Skip to content

The open-source AI voice studio. Clone, dictate, create.

docs: platform-qualified GPU asset names in developer docs; GPU settings footer says Windows and Linux

jamiepine2ce35092be0e
diff --git a/app/src/i18n/locales/en/translation.json b/app/src/i18n/locales/en/translation.json
index f8973fda..58186f53 100644
--- a/app/src/i18n/locales/en/translation.json
+++ b/app/src/i18n/locales/en/translation.json
@@ -1156,7 +1156,7 @@
"deleteCuda": "Failed to delete CUDA backend",
"deleteRocm": "Failed to delete ROCm backend"
},
- "footer": "Voicebox automatically detects and uses the best available GPU on your system. On Apple Silicon Macs, the MLX backend runs natively on the Neural Engine and GPU via Metal Performance Shaders (MPS), with no additional setup required. On Windows, you can download optional CUDA (NVIDIA) or ROCm (AMD) backends for hardware-accelerated inference. Intel XPU and DirectML are also supported where available through PyTorch. When no GPU is detected, Voicebox falls back to CPU — all engines still work, just slower.",
+ "footer": "Voicebox automatically detects and uses the best available GPU on your system. On Apple Silicon Macs, the MLX backend runs natively on the Neural Engine and GPU via Metal Performance Shaders (MPS), with no additional setup required. On Windows and Linux, you can download optional CUDA (NVIDIA) or ROCm (AMD) backends for hardware-accelerated inference. Intel XPU and DirectML are also supported where available through PyTorch. When no GPU is detected, Voicebox falls back to CPU — all engines still work, just slower.",
"rocm": {
"title": "AMD ROCm Backend",
"activeTitle": "ROCm Backend Active",
diff --git a/app/src/i18n/locales/es/translation.json b/app/src/i18n/locales/es/translation.json
index bd0da662..9d2d851e 100644
--- a/app/src/i18n/locales/es/translation.json
+++ b/app/src/i18n/locales/es/translation.json
@@ -1135,7 +1135,7 @@
"deleteCuda": "Error al eliminar el backend CUDA",
"deleteRocm": "Error al eliminar el backend ROCm"
},
- "footer": "Voicebox detecta y usa automáticamente la mejor GPU disponible en tu sistema. En Macs con Apple Silicon, el backend MLX se ejecuta de forma nativa en el Neural Engine y la GPU mediante Metal Performance Shaders (MPS), sin configuración adicional. En Windows, puedes descargar backends opcionales CUDA (NVIDIA) o ROCm (AMD) para inferencia acelerada por hardware. Intel XPU y DirectML también son compatibles cuando están disponibles a través de PyTorch. Cuando no se detecta ninguna GPU, Voicebox recurre a la CPU: todos los motores siguen funcionando, solo que más despacio.",
+ "footer": "Voicebox detecta y usa automáticamente la mejor GPU disponible en tu sistema. En Macs con Apple Silicon, el backend MLX se ejecuta de forma nativa en el Neural Engine y la GPU mediante Metal Performance Shaders (MPS), sin configuración adicional. En Windows y Linux, puedes descargar backends opcionales CUDA (NVIDIA) o ROCm (AMD) para inferencia acelerada por hardware. Intel XPU y DirectML también son compatibles cuando están disponibles a través de PyTorch. Cuando no se detecta ninguna GPU, Voicebox recurre a la CPU: todos los motores siguen funcionando, solo que más despacio.",
"activeBackend": {
"description": "La aceleración por GPU está habilitada actualmente."
},
diff --git a/app/src/i18n/locales/fr/translation.json b/app/src/i18n/locales/fr/translation.json
index 2815f9ff..f2af10b2 100644
--- a/app/src/i18n/locales/fr/translation.json
+++ b/app/src/i18n/locales/fr/translation.json
@@ -1128,7 +1128,7 @@
"switchCpu": "Échec du basculement vers CPU",
"deleteCuda": "Échec de la suppression du backend CUDA"
},
- "footer": "Voicebox détecte et utilise automatiquement le meilleur GPU disponible sur votre système. Sur les Mac Apple Silicon, le backend MLX fonctionne nativement sur le Neural Engine et le GPU via Metal Performance Shaders (MPS), sans configuration supplémentaire. Sur Windows et Linux avec GPU NVIDIA, vous pouvez télécharger un backend CUDA optionnel pour l'inférence accélérée par le matériel. AMD ROCm, Intel XPU et DirectML sont également pris en charge là où disponibles via PyTorch. Quand aucun GPU n'est détecté, Voicebox utilise le CPU — tous les moteurs fonctionnent, mais plus lentement."
