
+    exit 1
+fi
+
+echo ""
+echo "Found CUDA binary: $CUDA_BINARY"
+echo "Size: $(du -h "$CUDA_BINARY" | cut -f1)"
+echo ""
+
+# Test version
+VERSION="v0.1.12-test"
+PLATFORM="x86_64-pc-windows-msvc"
+FILENAME="voicebox-server-cuda-${PLATFORM}.exe"
+
+echo "Test upload configuration:"
+echo "  Version: $VERSION"
+echo "  Platform: $PLATFORM"
+echo "  Endpoint: $R2_ENDPOINT"
+echo "  Bucket: voicebox"
+echo "  Path: cuda/$VERSION/$FILENAME"
+echo ""
+
+read -p "Proceed with upload? (y/n) " -n 1 -r
+echo
+if [[ ! $REPLY =~ ^[Yy]$ ]]; then
+    echo "Aborted."
+    exit 1
+fi
+
+echo ""
+echo "Uploading to R2..."
+
+aws s3 cp "$CUDA_BINARY" \
+    "s3://voicebox/cuda/${VERSION}/${FILENAME}" \
+    --endpoint-url "$R2_ENDPOINT" \
+    --acl public-read
+
+if [ $? -eq 0 ]; then
+    echo ""
+    echo "============================================================"
+    echo "Upload successful!"
+    echo "============================================================"
+    echo ""
+    echo "Download URL:"
+    echo "https://downloads.voicebox.sh/cuda/${VERSION}/${FILENAME}"
+    echo ""
+    echo "Test with:"
+    echo "curl -I https://downloads.voicebox.sh/cuda/${VERSION}/${FILENAME}"
+    echo ""
+else
+    echo ""
+    echo "Upload failed!"
+    exit 1
+fi

backend/utils/cache.py

@@ -105,11 +105,46 @@ def clear_voice_prompt_cache() -> int:
     deleted_count = 0
     
     if cache_dir.exists():
+        # Delete prompt cache files
         for cache_file in cache_dir.glob("*.prompt"):
             try:
                 cache_file.unlink()
                 deleted_count += 1
             except Exception as e:
                 print(f"Failed to delete cache file {cache_file}: {e}")
+        
+        # Delete combined audio files
+        for audio_file in cache_dir.glob("combined_*.wav"):
+            try:
+                audio_file.unlink()
+                deleted_count += 1
+            except Exception as e:
+                print(f"Failed to delete combined audio file {audio_file}: {e}")
+    
+    return deleted_count
+
+
+def clear_profile_cache(profile_id: str) -> int:
+    """
+    Clear cache files for a specific profile.
+    
+    Args:
+        profile_id: Profile ID
+    
+    Returns:
+        Number of cache files deleted
+    """
+    cache_dir = _get_cache_dir()
+    deleted_count = 0
+    
+    if cache_dir.exists():
+        # Delete combined audio files for this profile
+        pattern = f"combined_{profile_id}_*.wav"
+        for audio_file in cache_dir.glob(pattern):
+            try:
+                audio_file.unlink()
+                deleted_count += 1
+            except Exception as e:
+                print(f"Failed to delete combined audio file {audio_file}: {e}")
     
     return deleted_count

backend/voicebox-server-cuda.spec

@@ -0,0 +1,48 @@
+# -*- mode: python ; coding: utf-8 -*-
+from PyInstaller.utils.hooks import collect_data_files
+from PyInstaller.utils.hooks import collect_submodules
+from PyInstaller.utils.hooks import copy_metadata
+
+datas = []
+hiddenimports = ['backend', 'backend.main', 'backend.config', 'backend.database', 'backend.models', 'backend.profiles', 'backend.history', 'backend.tts', 'backend.transcribe', 'backend.platform_detect', 'backend.backends', 'backend.backends.pytorch_backend', 'backend.utils.audio', 'backend.utils.cache', 'backend.utils.progress', 'backend.utils.hf_progress', 'backend.utils.validation', 'torch', 'transformers', 'fastapi', 'uvicorn', 'sqlalchemy', 'librosa', 'soundfile', 'qwen_tts', 'qwen_tts.inference', 'qwen_tts.inference.qwen3_tts_model', 'qwen_tts.inference.qwen3_tts_tokenizer', 'qwen_tts.core', 'qwen_tts.cli', 'pkg_resources.extern']
+datas += collect_data_files('qwen_tts')
+datas += copy_metadata('qwen-tts')
+hiddenimports += collect_submodules('qwen_tts')
+hiddenimports += collect_submodules('jaraco')
+
+
+a = Analysis(
+    ['server.py'],
+    pathex=[],
+    binaries=[],
+    datas=datas,
+    hiddenimports=hiddenimports,
+    hookspath=[],
+    hooksconfig={},
+    runtime_hooks=[],
+    excludes=[],
+    noarchive=False,
+    optimize=0,
+)
+pyz = PYZ(a.pure)
+
+exe = EXE(
+    pyz,
+    a.scripts,
+    a.binaries,
+    a.datas,
+    [],
+    name='voicebox-server-cuda',
+    debug=False,
+    bootloader_ignore_signals=False,
+    strip=False,
+    upx=True,
+    upx_exclude=[],
+    runtime_tmpdir=None,
+    console=True,
+    disable_windowed_traceback=False,
+    argv_emulation=False,
+    target_arch=None,
+    codesign_identity=None,
+    entitlements_file=None,
+)

backend/voicebox-server.spec

@@ -4,15 +4,11 @@ from PyInstaller.utils.hooks import collect_submodules
 from PyInstaller.utils.hooks import copy_metadata
 
