fix: select CUDA runtime by GPU architecture

hgmzhn · hgmzhn · commit 1b3b6b306447 · 2026-08-13T04:10:49.000+08:00
diff --git a/README.md b/README.md
@@ -136,7 +136,7 @@
 
 #### 下载发布便携包
 
-从 [GitHub Releases](https://github.com/hgmzhn/manga-translator-ui/releases) 下载 CPU、CUDA 13.0、CUDA 12.6 或实验性 ROCm 7.2.1 版本，解压后双击 `Win-Start.bat`。RTX 50 系列显卡请下载文件名含 `cuda13.0` 的版本；其他 NVIDIA 显卡可按兼容性选择 `cuda12.6`，支持 CUDA 13.0 及以上的驱动也能运行 CUDA 12.6 版本。包内已安装 Python 依赖和模型文件；ROCm 7.2.1 版本要求受支持的 AMD 显卡与 26.2.2 驱动。 📖 [发行版下载说明](https://hgmzhn.github.io/manga-translator-ui/zh/install/release-download)
+从 [GitHub Releases](https://github.com/hgmzhn/manga-translator-ui/releases) 下载 CPU、CUDA 13.0、CUDA 12.6 或实验性 ROCm 7.2.1 版本，解压后双击 `Win-Start.bat`。GeForce 10 系列（如 GTX 1060/1070/1080）必须下载文件名含 `cuda12.6` 的兼容版本；CUDA 13.0 仅支持 Turing（计算能力 7.5）及更新架构，RTX 20/30/40/50 系列在驱动支持 CUDA 13.0 时可选择 `cuda13.0`，RTX 50 系列推荐该版本。支持 CUDA 13.0 及以上的驱动也能运行 CUDA 12.6 包。包内已安装 Python 依赖和模型文件；ROCm 7.2.1 版本要求受支持的 AMD 显卡与 26.2.2 驱动。 📖 [发行版下载说明](https://hgmzhn.github.io/manga-translator-ui/zh/install/release-download)
 
 #### 从源码运行（开发者）
 
diff --git a/README_EN.md b/README_EN.md
@@ -136,7 +136,7 @@ No separate Python installation is required. Download the [portable package](htt
 
 #### Download a portable release
 
-Download a CPU, CUDA 13.0, CUDA 12.6, or experimental ROCm 7.2.1 build from [GitHub Releases](https://github.com/hgmzhn/manga-translator-ui/releases), extract it, and run `Win-Start.bat`. RTX 50-series cards must use the build whose filename contains `cuda13.0`; other NVIDIA GPUs may choose `cuda12.6` for compatibility, and drivers supporting CUDA 13.0 or newer can also run the CUDA 12.6 build. Python dependencies and model files are included. The ROCm 7.2.1 build requires a supported AMD GPU and driver 26.2.2. 📖 [Release download notes](https://hgmzhn.github.io/manga-translator-ui/en/install/release-download)
+Download a CPU, CUDA 13.0, CUDA 12.6, or experimental ROCm 7.2.1 build from [GitHub Releases](https://github.com/hgmzhn/manga-translator-ui/releases), extract it, and run `Win-Start.bat`. GeForce 10-series GPUs such as the GTX 1060/1070/1080 must use the compatibility build whose filename contains `cuda12.6`. CUDA 13.0 requires Turing (compute capability 7.5) or newer; RTX 20/30/40/50-series GPUs may use `cuda13.0` when the driver supports CUDA 13.0, and it is recommended for RTX 50-series GPUs. Drivers supporting CUDA 13.0 or newer can also run the CUDA 12.6 build. Python dependencies and model files are included. The ROCm 7.2.1 build requires a supported AMD GPU and driver 26.2.2. 📖 [Release download notes](https://hgmzhn.github.io/manga-translator-ui/en/install/release-download)
 