+ "footer": "Voicebox détecte et utilise automatiquement le meilleur GPU disponible sur votre système. Sur les Mac Apple Silicon, le backend MLX fonctionne nativement sur le Neural Engine et le GPU via Metal Performance Shaders (MPS), sans configuration supplémentaire. Sur Windows et Linux, vous pouvez télécharger des backends optionnels CUDA (NVIDIA) ou ROCm (AMD) pour l'inférence accélérée par le matériel. Intel XPU et DirectML sont également pris en charge là où disponibles via PyTorch. Quand aucun GPU n'est détecté, Voicebox utilise le CPU — tous les moteurs fonctionnent, mais plus lentement."
},
"logs": {
"title": "Journaux du serveur",
diff --git a/app/src/i18n/locales/it/translation.json b/app/src/i18n/locales/it/translation.json
index 3ec72b77..e63a296e 100644
--- a/app/src/i18n/locales/it/translation.json
+++ b/app/src/i18n/locales/it/translation.json
@@ -1137,7 +1137,7 @@
"deleteCuda": "Impossibile eliminare il backend CUDA",
"deleteRocm": "Impossibile eliminare il backend ROCm"
},
- "footer": "Voicebox rileva e utilizza automaticamente la migliore GPU disponibile sul tuo sistema. Sui Mac con chip Apple Silicon, il backend MLX viene eseguito nativamente sul Neural Engine e sulla GPU tramite Metal Performance Shaders (MPS), senza richiedere alcuna configurazione aggiuntiva. Su Windows, puoi scaricare i backend opzionali CUDA (NVIDIA) o ROCm (AMD) per l'inferenza con accelerazione hardware. Dove disponibili tramite PyTorch, sono supportati anche Intel XPU e DirectML. Quando non viene rilevata alcuna GPU, Voicebox si affida alla CPU — tutti i motori continuano a funzionare, solo più lentamente.",
+ "footer": "Voicebox rileva e utilizza automaticamente la migliore GPU disponibile sul tuo sistema. Sui Mac con chip Apple Silicon, il backend MLX viene eseguito nativamente sul Neural Engine e sulla GPU tramite Metal Performance Shaders (MPS), senza richiedere alcuna configurazione aggiuntiva. Su Windows e Linux, puoi scaricare i backend opzionali CUDA (NVIDIA) o ROCm (AMD) per l'inferenza con accelerazione hardware. Dove disponibili tramite PyTorch, sono supportati anche Intel XPU e DirectML. Quando non viene rilevata alcuna GPU, Voicebox si affida alla CPU — tutti i motori continuano a funzionare, solo più lentamente.",
"rocm": {
"title": "Backend AMD ROCm",
"activeTitle": "Backend ROCm attivo",
diff --git a/app/src/i18n/locales/ja/translation.json b/app/src/i18n/locales/ja/translation.json
index f6640e86..5ae4c514 100644
--- a/app/src/i18n/locales/ja/translation.json
+++ b/app/src/i18n/locales/ja/translation.json
@@ -1147,7 +1147,7 @@
"switchCpu": "CPU への切り替えに失敗しました",
"deleteCuda": "CUDA バックエンドの削除に失敗しました"
},
- "footer": "Voicebox はシステムで利用可能な最適な GPU を自動で検出し使用します。Apple Silicon Mac では、MLX バックエンドが Metal Performance Shaders(MPS)を介して Neural Engine と GPU 上でネイティブに動作し、追加のセットアップは不要です。NVIDIA GPU 搭載の Windows および Linux では、オプションの CUDA バックエンドをダウンロードしてハードウェアアクセラレーションによる推論が可能です。AMD ROCm、Intel XPU、DirectML も PyTorch を通じて利用可能な環境でサポートされます。GPU が検出されない場合、Voicebox は CPU にフォールバックし、すべてのエンジンはそのまま動作しますが速度は低下します。"