 datas = []
-hiddenimports = ['backend', 'backend.main', 'backend.config', 'backend.database', 'backend.models', 'backend.profiles', 'backend.history', 'backend.tts', 'backend.transcribe', 'backend.platform_detect', 'backend.backends', 'backend.backends.pytorch_backend', 'backend.utils.audio', 'backend.utils.cache', 'backend.utils.progress', 'backend.utils.hf_progress', 'backend.utils.validation', 'torch', 'transformers', 'fastapi', 'uvicorn', 'sqlalchemy', 'librosa', 'soundfile', 'qwen_tts', 'qwen_tts.inference', 'qwen_tts.inference.qwen3_tts_model', 'qwen_tts.inference.qwen3_tts_tokenizer', 'qwen_tts.core', 'qwen_tts.cli', 'pkg_resources.extern', 'backend.backends.mlx_backend', 'mlx', 'mlx.core', 'mlx.nn', 'mlx_audio', 'mlx_audio.tts', 'mlx_audio.stt']
+hiddenimports = ['backend', 'backend.main', 'backend.config', 'backend.database', 'backend.models', 'backend.profiles', 'backend.history', 'backend.tts', 'backend.transcribe', 'backend.platform_detect', 'backend.backends', 'backend.backends.pytorch_backend', 'backend.utils.audio', 'backend.utils.cache', 'backend.utils.progress', 'backend.utils.hf_progress', 'backend.utils.validation', 'torch', 'transformers', 'fastapi', 'uvicorn', 'sqlalchemy', 'librosa', 'soundfile', 'qwen_tts', 'qwen_tts.inference', 'qwen_tts.inference.qwen3_tts_model', 'qwen_tts.inference.qwen3_tts_tokenizer', 'qwen_tts.core', 'qwen_tts.cli', 'pkg_resources.extern']
 datas += collect_data_files('qwen_tts')
-datas += collect_data_files('mlx')
-datas += collect_data_files('mlx_audio')
 datas += copy_metadata('qwen-tts')
 hiddenimports += collect_submodules('qwen_tts')
 hiddenimports += collect_submodules('jaraco')
-hiddenimports += collect_submodules('mlx')
-hiddenimports += collect_submodules('mlx_audio')
 