 #### Run from source (developers)
 
diff --git a/doc/INSTALLATION.md b/doc/INSTALLATION.md
@@ -29,10 +29,10 @@
 
 - **内存**：16 GB RAM 或更多
 - **GPU**：
-  - **NVIDIA 显卡**：默认使用 CUDA 13.0；需要兼容版本时可使用 CUDA 12.6，支持 CUDA 13.0 及以上的驱动也能运行 CUDA 12.6 版本
+  - **NVIDIA 显卡**：GeForce 10 系必须使用 CUDA 12.6；CUDA 13.0 仅支持 Turing（计算能力 7.5）及更新架构，支持 CUDA 13.0 及以上的驱动也能运行 CUDA 12.6 版本
     - 建议显存：6 GB 或更多
     - 支持的 NVIDIA 显卡：GTX 1060 及以上
-    - RTX 50 系列必须下载 CUDA 13.0 版本。
+    - RTX 20/30/40/50 系列在驱动支持 CUDA 13.0 时可选择 CUDA 13.0；RTX 50 系列推荐此版本
   - **AMD 显卡**：支持 ROCm（实验性）
     - 支持的显卡：**仅 RX 7000/9000 系列（RDNA 3/4）**
     - ⚠️ RX 5000/6000 系列请使用 CPU 版本
@@ -70,7 +70,7 @@
 2. **强制同步最新代码**：同步失败会提示切换到另一条线路重试
 3. **检测显卡**：自动识别 NVIDIA / AMD / 集显；多显卡时列出让用户选择
 4. **选择 PyTorch 版本**：
-   - **NVIDIA**：默认 CUDA 13.0（`cuda12.6` 分支使用 CUDA 12.6）
+   - **NVIDIA**：按显卡型号、计算能力和驱动自动选择；GeForce 10 系强制 CUDA 12.6，Turing（计算能力 7.5）及更新架构在驱动支持时使用 CUDA 13.0
    - **AMD**：ROCm（实验性，**仅 RX 7000/9000 系列**）
    - **其他/集显**：CPU 版本
 5. **uv 高速批量安装依赖**：
@@ -123,10 +123,12 @@
 ### 2. 选择版本
 
 - `manga-translator-cpu-vX.Y.Z.7z.001`：兼容性最好，不需要独立显卡。
-- `manga-translator-cuda13.0-vX.Y.Z.7z.001`：NVIDIA CUDA 13.0 版本；RTX 50 系列必须选择此版本。
-- `manga-translator-cuda12.6-vX.Y.Z.7z.001`：NVIDIA CUDA 12.6 版本；适用于 CUDA 12.x，也可运行在支持 CUDA 13.0 及以上的驱动上。
+- `manga-translator-cuda13.0-vX.Y.Z.7z.001`：适用于 Turing（计算能力 7.5）及更新的 NVIDIA 显卡，并且驱动支持 CUDA 13.0；RTX 50 系列推荐此版本。
+- `manga-translator-cuda12.6-vX.Y.Z.7z.001`：GeForce 10 系列必须选择此版本；也适用于其他需要兼容版本的 NVIDIA 显卡，支持 CUDA 13.0 及以上的驱动可向下兼容运行。
 - `manga-translator-rocm7.2.1-vX.Y.Z.7z.001`：实验性 Radeon ROCm 7.2.1 版本，需要受支持的 AMD 显卡和 AMD 26.2.2 驱动。
 
+> CUDA 13.0 已不支持 Turing 之前的 NVIDIA 架构。GTX 1060/1070/1080 等 GeForce 10 系列不要下载 CUDA 13.0 包，必须使用 CUDA 12.6 包。
+
 ### 3. 分卷解压
 
 下载同一版本的全部 `.7z.001`、`.002` 等分卷到同一目录，保持原文件名不变，然后解压 `.001`。缺少任一分卷都会导致解压失败。
@@ -591,7 +593,7 @@ A: 运行 `./Unix-Install-or-Update.sh`，在 Python 菜单中选择 [2] 更新
 **问题**：GPU 版本运行时崩溃或报错
 
 **解决方法**：
-1. 确认显卡和驱动支持 CUDA 13.0；CUDA 12.6 包则确认支持 CUDA 12.6
+1. 确认版本匹配：GeForce 10 系必须使用 CUDA 12.6；CUDA 13.0 需要 Turing（计算能力 7.5）及更新架构和支持 CUDA 13.0 的驱动
 2. 安装或更新 NVIDIA 显卡驱动
 3. 如需开发工具链，下载并安装 [CUDA Toolkit 12.x](https://developer.nvidia.com/cuda-downloads)
 4. 如果仍然失败，使用 CPU 版本
diff --git a/doc/en/INSTALLATION.md b/doc/en/INSTALLATION.md
@@ -31,10 +31,10 @@ This document provides detailed installation steps, system requirements, first-r
 