+ "footer": "Voicebox はシステムで利用可能な最適な GPU を自動で検出し使用します。Apple Silicon Mac では、MLX バックエンドが Metal Performance Shaders(MPS)を介して Neural Engine と GPU 上でネイティブに動作し、追加のセットアップは不要です。Windows および Linux では、オプションの CUDA(NVIDIA)または ROCm(AMD)バックエンドをダウンロードしてハードウェアアクセラレーションによる推論が可能です。Intel XPU、DirectML も PyTorch を通じて利用可能な環境でサポートされます。GPU が検出されない場合、Voicebox は CPU にフォールバックし、すべてのエンジンはそのまま動作しますが速度は低下します。"
},
"logs": {
"title": "サーバーログ",
diff --git a/app/src/i18n/locales/ko/translation.json b/app/src/i18n/locales/ko/translation.json
index 284e082f..88d438dd 100644
--- a/app/src/i18n/locales/ko/translation.json
+++ b/app/src/i18n/locales/ko/translation.json
@@ -1133,7 +1133,7 @@
"deleteCuda": "CUDA 백엔드 삭제 실패",
"deleteRocm": "ROCm 백엔드 삭제 실패"
},
- "footer": "Voicebox는 시스템에서 사용 가능한 최고의 GPU를 자동으로 감지하여 사용합니다. Apple Silicon Mac에서는 Metal Performance Shaders(MPS)를 통해 MLX 백엔드가 Neural Engine과 GPU에서 기본 실행되며 추가 설정이 필요하지 않습니다. Windows에서는 선택적 CUDA(NVIDIA) 또는 ROCm(AMD) 백엔드를 다운로드하여 하드웨어 가속 추론을 사용할 수 있습니다. Intel XPU와 DirectML도 PyTorch를 통해 지원됩니다. GPU가 감지되지 않으면 Voicebox가 CPU로 대체됩니다 — 모든 엔진이 작동하지만 더 느립니다.",
+ "footer": "Voicebox는 시스템에서 사용 가능한 최고의 GPU를 자동으로 감지하여 사용합니다. Apple Silicon Mac에서는 Metal Performance Shaders(MPS)를 통해 MLX 백엔드가 Neural Engine과 GPU에서 기본 실행되며 추가 설정이 필요하지 않습니다. Windows 및 Linux에서는 선택적 CUDA(NVIDIA) 또는 ROCm(AMD) 백엔드를 다운로드하여 하드웨어 가속 추론을 사용할 수 있습니다. Intel XPU와 DirectML도 PyTorch를 통해 지원됩니다. GPU가 감지되지 않으면 Voicebox가 CPU로 대체됩니다 — 모든 엔진이 작동하지만 더 느립니다.",
"rocm": {
"title": "AMD ROCm 백엔드",
"activeTitle": "ROCm 백엔드 활성",
diff --git a/app/src/i18n/locales/pt-BR/translation.json b/app/src/i18n/locales/pt-BR/translation.json
index b4972a90..08975f28 100644
--- a/app/src/i18n/locales/pt-BR/translation.json
+++ b/app/src/i18n/locales/pt-BR/translation.json
@@ -1128,7 +1128,7 @@
"switchCpu": "Falha ao mudar para CPU",
"deleteCuda": "Falha ao excluir o backend CUDA"
},
- "footer": "O Voicebox detecta e usa automaticamente a melhor GPU disponível no seu sistema. Em Macs com Apple Silicon, o backend MLX roda nativamente no Neural Engine e na GPU via Metal Performance Shaders (MPS), sem configuração adicional. No Windows e no Linux com GPUs NVIDIA, você pode baixar um backend CUDA opcional para inferência acelerada por hardware. AMD ROCm, Intel XPU e DirectML também são suportados quando disponíveis através do PyTorch. Quando nenhuma GPU é detectada, o Voicebox recorre à CPU — todos os motores ainda funcionam, só que mais devagar."
+ "footer": "O Voicebox detecta e usa automaticamente a melhor GPU disponível no seu sistema. Em Macs com Apple Silicon, o backend MLX roda nativamente no Neural Engine e na GPU via Metal Performance Shaders (MPS), sem configuração adicional. No Windows e no Linux, você pode baixar backends opcionais CUDA (NVIDIA) ou ROCm (AMD) para inferência acelerada por hardware. Intel XPU e DirectML também são suportados quando disponíveis através do PyTorch. Quando nenhuma GPU é detectada, o Voicebox recorre à CPU — todos os motores ainda funcionam, só que mais devagar."