 
 a = Analysis(

docs/cuda-distribution-problem.md

@@ -0,0 +1,620 @@
+# CUDA Distribution Problem - Complete Analysis
+
+## Table of Contents
+1. [Problem Overview](#problem-overview)
+2. [Root Cause](#root-cause)
+3. [Attempted Solutions](#attempted-solutions)
+4. [Current Status](#current-status)
+5. [Available Options](#available-options)
+6. [Technical Details](#technical-details)
+7. [Cost Analysis](#cost-analysis)
+8. [Recommendations](#recommendations)
+
+---
+
+## Problem Overview
+
+### Timeline of Issues
+
+**Original Problem (v0.1.0 - v0.1.11)**
+- Single server binary with CUDA support
+- Size: ~2.9GB
+- Issue: MSI installer build fails in GitHub Actions CI
+- Error: WiX Toolset cannot handle 3GB files efficiently
+
+**First Solution: Dual Binary System (v0.1.12)**
+- Split into CPU (295MB) and CUDA (2.37GB) binaries
+- CPU ships with installer
+- CUDA as optional download
+- Issue: GitHub Release assets have 2GB limit
+
+**Current Problem (Discovered during implementation)**
+- GitHub Release Asset Limit: **2GB hard maximum**
+- CUDA binary: **2.37GB** (370MB over limit)
+- Cannot upload to GitHub Releases
+
+---
+
+## Root Cause
+
+### Why Is The CUDA Binary So Large?
+
+The size difference between CPU and CUDA builds:
+
+| Component | CPU Build | CUDA Build | Difference |
+|-----------|-----------|------------|------------|
+| PyTorch Core | ~150MB | ~150MB | - |
+| CPU Libraries (MKL/OpenBLAS) | ~100MB | - | -100MB |
+| CUDA Runtime | - | ~500MB | +500MB |
+| cuBLAS | - | ~350MB | +350MB |
+| cuDNN | - | ~1.2GB | +1.2GB |
+| NVRTC (CUDA Compiler) | - | ~90MB | +90MB |
+| Other CUDA libs | - | ~100MB | +100MB |
+| **Total** | **~295MB** | **~2.37GB** | **+2.07GB** |
+
+### CUDA Dependencies Breakdown
+
+```
+torch/lib/ (CUDA build):
+├── cudart64_12.dll          (~0.5 MB)   - CUDA Runtime
+├── cublas64_12.dll          (~100 MB)   - Basic Linear Algebra
+├── cublasLt64_12.dll        (~200 MB)   - Linear Algebra (optimized)
+├── cudnn64_9.dll            (~800 MB)   - Deep Neural Networks
+├── cudnn_*_infer64_9.dll    (~400 MB)   - DNN Inference ops
+├── nvrtc64_*.dll            (~50 MB)    - Runtime Compiler
+├── nvrtc-builtins64_*.dll   (~40 MB)    - Compiler builtins
+├── torch_cuda.dll           (~200 MB)   - PyTorch CUDA bridge
+└── c10_cuda.dll             (~20 MB)    - Core CUDA utilities
+```
+
+**Why These Are Required:**
+- cuDNN is essential for neural network operations
+- cuBLAS handles all matrix operations (core of ML)
+- Cannot split or remove without breaking functionality
+
+---
+
+## Attempted Solutions
+
+### Solution 1: Dual Binary System ✅ (Partially Successful)
+
+**Goal**: Split CPU and CUDA into separate downloads
+
+**Implementation**:
+```bash
+# Build CPU-only (295MB)
+pip install torch --index-url https://download.pytorch.org/whl/cpu
+python build_binary.py cpu
+
+# Build CUDA (2.37GB)
+pip install torch --index-url https://download.pytorch.org/whl/cu121
+python build_binary.py cuda
+```
+
+**Results**:
+- ✅ CPU binary: 295MB (fits in installer)
+- ✅ CI builds successfully
+- ✅ Installer size reduced from 3GB to ~500MB
+- ❌ CUDA binary still too large for GitHub
+
+**See**: `docs/dual-server-binaries.md`
+
+### Solution 2: Compression Testing ❌ (Failed)
+
+**Goal**: Compress CUDA binary to fit under 2GB
+
+**Method**: 7z with maximum compression settings
+```bash
+7z a -t7z -m0=lzma2 -mx=9 -mfb=64 -md=32m -ms=on \
+  voicebox-server-cuda.7z voicebox-server-cuda.exe
+```
+
+**Results**:
+```
+Original:     2.37 GB (2,545,086,396 bytes)
+Compressed:   2.35 GB (2,519,381,264 bytes)
+Compression:  1.0% (only 24.5MB saved)
+GitHub Limit: 2.00 GB (2,147,483,648 bytes)
+Over by:      354.67 MB
+
+Status: FAILED - Still exceeds limit by 354MB
+```
+
+**Why Compression Failed**:
+- CUDA binaries are already optimized machine code
+- No redundant data to compress
+- Neural network kernels are highly compact
+- Libraries are already stripped of debug symbols
+
+**Conclusion**: Compression is not viable
+
+---
+
+## Current Status
+
+### What Works
+- ✅ CPU binary builds successfully (295MB)
+- ✅ CUDA binary builds successfully (2.37GB)
+- ✅ Build scripts for both variants
+- ✅ CI workflow updated for dual binaries
+- ✅ Installer can be created with CPU binary
+
+### What Doesn't Work
+- ❌ Cannot upload CUDA binary to GitHub Releases (exceeds 2GB limit)
+- ❌ Compression doesn't reduce size enough
+- ❌ No automated distribution path for CUDA binary
+
+### Branch Status
+- Branch: `feat/dual-server-binaries`
+- Commits: Implementation complete
+- Testing: Local builds successful
+- Blocker: CUDA distribution path
+
+---
+
+## Available Options
+
+### Option 1: AWS S3 Hosting (Recommended)
+
+**Description**: Host CUDA binary in Amazon S3 bucket
+
+**Pros**:
+- ✅ No file size limits (can handle multi-GB files)
+- ✅ Fast global CDN (CloudFront)
+- ✅ Reliable (99.99% uptime)
+- ✅ Pay only for usage
+- ✅ Easy CI integration
+- ✅ Version control (keep multiple releases)
+
+**Cons**:
+- ❌ Requires AWS account
+- ❌ Monthly costs (~$1-5/month)
+- ❌ Additional infrastructure to manage
+
+**Cost Estimate**:
+```
+Storage: 2.37 GB × $0.023/GB = $0.05/month
+Transfer: 100 downloads × 2.37GB × $0.09/GB = $21.33/month
+Total: ~$21-25/month for 100 downloads
+      ~$2-5/month for 10-20 downloads
+```
+
+**Implementation**:
+```yaml
+# .github/workflows/release.yml
+- name: Upload CUDA to S3
+  env:
+    AWS_ACCESS_KEY_ID: ${{ secrets.AWS_ACCESS_KEY_ID }}
+    AWS_SECRET_ACCESS_KEY: ${{ secrets.AWS_SECRET_ACCESS_KEY }}
+  run: |
+    aws s3 cp backend/cuda-release/voicebox-server-cuda-*.exe \
+      s3://voicebox-releases/cuda/${{ github.ref_name }}/ \
+      --acl public-read
+
+    # Generate download URL
+    echo "CUDA_URL=https://voicebox-releases.s3.amazonaws.com/cuda/${{ github.ref_name }}/voicebox-server-cuda-x86_64-pc-windows-msvc.exe" >> release_notes.txt
+```
+
+**User Experience**:
+1. Install app normally (500MB installer)
+2. App detects NVIDIA GPU
+3. Shows: "Download CUDA support? (2.4GB)"
+4. Downloads from S3: `https://voicebox-releases.s3.amazonaws.com/cuda/v0.1.12/voicebox-server-cuda.exe`
+5. Saves to `%APPDATA%/voicebox/binaries/`
+6. App restarts with CUDA server
+
+---
+
+### Option 2: Azure Blob Storage
+
+**Description**: Microsoft Azure alternative to S3
+
+**Pros**:
+- ✅ Similar to S3 (no size limits, CDN, reliable)
+- ✅ Good if already using Azure
+- ✅ Competitive pricing
+- ✅ Global CDN with Azure CDN
+
+**Cons**:
+- ❌ Requires Azure account
+- ❌ Similar monthly costs
+- ❌ Less common in open source projects
+
+**Cost Estimate**:
+```
+Storage: $0.018/GB = $0.04/month
+Transfer: ~$20-25/month for 100 downloads
+```
+
+**Implementation**:
+```yaml
+- name: Upload to Azure Blob
+  env:
+    AZURE_STORAGE_CONNECTION_STRING: ${{ secrets.AZURE_STORAGE }}
+  run: |
+    az storage blob upload \
+      --account-name voiceboxreleases \
+      --container-name cuda-binaries \
+      --name v${{ github.ref_name }}/voicebox-server-cuda.exe \
+      --file backend/cuda-release/voicebox-server-cuda-*.exe \
+      --tier Hot
+```
+
+---
+
+### Option 3: Cloudflare R2
+
+**Description**: Cloudflare's S3-compatible object storage
+
+**Pros**:
+- ✅ S3-compatible API
+- ✅ **FREE egress (no bandwidth charges!)**
+- ✅ Cheaper than S3/Azure
+- ✅ Cloudflare CDN included
+- ✅ Good for open source projects
+
+**Cons**:
+- ❌ Requires Cloudflare account
+- ❌ Newer service (less mature than S3)
+
+**Cost Estimate**:
+```
+Storage: $0.015/GB = $0.04/month
+Egress: $0.00 (FREE!)
+Class A ops: Negligible
+Total: ~$0.04/month (essentially free!)
+```
+
+**Why This Is Attractive**:
+- Zero bandwidth costs (huge savings)
+- Perfect for open source distribution
+- S3-compatible (easy migration if needed)
+
+**Implementation**:
+Same as S3 (R2 is S3-compatible):
+```yaml
+- name: Upload to R2
+  env:
+    AWS_ACCESS_KEY_ID: ${{ secrets.R2_ACCESS_KEY_ID }}
+    AWS_SECRET_ACCESS_KEY: ${{ secrets.R2_SECRET_ACCESS_KEY }}
+    AWS_ENDPOINT_URL: https://<account-id>.r2.cloudflarestorage.com
+  run: |
+    aws s3 cp backend/cuda-release/voicebox-server-cuda-*.exe \
+      s3://voicebox-releases/cuda/${{ github.ref_name }}/ \
+      --endpoint-url=$AWS_ENDPOINT_URL
+```
+
+---
+
+### Option 4: GitHub Packages (Container Registry)
+
+**Description**: Package CUDA binary as OCI/Docker artifact
+
+**Pros**:
+- ✅ Stays in GitHub ecosystem
+- ✅ No additional accounts needed
+- ✅ Free for public repos
+
+**Cons**:
+- ❌ Complex for desktop app distribution
+- ❌ Users need to extract from container
+- ❌ Awkward UX (not designed for binary distribution)