 - **Memory**: 16 GB RAM or more
 - **GPU**:
-  - **NVIDIA GPU**: CUDA 13.0 is the default; CUDA 12.6 is available for compatibility and also runs on drivers supporting CUDA 13.0 or newer
+  - **NVIDIA GPU**: GeForce 10-series GPUs must use CUDA 12.6; CUDA 13.0 requires Turing (compute capability 7.5) or newer. Drivers supporting CUDA 13.0 or newer can also run the CUDA 12.6 build
     - Recommended VRAM: 6 GB or more
     - Typical supported class: GTX 1060 and above
-    - RTX 50-series cards must download the CUDA 13.0 build.
+    - RTX 20/30/40/50-series GPUs may use CUDA 13.0 when the driver supports it; CUDA 13.0 is recommended for RTX 50-series GPUs
   - **AMD GPU**: ROCm support is experimental
     - Supported cards: **RX 7000 / 9000 only**
     - ⚠️ RX 5000 / 6000 should use the CPU build
@@ -72,7 +72,7 @@ Double-click `Win-Install-or-Update.bat` and choose **[1] Install** in the maint
 2. **Force-sync the latest code**: after a successful sync, the maintenance launcher reloads the updated code before continuing
 3. **GPU detection**: automatically detects NVIDIA / AMD / integrated graphics; with multiple GPUs you get a list to pick from
 4. **Choose a PyTorch build**:
-   - **NVIDIA**: CUDA 13.0 by default (the `cuda12.6` branch uses CUDA 12.6)
+   - **NVIDIA**: selected automatically from the GPU model, compute capability, and driver; GeForce 10-series GPUs are forced to CUDA 12.6, while Turing (compute capability 7.5) or newer uses CUDA 13.0 when supported by the driver
    - **AMD**: ROCm, experimental, **RX 7000 / 9000 series only**
    - **Other / integrated graphics**: CPU build
 5. **Fast batch dependency install with uv**:
@@ -125,10 +125,12 @@ Go to [GitHub Releases](https://github.com/hgmzhn/manga-translator-ui/releases).
 ### 2. Choose a build
 
 - `manga-translator-cpu-vX.Y.Z.7z.001`: best compatibility; no dedicated GPU required.
-- `manga-translator-cuda13.0-vX.Y.Z.7z.001`: NVIDIA CUDA 13.0; RTX 50-series cards must use this build.
-- `manga-translator-cuda12.6-vX.Y.Z.7z.001`: NVIDIA CUDA 12.6; supports CUDA 12.x systems and also runs on drivers supporting CUDA 13.0 or newer.
+- `manga-translator-cuda13.0-vX.Y.Z.7z.001`: for Turing (compute capability 7.5) or newer NVIDIA GPUs whose driver supports CUDA 13.0; recommended for RTX 50-series GPUs.
+- `manga-translator-cuda12.6-vX.Y.Z.7z.001`: required for GeForce 10-series GPUs and suitable for other NVIDIA GPUs needing compatibility; drivers supporting CUDA 13.0 or newer can run it through backward compatibility.
 - `manga-translator-rocm7.2.1-vX.Y.Z.7z.001`: experimental Radeon ROCm 7.2.1; requires a supported AMD GPU and AMD driver 26.2.2.
 
+> CUDA 13.0 no longer supports NVIDIA architectures before Turing. GeForce 10-series GPUs such as the GTX 1060/1070/1080 must use the CUDA 12.6 package and must not download the CUDA 13.0 package.
+
 ### 3. Extract split volumes
 
 Download every `.7z.001`, `.002`, and later volume for the selected build into one directory without renaming them, then extract `.001`. Missing any volume makes extraction fail.
@@ -565,7 +567,7 @@ If you want to fine-tune the result later, open it in `Editor View`.
 