},
"logs": {
"title": "Logs do Servidor",
diff --git a/app/src/i18n/locales/zh-CN/translation.json b/app/src/i18n/locales/zh-CN/translation.json
index 1c2dbbb1..ff827a0c 100644
--- a/app/src/i18n/locales/zh-CN/translation.json
+++ b/app/src/i18n/locales/zh-CN/translation.json
@@ -1128,7 +1128,7 @@
"switchCpu": "切换到 CPU 失败",
"deleteCuda": "删除 CUDA 后端失败"
},
- "footer": "Voicebox 会自动检测并使用系统上可用的最佳 GPU。在 Apple Silicon Mac 上,MLX 后端通过 Metal Performance Shaders (MPS) 在神经引擎和 GPU 上原生运行,无需额外设置。在配备 NVIDIA GPU 的 Windows 和 Linux 上,您可以下载可选的 CUDA 后端以获得硬件加速推理。AMD ROCm、Intel XPU 和 DirectML 也通过 PyTorch 获得支持。未检测到 GPU 时,Voicebox 会退回到 CPU——所有引擎仍可工作,只是速度较慢。"
+ "footer": "Voicebox 会自动检测并使用系统上可用的最佳 GPU。在 Apple Silicon Mac 上,MLX 后端通过 Metal Performance Shaders (MPS) 在神经引擎和 GPU 上原生运行,无需额外设置。在 Windows 和 Linux 上,您可以下载可选的 CUDA (NVIDIA) 或 ROCm (AMD) 后端以获得硬件加速推理。Intel XPU 和 DirectML 也通过 PyTorch 获得支持。未检测到 GPU 时,Voicebox 会退回到 CPU——所有引擎仍可工作,只是速度较慢。"
},
"logs": {
"title": "服务器日志",
diff --git a/app/src/i18n/locales/zh-TW/translation.json b/app/src/i18n/locales/zh-TW/translation.json
index 3020686f..ed1e6bbb 100644
--- a/app/src/i18n/locales/zh-TW/translation.json
+++ b/app/src/i18n/locales/zh-TW/translation.json
@@ -1128,7 +1128,7 @@
"switchCpu": "切換到 CPU 失敗",
"deleteCuda": "刪除 CUDA 後端失敗"
},
- "footer": "Voicebox 會自動偵測並使用系統上可用的最佳 GPU。在 Apple Silicon Mac 上,MLX 後端透過 Metal Performance Shaders (MPS) 在神經引擎與 GPU 上原生執行,無需額外設定。在配備 NVIDIA GPU 的 Windows 與 Linux 上,可以下載選用的 CUDA 後端以取得硬體加速推論。AMD ROCm、Intel XPU 與 DirectML 也透過 PyTorch 獲得支援。未偵測到 GPU 時,Voicebox 會退回到 CPU——所有引擎仍可運作,只是速度較慢。"
+ "footer": "Voicebox 會自動偵測並使用系統上可用的最佳 GPU。在 Apple Silicon Mac 上,MLX 後端透過 Metal Performance Shaders (MPS) 在神經引擎與 GPU 上原生執行,無需額外設定。在 Windows 與 Linux 上,可以下載選用的 CUDA (NVIDIA) 或 ROCm (AMD) 後端以取得硬體加速推論。Intel XPU 與 DirectML 也透過 PyTorch 獲得支援。未偵測到 GPU 時,Voicebox 會退回到 CPU——所有引擎仍可運作,只是速度較慢。"
},
"logs": {
"title": "伺服器日誌",
diff --git a/docs/content/docs/developer/autoupdater.mdx b/docs/content/docs/developer/autoupdater.mdx
index d09ba58b..167b5fb8 100644
--- a/docs/content/docs/developer/autoupdater.mdx
+++ b/docs/content/docs/developer/autoupdater.mdx
@@ -156,8 +156,8 @@ The CUDA-enabled backend is distributed separately from the main app because bun
Since v0.4, the CUDA binary is packaged as **two archives** instead of one:
-- **Server core** (`voicebox-server-cuda.tar.gz`) — the Python server + PyTorch code, changes every release.