+- ❌ Requires Docker understanding
+
+**Not Recommended**: Containers aren't designed for desktop app binaries
+
+---
+
+### Option 5: Self-Hosted Server
+
+**Description**: Host on your own VPS/server
+
+**Pros**:
+- ✅ Full control
+- ✅ No cloud provider dependency
+- ✅ Predictable costs
+
+**Cons**:
+- ❌ Requires server maintenance
+- ❌ Bandwidth costs can be high
+- ❌ Uptime responsibility
+- ❌ Scaling challenges
+
+**Cost Estimate**:
+```
+VPS: $5-20/month (DigitalOcean, Linode)
+Bandwidth: $0.01-0.02/GB
+Total: $10-50/month depending on traffic
+```
+
+---
+
+### Option 6: Manual Distribution
+
+**Description**: Don't automate - provide manual download instructions
+
+**Pros**:
+- ✅ Zero cost
+- ✅ Zero infrastructure
+- ✅ Simple
+
+**Cons**:
+- ❌ Poor user experience
+- ❌ Manual upload to file host each release
+- ❌ Users must manually download and install
+- ❌ No automatic updates for CUDA binary
+- ❌ Increases support burden
+
+**Implementation**:
+```
+Release notes:
+"Windows users with NVIDIA GPUs can download CUDA support:
+1. Download voicebox-server-cuda.exe from [Google Drive/Mega/etc]
+2. Place in C:\Users\<YourName>\AppData\Roaming\voicebox\binaries\
+3. Restart the app"
+```
+
+**Not Recommended**: Creates friction, support issues
+
+---
+
+### Option 7: Split CUDA Binary
+
+**Description**: Break CUDA binary into multiple <2GB chunks
+
+**Technical Approach**:
+```python
+# Split binary
+split -b 2000M voicebox-server-cuda.exe cuda_part_
+
+# Upload parts to GitHub (each <2GB)
+cuda_part_aa (2.0 GB)
+cuda_part_ab (0.37 GB)
+
+# App downloads and reassembles
+cat cuda_part_* > voicebox-server-cuda.exe
+```
+
+**Pros**:
+- ✅ Stays on GitHub
+- ✅ No external hosting
+
+**Cons**:
+- ❌ Complex download logic (multiple files)
+- ❌ Integrity checking required
+- ❌ More points of failure
+- ❌ Users must wait for multiple downloads
+- ❌ Still hacky solution
+
+**Complexity**: Medium-High
+
+---
+
+## Technical Details
+
+### Current Build Output
+
+```
+backend/dist/
+├── voicebox-server.exe          295 MB  (CPU-only)
+└── voicebox-server-cuda.exe    2.37 GB  (CUDA)
+
+# After compression test:
+backend/dist/
+└── voicebox-server-cuda.7z     2.35 GB  (not viable)
+```
+
+### CI Workflow Changes Required
+
+For external hosting (S3/R2/Azure):
+
+```yaml
+# Current workflow (fails)
+- name: Upload CUDA server binary (Windows only)
+  if: matrix.platform == 'windows-latest'
+  uses: softprops/action-gh-release@v1
+  with:
+    files: backend/cuda-release/voicebox-server-cuda-*.exe  # ❌ Fails: >2GB
+    draft: true
+
+# New workflow (S3 example)
+- name: Upload CUDA to S3 (Windows only)
+  if: matrix.platform == 'windows-latest'
+  env:
+    AWS_ACCESS_KEY_ID: ${{ secrets.AWS_ACCESS_KEY_ID }}
+    AWS_SECRET_ACCESS_KEY: ${{ secrets.AWS_SECRET_ACCESS_KEY }}
+  run: |
+    aws s3 cp backend/cuda-release/voicebox-server-cuda-*.exe \
+      s3://voicebox-releases/cuda/${{ github.ref_name }}/ \
+      --acl public-read
+
+    # Generate release notes with download URL
+    cat >> release_notes.md <<EOF
+
+    ### GPU Acceleration (Windows)
+    Download CUDA support for NVIDIA GPUs:
+    [voicebox-server-cuda.exe](https://voicebox-releases.s3.amazonaws.com/cuda/${{ github.ref_name }}/voicebox-server-cuda-x86_64-pc-windows-msvc.exe)
+    Size: 2.37 GB
+    EOF
+```
+
+### App Changes Required
+
+**Frontend (Tauri)**: Download manager
+```typescript
+// src/lib/cuda-downloader.ts
+const CUDA_DOWNLOAD_URL =
+  "https://voicebox-releases.s3.amazonaws.com/cuda/v{VERSION}/voicebox-server-cuda.exe";
+
+async function downloadCudaBinary(version: string) {
+  const url = CUDA_DOWNLOAD_URL.replace("{VERSION}", version);
+  const savePath = path.join(app.getPath("userData"), "binaries", "voicebox-server-cuda.exe");
+
+  // Download with progress
+  await downloadFile(url, savePath, (progress) => {
+    // Update UI: "Downloading CUDA support: 45% (1.2GB / 2.4GB)"
+  });
+
+  // Verify checksum
+  const checksum = await calculateChecksum(savePath);
+  if (checksum !== EXPECTED_CHECKSUM) {
+    throw new Error("Download corrupted");
+  }
+}
+```
+
+**Backend**: Already supports both binaries (no changes needed)
+
+---
+
+## Cost Analysis
+
+### Monthly Cost Comparison (100 downloads/month)
+
+| Option | Storage | Bandwidth | Total/Month | Notes |
+|--------|---------|-----------|-------------|-------|
+| **Cloudflare R2** | $0.04 | $0.00 | **$0.04** | Best for open source |
+| AWS S3 | $0.05 | $21.33 | $21.38 | Good reliability |
+| Azure Blob | $0.04 | $20.00 | $20.04 | Azure ecosystem |
+| Self-hosted VPS | $10.00 | $2.37 | $12.37 | Maintenance overhead |
+| Manual | $0.00 | $0.00 | $0.00 | Poor UX |
+
+### Annual Cost Comparison
+
+| Option | Year 1 | Year 2+ | Notes |
+|--------|--------|---------|-------|
+| **Cloudflare R2** | **$0.50** | **$0.50** | Essentially free |
+| AWS S3 | $256 | $256 | Predictable |
+| Self-hosted | $144 | $144 | Time cost |
+
+**Recommendation**: Cloudflare R2 (free egress = huge savings)
+
+---
+
+## Recommendations
+
+### Recommended Solution: Cloudflare R2
+
+**Why**:
+1. **Cost**: Essentially free (~$0.04/month)
+2. **Bandwidth**: Zero egress charges (unlimited downloads)
+3. **CDN**: Cloudflare's global network included
+4. **Compatibility**: S3-compatible API (easy to use)
+5. **Perfect for open source**: No surprise bandwidth bills
+
+### Implementation Priority
+
+**Phase 1: Setup (1-2 hours)**
+1. Create Cloudflare R2 account
+2. Create bucket: `voicebox-releases`
+3. Generate API credentials
+4. Add to GitHub Secrets
+
+**Phase 2: CI Integration (1-2 hours)**
+1. Update `.github/workflows/release.yml`
+2. Add R2 upload step
+3. Generate release notes with download URL
+4. Test with draft release
+
+**Phase 3: App Integration (4-6 hours)**
+1. Add GPU detection on startup
+2. Implement download manager UI
+3. Add progress indicators
+4. Implement checksum verification
+5. Server restart logic
+
+**Phase 4: Documentation (1 hour)**
+1. Update README with GPU instructions
+2. Add troubleshooting guide
+3. Document manual download process
+
+**Total Time**: ~8-12 hours of development
+
+### Alternative: AWS S3 (If Already Using AWS)
+
+If you're already using AWS for other infrastructure, S3 is also a solid choice:
+- More mature than R2
+- Extensive documentation
+- Familiar tooling
+- ~$20/month for moderate usage
+
+---
+
+## Open Questions
+
+1. **Expected Download Volume**: How many CUDA downloads per month?
+   - Affects cost calculations
+   - Determines if R2's free egress is significant
+
+2. **Update Strategy**: How to handle CUDA updates?
+   - Option A: Version in URL path (keep all versions)
+   - Option B: Overwrite latest (save space)
+
+3. **Fallback Strategy**: What if cloud provider is down?
+   - Mirror on multiple providers?
+   - Graceful degradation to CPU?
+
+4. **Telemetry**: Track CUDA download stats?
+   - Helps with cost forecasting
+   - User behavior insights
+
+---
+
+## Next Steps
+
+1. **Research Phase** (You are here)
+   - Evaluate cloud providers
+   - Check terms of service
+   - Test account creation
+
+2. **Decision Phase**
+   - Choose provider (Cloudflare R2 recommended)
+   - Set up account
+   - Configure billing alerts
+
+3. **Implementation Phase**
+   - Update CI workflow
+   - Implement download manager
+   - Test end-to-end flow
+
+4. **Launch Phase**
+   - Deploy to production
+   - Monitor downloads
+   - Gather user feedback
+
+---
+
+## References
+
+- **GitHub Release Limits**: https://docs.github.com/en/repositories/releasing-projects-on-github/about-releases
+- **Cloudflare R2 Pricing**: https://developers.cloudflare.com/r2/pricing/
+- **AWS S3 Pricing**: https://aws.amazon.com/s3/pricing/
+- **Compression Test Results**: `backend/test_cuda_compression.py`
+- **Dual Binary Implementation**: `docs/dual-server-binaries.md`
+
+---
+
+## Appendix: Alternative Approaches Considered
+
+### A. Dynamic CUDA Loading
+**Idea**: Load CUDA DLLs dynamically at runtime
+**Why Not**: PyTorch requires CUDA DLLs at import time, can't lazy-load
+
+### B. CUDA as Separate Package
+**Idea**: Python package with just CUDA libs
+**Why Not**: Still 2GB+, same problem
+
+### C. Model Quantization
+**Idea**: Use smaller quantized models
+**Why Not**: Doesn't reduce CUDA runtime size
+
+### D. Docker Distribution
+**Idea**: Distribute as Docker container
+**Why Not**: Poor fit for desktop app, requires Docker installed
+
+---
+
+**Document Version**: 1.0
+**Last Updated**: 2026-01-31
+**Status**: Research Phase
+**Next Review**: After cloud provider decision