 **Try this**
 
-1. Confirm the GPU and driver support CUDA 13.0; the CUDA 12.6 build requires CUDA 12.6
+1. Confirm the build matches the GPU: GeForce 10-series GPUs must use CUDA 12.6; CUDA 13.0 requires Turing (compute capability 7.5) or newer plus a driver that supports CUDA 13.0
 2. Update the NVIDIA driver
 3. If the issue is ONNX-specific, open `Settings` -> `General` and enable `Disable ONNX GPU Acceleration`
 4. If the machine is not compatible, switch to the CPU build
diff --git a/doc/wiki/en/install/release-download.md b/doc/wiki/en/install/release-download.md
@@ -30,12 +30,14 @@ Open the latest version on [GitHub Releases](https://github.com/hgmzhn/manga-tra
 | Release filename prefix | Best for | Runtime |
 | --- | --- | --- |
 | `manga-translator-cpu-vX.Y.Z.7z.*` | Any Windows x64 computer; use it without a discrete GPU or when compatibility is uncertain | CPU PyTorch / ONNX Runtime |
-| `manga-translator-cuda13.0-vX.Y.Z.7z.*` | NVIDIA GPUs supporting CUDA 13.0 or newer; RTX 50-series cards must use this build | CUDA 13.0 / PyTorch cu130 |
-| `manga-translator-cuda12.6-vX.Y.Z.7z.*` | NVIDIA GPUs needing the CUDA 12.6 compatibility build; drivers supporting CUDA 13.0 or newer can also run it | CUDA 12.6 / PyTorch cu126 |
+| `manga-translator-cuda13.0-vX.Y.Z.7z.*` | Turing (compute capability 7.5) or newer NVIDIA GPUs whose driver supports CUDA 13.0; recommended for RTX 50-series GPUs | CUDA 13.0 / PyTorch cu130 |
+| `manga-translator-cuda12.6-vX.Y.Z.7z.*` | Required for GeForce 10-series GPUs; also for other NVIDIA GPUs needing the compatibility build. Drivers supporting CUDA 13.0 or newer can run it through backward compatibility | CUDA 12.6 / PyTorch cu126 |
 | `manga-translator-rocm7.2.1-vX.Y.Z.7z.*` | AMD GPUs supported by Windows ROCm | Experimental Radeon ROCm 7.2.1; AMD driver 26.2.2 is required |
 
 The ROCm 7.2.1 package works only with supported AMD GPUs. Use the CPU package when uncertain, and never mix volumes from different runtimes.
 
+> CUDA 13.0 no longer supports NVIDIA architectures before Turing. The installer explicitly recognizes GeForce 10-series model names and forces CUDA 12.6; it also conservatively falls back to CUDA 12.6 when compute capability cannot be detected.
+
 ## Download and extract the volumes
 
 1. Download **all** `.7z.001`, `.7z.002`, and later volumes for the selected variant into one directory without renaming them.
diff --git a/doc/wiki/en/install/requirements.md b/doc/wiki/en/install/requirements.md
@@ -23,8 +23,8 @@ The current source of dependency truth is `pyproject.toml` and `uv.lock`. `requi
 | RAM | 8 GB | 16 GB or more |
 | Disk | 5 GB free space | 10 GB SSD |
 | Python (source) | 3.12 (`>=3.12,<3.13`) | 3.12 |
-| NVIDIA GPU | GTX 1060 or newer with 6 GB VRAM; driver must support CUDA 13.0 by default, or CUDA 12.6 for the alternate build | More VRAM is better |
-> RTX 50-series NVIDIA cards must use the NVIDIA GPU package; select the CUDA 13.0 default package or the CUDA 12.6 package only when its driver supports that runtime.
+| NVIDIA GPU | GTX 1060 or newer with 6 GB VRAM; GeForce 10-series GPUs must use CUDA 12.6, while CUDA 13.0 requires Turing (compute capability 7.5) or newer | More VRAM is better |
+> GeForce 10-series GPUs such as the GTX 1060/1070/1080 must use the CUDA 12.6 package even when a newer driver reports CUDA 13 support. RTX 20/30/40/50-series GPUs may use CUDA 13.0 when the driver supports it; CUDA 13.0 is recommended for RTX 50-series GPUs.
 | AMD GPU | RX 7000/9000 series only (RDNA 3/4); ROCm is experimental. Use the CPU build on RX 5000/6000 | — |
 