-- **CUDA libs** (`cuda-libs-cu128-v1.tar.gz`) — the heavy NVIDIA CUDA/cuDNN DLLs, only re-downloaded when the CUDA toolkit major version changes.
+- **Server core** (`voicebox-server-cuda-<platform>.tar.gz`, where `<platform>` is `windows-x86_64` or `linux-x86_64`) — the Python server + PyTorch code, changes every release.
+- **CUDA libs** (`cuda-libs-<platform>-cu128-v1.tar.gz`) — the heavy NVIDIA CUDA/cuDNN DLLs, only re-downloaded when the CUDA toolkit major version changes.
This means most Voicebox upgrades only re-download the ~945 MB server core, not the full ~2.5 GB bundle.
@@ -203,7 +203,7 @@ Downloads report progress via Server-Sent Events (SSE):
GET /backend/cuda-progress
event: progress
-data: {"current": 52428800, "total": 945000000, "filename": "voicebox-server-cuda.tar.gz", "status": "downloading"}
+data: {"current": 52428800, "total": 945000000, "filename": "voicebox-server-cuda-windows-x86_64.tar.gz", "status": "downloading"}
```
The frontend subscribes to this endpoint to show real-time progress, including which archive (server core vs libs) is currently downloading.
@@ -212,7 +212,9 @@ The frontend subscribes to this endpoint to show real-time progress, including w
For each CUDA-capable release, these files are uploaded to GitHub:
-- `voicebox-server-cuda.tar.gz` — server-core archive
-- `voicebox-server-cuda.tar.gz.sha256` — checksum
-- `cuda-libs-cu128-v1.tar.gz` — CUDA runtime libs (only when the libs version bumps)
-- `cuda-libs-cu128-v1.tar.gz.sha256` — checksum
+- `voicebox-server-cuda-<platform>.tar.gz` — server-core archive
+- `voicebox-server-cuda-<platform>.tar.gz.sha256` — checksum
+- `cuda-libs-<platform>-cu128-v1.tar.gz` — CUDA runtime libs (only when the libs version bumps)
+- `cuda-libs-<platform>-cu128-v1.tar.gz.sha256` — checksum
+
+`<platform>` is `windows-x86_64` (built by `build-cuda-windows`) or `linux-x86_64` (built by the `build-gpu-linux` matrix job, which also publishes `voicebox-server-rocm-linux-x86_64.tar.gz` and `rocm-libs-linux-x86_64-rocm7.2-v1.tar.gz`). The client picks the token with `server_asset_platform()` in `backend/utils/platform_detect.py`; macOS and Linux arm64 have no GPU assets and are refused before any download starts.
diff --git a/docs/content/docs/developer/building.mdx b/docs/content/docs/developer/building.mdx
index ed80f01d..a515edf4 100644
--- a/docs/content/docs/developer/building.mdx
+++ b/docs/content/docs/developer/building.mdx
@@ -162,10 +162,12 @@ The `build-cuda-windows` job runs separately:
1. Install PyTorch with CUDA 12.8
2. Build with `build_binary.py --cuda` (produces `--onedir` output)
3. Package with `scripts/package_cuda.py` into two archives:
- - `voicebox-server-cuda.tar.gz` — server core (~945 MB)
- - `cuda-libs-cu128-v1.tar.gz` — NVIDIA runtime libraries (~1.7 GB, cached independently)
+ - `voicebox-server-cuda-windows-x86_64.tar.gz` — server core (~945 MB)
+ - `cuda-libs-windows-x86_64-cu128-v1.tar.gz` — NVIDIA runtime libraries (~1.7 GB, cached independently)
4. Upload archives as release artifacts
+The `build-gpu-linux` matrix job does the same on `ubuntu-latest` for CUDA (cu128) and ROCm (PyTorch rocm7.2 index), producing `voicebox-server-{cuda,rocm}-linux-x86_64.tar.gz` and `{cuda,rocm}-libs-linux-x86_64-<version>.tar.gz`.
+
This binary is downloaded on-demand by users who enable CUDA in settings. The CUDA libs archive is only re-downloaded when the CUDA toolkit version changes, not on every app update.
## Troubleshooting