docs/dual-server-binaries.md

@@ -0,0 +1,177 @@
+# Dual Server Binary System
+
+## Overview
+
+Voicebox now uses a dual-binary approach to manage the size difference between CPU-only and CUDA-enabled builds:
+
+- **CPU Binary** (~500MB): Ships with the installer by default
+- **CUDA Binary** (~3GB): Downloaded on-demand for GPU users
+
+## Problem Solved
+
+Previously, bundling PyTorch with CUDA support created a 3GB server binary, which:
+- Made the installer too large (failed CI builds with WiX)
+- Forced all users to download CUDA libraries even without NVIDIA GPUs
+- Created poor user experience
+
+## Solution
+
+### Build Process
+
+**Two separate binaries are built:**
+
+1. **voicebox-server.exe** (CPU)
+   - Built with: `pip install torch --index-url https://download.pytorch.org/whl/cpu`
+   - Size: ~500MB
+   - Works on all Windows machines
+   - Included in the installer by default
+
+2. **voicebox-server-cuda.exe** (CUDA)
+   - Built with: `pip install torch --index-url https://download.pytorch.org/whl/cu121`
+   - Size: ~3GB
+   - Requires NVIDIA GPU + drivers
+   - Uploaded as separate GitHub Release asset
+
+### User Experience
+
+**First Launch:**
+1. User installs app (~500MB download)
+2. App starts with CPU server
+3. If NVIDIA GPU detected:
+   - Show notification: "Download CUDA support for 4-5x faster inference?"
+   - User clicks "Download"
+   - Download voicebox-server-cuda.exe from GitHub (~3GB)
+   - Save to `%APPDATA%/voicebox/binaries/`
+   - Restart server with CUDA version
+
+**Settings Panel:**
+- Toggle between CPU/CUDA modes
+- Download CUDA if not already installed
+- Show current inference backend
+
+### Build Scripts
+
+**Windows:**
+```bash
+cd backend
+
+# Build CPU only
+build_cpu.bat
+
+# Build CUDA only
+build_cuda.bat
+
+# Build both
+build_both.bat
+```
+
+**Unix (macOS/Linux):**
+```bash
+cd backend
+
+# Build CPU only
+./build_cpu.sh
+```
+
+### CI/CD Workflow
+
+**GitHub Actions (.github/workflows/release.yml):**
+
+1. Install CPU PyTorch
+2. Build CPU server → Copy to Tauri binaries
+3. Install CUDA PyTorch
+4. Build CUDA server → Save for upload
+5. Build Tauri app (bundles CPU server)
+6. Upload CUDA server as separate release asset
+
+### File Structure
+
+```
+Release Assets:
+├── Voicebox_0.1.12_x64_en-US.msi           (~500MB - includes CPU server)
+├── voicebox-server-cuda-x86_64-pc-windows-msvc.exe   (~3GB - optional download)
+└── latest.json                              (updater manifest)
+```
+
+## Implementation Details
+
+### Modified Files
+
+1. **backend/build_binary.py**
+   - Added `variant` parameter ('cpu' or 'cuda')
+   - Outputs different binary names based on variant
+
+2. **backend/build_cpu.bat** (new)
+   - Installs CPU PyTorch
+   - Builds CPU binary
+   - Restores CUDA PyTorch for dev
+
+3. **backend/build_cuda.bat** (new)
+   - Ensures CUDA PyTorch is installed
+   - Builds CUDA binary
+
+4. **.github/workflows/release.yml**
+   - Build CPU binary first (for installer)
+   - Build CUDA binary second (for upload)
+   - Upload CUDA binary as additional release asset
+   - Updated release notes to explain GPU acceleration
+
+### Future Frontend Work
+
+**TODO: Implement CUDA download in the app**
+
+Location: `tauri/src/`
+
+Features needed:
+1. GPU detection on startup
+2. Download manager for CUDA binary
+3. Server binary path switcher
+4. Settings UI for CPU/CUDA toggle
+5. Progress indicator for 3GB download
+
+API endpoints needed (already exist):
+- `/health` - Shows GPU availability
+- Server restart mechanism
+
+## Benefits
+
+✓ **Smaller installer**: ~500MB instead of 3GB
+✓ **Faster CI builds**: WiX can handle 500MB easily
+✓ **User choice**: CPU users don't download unnecessary files
+✓ **Better UX**: Optional performance upgrade for GPU users
+✓ **Cost savings**: Reduced bandwidth for users without GPUs
+
+## Testing
+
+**Test CPU build:**
+```bash
+cd backend
+python build_binary.py cpu
+./dist/voicebox-server.exe --version
+```
+
+**Test CUDA build:**
+```bash
+cd backend
+python build_binary.py cuda
+./dist/voicebox-server-cuda.exe --version
+```
+
+**Verify size:**
+```bash
+ls -lh backend/dist/
+# Should see:
+# voicebox-server.exe       ~500MB
+# voicebox-server-cuda.exe  ~3GB
+```
+
+**Test server startup:**
+```bash
+# CPU version
+./backend/dist/voicebox-server.exe
+# Check logs: Should show CPU inference
+
+# CUDA version (requires NVIDIA GPU)
+./backend/dist/voicebox-server-cuda.exe
+# Check logs: Should show CUDA inference
+```