 > Windows AMD users can choose the experimental AMD portable release or install through the maintenance script. Both require a supported GPU, Radeon ROCm 7.2.1, and AMD driver 26.2.2; this remains an experimental Windows path.
diff --git a/doc/wiki/zh/install/release-download.md b/doc/wiki/zh/install/release-download.md
@@ -30,12 +30,14 @@ lastUpdated: true
 | 发布文件前缀 | 适用设备 | 运行时 |
 | --- | --- | --- |
 | `manga-translator-cpu-vX.Y.Z.7z.*` | 所有 Windows x64 电脑；无独显或不确定兼容性时选它 | CPU PyTorch / ONNX Runtime |
-| `manga-translator-cuda13.0-vX.Y.Z.7z.*` | 支持 CUDA 13.0 或更高版本的 NVIDIA 显卡；RTX 50 系列必须选此包 | CUDA 13.0 / PyTorch cu130 |
-| `manga-translator-cuda12.6-vX.Y.Z.7z.*` | 需要 CUDA 12.6 兼容版本的 NVIDIA 显卡；支持 CUDA 13.0 及以上的驱动也可运行 | CUDA 12.6 / PyTorch cu126 |
+| `manga-translator-cuda13.0-vX.Y.Z.7z.*` | Turing（计算能力 7.5）及更新的 NVIDIA 显卡，并且驱动支持 CUDA 13.0；RTX 50 系列推荐此包 | CUDA 13.0 / PyTorch cu130 |
+| `manga-translator-cuda12.6-vX.Y.Z.7z.*` | GeForce 10 系列必须选此包；也适用于其他需要兼容版本的 NVIDIA 显卡，支持 CUDA 13.0 及以上的驱动可向下兼容运行 | CUDA 12.6 / PyTorch cu126 |
 | `manga-translator-rocm7.2.1-vX.Y.Z.7z.*` | 受 Windows ROCm 支持的 AMD 显卡 | 实验性 Radeon ROCm 7.2.1；要求 AMD 26.2.2 驱动 |
 
 ROCm 7.2.1 包只适用于其支持范围内的 AMD 显卡。无法确认兼容性时先使用 CPU 包；不要把不同运行时的分卷混在一起。
 
+> CUDA 13.0 已不支持 Turing 之前的 NVIDIA 架构。安装器会优先按显卡型号识别 GeForce 10 系列，并强制选择 CUDA 12.6；无法读取计算能力时也会保守回退到 CUDA 12.6。
+
 ## 下载和解压分卷
 
 1. 下载所选版本的 `.7z.001`、`.7z.002` 等**全部分卷**，放在同一目录并保持原文件名。
diff --git a/doc/wiki/zh/install/requirements.md b/doc/wiki/zh/install/requirements.md
@@ -23,8 +23,8 @@ lastUpdated: true
 | 内存 | 8 GB | 16 GB 或更多 |
 | 磁盘 | 5 GB 可用空间 | 10 GB SSD |
 | Python（源码版） | 3.12（`>=3.12,<3.13`） | 3.12 |
-| NVIDIA GPU | GTX 1060 及以上、6 GB 显存；默认驱动需支持 CUDA 13.0，备用版本需支持 CUDA 12.6 | 显存更大更佳 |
-> RTX 50 系列 NVIDIA 显卡必须使用 NVIDIA GPU 包；默认选择 CUDA 13.0 包，只有在驱动支持该运行时时才选择 CUDA 12.6 包。
+| NVIDIA GPU | GTX 1060 及以上、6 GB 显存；GeForce 10 系必须使用 CUDA 12.6，CUDA 13.0 仅支持 Turing（计算能力 7.5）及更新架构 | 显存更大更佳 |
+> GeForce 10 系列（如 GTX 1060/1070/1080）即使新驱动显示支持 CUDA 13，也必须选择 CUDA 12.6 包。RTX 20/30/40/50 系列在驱动支持 CUDA 13.0 时可选择 CUDA 13.0 包；RTX 50 系列推荐 CUDA 13.0。
 | AMD GPU | 仅 RX 7000/9000 系列（RDNA 3/4），ROCm 为实验性支持；RX 5000/6000 请使用 CPU 版 | — |
 
 > Windows AMD 可选择实验性的 AMD 发布便携包或通过维护脚本安装；两者均要求受支持的显卡、Radeon ROCm 7.2.1 和 AMD 26.2.2 驱动。ROCm 在 Windows 上仍属实验性路径。
diff --git a/packaging/launch.py b/packaging/launch.py
diff --git a/test/test_dependency_groups.py b/test/test_dependency_groups.py