docs/github-2gb-limit-issue.md

@@ -0,0 +1,122 @@
+# GitHub 2GB Release Asset Limit Issue
+
+## Problem
+
+The CUDA server binary upload fails in CI with:
+```
+Error: File size (2543828017) is greater than 2 GiB
+```
+
+GitHub release assets have a hard limit of 2GB per file. Our CUDA binary is ~2.5GB, which exceeds this limit.
+
+## Background
+
+The dual-server binary system (see `dual-server-binaries.md`) creates two binaries:
+- **CPU binary**: ~500MB ✅ Works fine
+- **CUDA binary**: ~2.5GB ❌ Exceeds GitHub limit
+
+## Attempted Solution: Compression
+
+We're testing 7z compression with maximum settings to see if we can squeeze the CUDA binary under 2GB.
+
+### Test Script
+
+Run `backend/test_cuda_compression.py` to test compression locally:
+
+```bash
+cd backend
+python test_cuda_compression.py
+```
+
+This will:
+1. Find the CUDA binary in `dist/`
+2. Compress it with 7z (maximum compression)
+3. Report if the compressed size fits under 2GB
+
+### Expected Compression
+
+PyTorch CUDA binaries typically compress well since they contain:
+- Repeated patterns in neural network weights
+- Debug symbols and metadata
+- Redundant CUDA libraries
+
+Estimated compression: 30-40% reduction
+- Original: ~2.5GB
+- Target: <2GB
+- Required compression: >20%
+
+## Fallback: External Hosting
+
+If compression doesn't work, we'll need to host the CUDA binary externally:
+
+### Option 1: AWS S3
+```yaml
+- name: Upload CUDA binary to S3
+  run: |
+    aws s3 cp backend/cuda-release/voicebox-server-cuda-*.exe \
+      s3://voicebox-releases/cuda-binaries/${{ github.ref_name }}/
+```
+
+### Option 2: Azure Blob Storage
+```yaml
+- name: Upload to Azure Blob
+  run: |
+    az storage blob upload \
+      --account-name voiceboxreleases \
+      --container-name cuda-binaries \
+      --file backend/cuda-release/voicebox-server-cuda-*.exe
+```
+
+### Option 3: GitHub Packages (Container Registry)
+Package as a container image, though this adds complexity for desktop app distribution.
+
+## Implementation Plan
+
+1. **Test compression locally** ← Current step
+2. **If compression works (<2GB)**:
+   - Update CI to compress before upload
+   - Update app to handle .7z downloads
+   - Add extraction step in download manager
+
+3. **If compression fails (≥2GB)**:
+   - Set up external storage (likely S3)
+   - Update CI to upload to S3
+   - Provide download URL in release notes
+   - Update app download manager to fetch from S3
+
+## CI Workflow Changes (if compression works)
+
+```yaml
+- name: Compress CUDA binary (Windows only)
+  if: matrix.platform == 'windows-latest'
+  shell: bash
+  run: |
+    cd backend/cuda-release
+    7z a -t7z -m0=lzma2 -mx=9 -mfb=64 -md=32m -ms=on \
+      voicebox-server-cuda-x86_64-pc-windows-msvc.7z \
+      voicebox-server-cuda-*.exe
+
+- name: Upload compressed CUDA server (Windows only)
+  if: matrix.platform == 'windows-latest'
+  uses: softprops/action-gh-release@v1
+  with:
+    files: backend/cuda-release/*.7z
+```
+
+## User Experience Impact
+
+### With Compression
+- Download: `voicebox-server-cuda-*.7z` (~1.5-1.8GB)
+- App extracts automatically
+- One extra step but manageable
+
+### With External Hosting
+- Download from S3/Azure URL
+- No GitHub release asset dependency
+- Potentially faster download speeds (CDN)
+
+## Status
+
+🔄 **Testing compression locally to determine viability**
+
+Results pending from local test run.

docs/r2-setup.md

@@ -0,0 +1,274 @@
+# Cloudflare R2 Setup Guide
+
+## Overview
+
+The CUDA binary (2.4GB) is hosted on Cloudflare R2 at `downloads.voicebox.sh` instead of GitHub Releases (which has a 2GB limit).
+
+## R2 Bucket Configuration
+
+✅ **Completed:**
+- Bucket created: `voicebox`
+- Custom domain configured: `downloads.voicebox.sh`
+
+## GitHub Secrets Required
+
+Add these secrets to your GitHub repository:
+
+### 1. R2_ACCESS_KEY_ID
+
+Your Cloudflare R2 API Access Key ID
+
+**How to get it:**
+1. Go to Cloudflare Dashboard → R2
+2. Click "Manage R2 API Tokens"
+3. Create API Token with "Object Read & Write" permissions
+4. Copy the "Access Key ID"
+
+**Add to GitHub:**
+```
+Repository Settings → Secrets and variables → Actions → New repository secret
+Name: R2_ACCESS_KEY_ID
+Value: <your-access-key-id>
+```
+
+### 2. R2_SECRET_ACCESS_KEY
+
+Your Cloudflare R2 Secret Access Key
+
+**How to get it:**
+- Same process as above
+- Copy the "Secret Access Key" (shown only once!)
+- Store it securely
+
+**Add to GitHub:**
+```
+Name: R2_SECRET_ACCESS_KEY
+Value: <your-secret-access-key>
+```
+
+### 3. R2_ENDPOINT
+
+Your Cloudflare R2 endpoint URL
+
+**Format:**
+```
+https://<account-id>.r2.cloudflarestorage.com
+```
+
+**How to find your account ID:**
+- Cloudflare Dashboard → R2
+- Look at the URL or bucket settings
+- Should be a string of letters/numbers
+
+**Add to GitHub:**
+```
+Name: R2_ENDPOINT
+Value: https://<your-account-id>.r2.cloudflarestorage.com
+```
+
+## Bucket Structure
+
+After CI uploads, the bucket will have this structure:
+
+```
+voicebox/
+└── cuda/
+    ├── v0.1.12/
+    │   └── voicebox-server-cuda-x86_64-pc-windows-msvc.exe
+    ├── v0.1.13/
+    │   └── voicebox-server-cuda-x86_64-pc-windows-msvc.exe
+    └── v0.2.0/
+        └── voicebox-server-cuda-x86_64-pc-windows-msvc.exe
+```
+
+## Public Access
+
+Files are uploaded with `--acl public-read`, making them accessible at:
+
+```
+https://downloads.voicebox.sh/cuda/v{VERSION}/voicebox-server-cuda-x86_64-pc-windows-msvc.exe
+```
+
+**Example:**
+```
+https://downloads.voicebox.sh/cuda/v0.1.12/voicebox-server-cuda-x86_64-pc-windows-msvc.exe
+```
+
+## Testing the Setup
+
+### Local Test Upload
+
+Before running the CI, test uploading locally:
+
+```bash
+# Set environment variables
+export AWS_ACCESS_KEY_ID="your-r2-access-key-id"
+export AWS_SECRET_ACCESS_KEY="your-r2-secret-access-key"
+export R2_ENDPOINT="https://your-account-id.r2.cloudflarestorage.com"
+
+# Install AWS CLI
+pip install awscli
+
+# Test upload (use a small test file first)
+echo "test" > test.txt
+aws s3 cp test.txt \
+  s3://voicebox/test/test.txt \
+  --endpoint-url $R2_ENDPOINT \
+  --acl public-read
+
+# Verify it's accessible
+curl https://downloads.voicebox.sh/test/test.txt
+
+# If successful, try the actual CUDA binary
+aws s3 cp backend/dist/voicebox-server-cuda.exe \
+  s3://voicebox/cuda/v0.1.12-test/voicebox-server-cuda-x86_64-pc-windows-msvc.exe \
+  --endpoint-url $R2_ENDPOINT \
+  --acl public-read
+```
+
+### Verify Upload
+
+Check if the file is accessible:
+
+```bash
+curl -I https://downloads.voicebox.sh/cuda/v0.1.12-test/voicebox-server-cuda-x86_64-pc-windows-msvc.exe
+```
+
+Should return:
+```
+HTTP/2 200
+content-length: 2545086396
+content-type: application/x-msdownload
+...
+```
+
+## CI Workflow
+
+The workflow now:
+
+1. **Builds CPU binary** → Includes in installer
+2. **Builds CUDA binary** → Uploads to R2
+3. **Release notes** → Include R2 download link
+
+### CI Steps (Windows)
+
+```yaml
+- name: Build CUDA Python server (Windows only)
+  # Builds the CUDA binary
+
+- name: Upload CUDA server to Cloudflare R2 (Windows only)
+  env:
+    AWS_ACCESS_KEY_ID: ${{ secrets.R2_ACCESS_KEY_ID }}
+    AWS_SECRET_ACCESS_KEY: ${{ secrets.R2_SECRET_ACCESS_KEY }}
+    R2_ENDPOINT: ${{ secrets.R2_ENDPOINT }}
+  run: |
+    aws s3 cp backend/cuda-release/voicebox-server-cuda-*.exe \
+      s3://voicebox/cuda/${VERSION}/... \
+      --endpoint-url $R2_ENDPOINT \
+      --acl public-read
+```
+
+## Cost Tracking
+
+Monitor your R2 usage:
+
+**Cloudflare Dashboard → R2 → voicebox → Metrics**
+
+Expected costs (per month):
+- Storage: 2.4GB × $0.015/GB = **$0.036**
+- Egress: **$0.00** (free!)
+- Class A ops: ~100 × $4.50/million = **$0.00**
+
+**Total: ~$0.04/month** (essentially free!)
+
+## Troubleshooting
+
+### Upload fails: "Access Denied"
+
+**Solution:** Check API token permissions
+- Must have "Object Read & Write" on the bucket
+- Regenerate token if needed
+
+### File not accessible at downloads.voicebox.sh
+
+**Solution:** Check custom domain configuration
+- R2 Dashboard → Bucket → Settings → Custom Domains
+- Ensure `downloads.voicebox.sh` is properly configured
+- DNS may take time to propagate
+
+### "endpoint-url" not recognized
+
+**Solution:** Make sure AWS CLI is updated
+```bash
+pip install --upgrade awscli
+```
+
+### File uploaded but wrong permissions
+
+**Solution:** Re-upload with `--acl public-read`
+```bash
+aws s3 cp ... --acl public-read
+```
+
+Or set bucket default permissions in R2 Dashboard.
+
+## Security Notes
+
+### API Token Permissions
+
+✅ **Recommended:**
+- Object Read & Write only
+- No admin permissions needed
+- Scoped to `voicebox` bucket only
+
+❌ **Avoid:**
+- Account-wide permissions
+- Account admin access
+- Worker edit permissions
+
+### Secret Rotation
+
+Rotate API tokens every 6-12 months:
+1. Create new API token
+2. Update GitHub secrets
+3. Verify CI still works
+4. Delete old token
+
+## Maintenance
+
+### Cleaning Old Versions
+
+Optional: Delete old CUDA binaries to save storage costs
+
+```bash
+# List all versions
+aws s3 ls s3://voicebox/cuda/ \
+  --endpoint-url $R2_ENDPOINT
+
+# Delete old version
+aws s3 rm s3://voicebox/cuda/v0.1.0/ \
+  --recursive \
+  --endpoint-url $R2_ENDPOINT
+```
+
+### Monitoring
+
+Set up Cloudflare notifications:
+- Storage approaching limits
+- Unusual traffic patterns
+- High operation counts
+
+## Next Steps
+
+1. ✅ Bucket configured
+2. ⏳ Add GitHub secrets (R2_ACCESS_KEY_ID, R2_SECRET_ACCESS_KEY, R2_ENDPOINT)
+3. ⏳ Test local upload
+4. ⏳ Push branch and create test release
+5. ⏳ Verify CUDA binary accessible from downloads.voicebox.sh
+6. ⏳ Implement frontend download manager
+
+---
+
+**Status**: Ready for testing
+**Cost**: ~$0.04/month
+**Bandwidth**: Free (unlimited)

landing/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@voicebox/landing",
-  "version": "0.1.11",
+  "version": "0.1.12",
   "description": "Landing page for voicebox.sh",
   "scripts": {
     "dev": "bun --bun next dev --turbo",

package.json

@@ -1,6 +1,6 @@
 {
   "name": "voicebox",
-  "version": "0.1.11",
+  "version": "0.1.12",
   "private": true,
   "workspaces": [
     "app",

tauri/package.json

@@ -1,7 +1,7 @@
 {
   "name": "@voicebox/tauri",
   "private": true,
-  "version": "0.1.11",
+  "version": "0.1.12",
   "type": "module",
   "scripts": {
     "dev": "vite",

tauri/src-tauri/Cargo.lock

@@ -5041,7 +5041,7 @@ checksum = "0b928f33d975fc6ad9f86c8f283853ad26bdd5b10b7f1542aa2fa15e2289105a"
 
 [[package]]
 name = "voicebox"
-version = "0.1.11"
+version = "0.1.12"
 dependencies = [
  "base64 0.22.1",
  "core-foundation-sys",

tauri/src-tauri/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "voicebox"
-version = "0.1.11"
+version = "0.1.12"
 description = "A production-quality desktop app for Qwen3-TTS voice cloning and generation"
 authors = ["you"]
 license = ""

tauri/src-tauri/tauri.conf.json

@@ -1,7 +1,7 @@
 {
   "$schema": "https://schema.tauri.app/config/2",
   "productName": "Voicebox",
-  "version": "0.1.11",
+  "version": "0.1.12",
   "identifier": "sh.voicebox.app",
   "build": {
     "beforeDevCommand": "bun run dev",

tauri/vite.config.ts

@@ -35,5 +35,15 @@ export default defineConfig({
     minify: !process.env.TAURI_DEBUG,
     sourcemap: !!process.env.TAURI_DEBUG,
     outDir: 'dist',
+    rollupOptions: {
+      external: [
+        '@tauri-apps/api',
+        '@tauri-apps/plugin-dialog',
+        '@tauri-apps/plugin-fs',
+        '@tauri-apps/plugin-process',
+        '@tauri-apps/plugin-shell',
+        '@tauri-apps/plugin-updater',
+      ],
+    },
   },
 });

test_cuda.py

@@ -0,0 +1,77 @@
+"""Test CUDA detection in voicebox backend"""
+import sys
+import torch
+
+print("=" * 60)
+print("PyTorch CUDA Detection Test")
+print("=" * 60)
+
+# Basic torch info
+print(f"\nPyTorch version: {torch.__version__}")
+print(f"CUDA available: {torch.cuda.is_available()}")
+
+if torch.cuda.is_available():
+    print(f"CUDA version: {torch.version.cuda}")
+    print(f"GPU count: {torch.cuda.device_count()}")
+    print(f"Current GPU: {torch.cuda.current_device()}")
+    print(f"GPU name: {torch.cuda.get_device_name(0)}")
+    print(f"GPU memory: {torch.cuda.get_device_properties(0).total_memory / 1024**3:.2f} GB")
+else:
+    print("\nNo CUDA available - would run on CPU")
+
+# Test backend device selection
+print("\n" + "=" * 60)
+print("Backend Device Selection")
+print("=" * 60)
+
+# Simulate the _get_device method from pytorch_backend.py
+def _get_device() -> str:
+    """Get the best available device."""
+    if torch.cuda.is_available():
+        return "cuda"
+    elif hasattr(torch.backends, 'mps') and torch.backends.mps.is_available():
+        # MPS can have issues, use CPU for stability
+        return "cpu"
+    return "cpu"
+
+selected_device = _get_device()
+print(f"\nSelected device: {selected_device}")
+print(f"Would use dtype: {'torch.bfloat16' if selected_device != 'cpu' else 'torch.float32'}")
+
+# Test actual tensor creation on device
+print("\n" + "=" * 60)
+print("Testing Tensor Creation on Device")
+print("=" * 60)
+
+try:
+    test_tensor = torch.randn(1000, 1000).to(selected_device)
+    print(f"\n[OK] Successfully created tensor on {selected_device}")
+    print(f"  Tensor device: {test_tensor.device}")
+    print(f"  Tensor dtype: {test_tensor.dtype}")
+
+    # Test computation
+    result = test_tensor @ test_tensor.T
+    print(f"[OK] Successfully performed computation on {selected_device}")
+
+    if selected_device == "cuda":
+        print(f"\nCUDA memory allocated: {torch.cuda.memory_allocated() / 1024**2:.2f} MB")
+        print(f"CUDA memory reserved: {torch.cuda.memory_reserved() / 1024**2:.2f} MB")
+
+except Exception as e:
+    print(f"\n[ERROR] {e}")
+
+print("\n" + "=" * 60)
+print("Summary")
+print("=" * 60)
+
+if selected_device == "cuda":
+    print("\n[SUCCESS] CUDA IS WORKING!")
+    print("  The backend will use your NVIDIA GPU for inference")
+    print(f"  GPU: {torch.cuda.get_device_name(0)}")
+    print(f"  This will be significantly faster than CPU")
+else:
+    print("\n[FAIL] CUDA is not available")
+    print("  The backend will use CPU for inference")
+    print("  This will be slower than GPU")
+
+print("\n" + "=" * 60)

web/package.json

@@ -1,7 +1,7 @@
 {
   "name": "@voicebox/web",
   "private": true,
-  "version": "0.1.11",
+  "version": "0.1.12",
   "type": "module",
   "scripts": {
     "dev": "vite",