RTX5060显卡+windows CUDA12.8+cuDNN8.9.7+pytorch安装

为什么英伟达50系列显卡要安装cuda12.8

可以看文章（https://zhuanlan.zhihu.com/p/1970666740221450142）

安装cuda

https://developer.nvidia.com/cuda-12-8-1-download-archive?target_os=Windows&target_arch=x86_64&target_version=11&target_type=exe_local

下载到本地，双击安装。过程中需要注意：自定义安装，不勾选里面的 Visual Studio Integration 。

安装cuDNN

https://developer.nvidia.com/rdp/cudnn-archive

解压缩后会看到文件目录结构如下。

测试cuda+cuDNN是否成功

进入到上面的安装文件夹下的 extra/demo_suit 文件夹(如：C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.8\extras\demo_suite）下，然后在地址栏输入 cmd，打开命名提示行。运行 bandwidthTest.exe 和 deviceQuery.exe ，结果如下，得到两个 PASS，就基本是安装成功了。

PS C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.8>cd.\extras\demo_suite\PS C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.8\extras\demo_suite>.\bandwidthTest.exe[CUDA Bandwidth Test]- Starting... Running on... Device0: NVIDIA GeForce RTX5060Ti Quick Mode Host to Device Bandwidth,1Device(s)PINNED Memory Transfers Transfer Size(Bytes)Bandwidth(MB/s)3355443212839.0Device to Host Bandwidth,1Device(s)PINNED Memory Transfers Transfer Size(Bytes)Bandwidth(MB/s)3355443213858.9Device to Device Bandwidth,1Device(s)PINNED Memory Transfers Transfer Size(Bytes)Bandwidth(MB/s)33554432370719.0Result=PASS NOTE: The CUDA Samples are not meantforperformance measurements. Results may vary when GPU Boost is enabled.

PS C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.8\extras\demo_suite>.\deviceQuery.exe C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.8\extras\demo_suite\deviceQuery.exe Starting... CUDA Device Query(Runtime API)Detected1CUDA Capable device(s)Device0:"NVIDIA GeForce RTX 5060 Ti"CUDA Driver Version / Runtime Version13.0/12.8CUDA Capability Major/Minor version number:12.0Total amount of global memory:8151MBytes(8546484224bytes)MapSMtoCoresforSM12.0is undefined. Default to use128Cores/SM MapSMtoCoresforSM12.0is undefined. Default to use128Cores/SM(36)Multiprocessors,(128)CUDA Cores/MP:4608CUDA Cores GPU Max Clock rate:2572MHz(2.57GHz)Memory Clock rate:14001Mhz Memory Bus Width:128-bit L2 Cache Size:33554432bytes Maximum Texture Dimension Size(x,y,z)1D=(131072),2D=(131072,65536),3D=(16384,16384,16384)Maximum Layered 1D Texture Size,(num)layers1D=(32768),2048layers Maximum Layered 2D Texture Size,(num)layers2D=(32768,32768),2048layers Total amount of constant memory: zu bytes Total amount of shared memory per block: zu bytes Total number of registers available per block:65536Warp size:32Maximum number of threads per multiprocessor:1536Maximum number of threads per block:1024Max dimension size of a thread block(x,y,z):(1024,1024,64)Max dimension size of a grid size(x,y,z):(2147483647,65535,65535)Maximum memory pitch: zu bytes Texture alignment: zu bytes Concurrent copy and kernel execution: Yes with1copy engine(s)Runtimelimit on kernels: Yes Integrated GPU sharing Host Memory: No Supporthostpage-locked memory mapping: Yes Alignment requirementforSurfaces: Yes Device has ECC support: Disabled CUDA Device Driver Mode(TCC or WDDM): WDDM(Windows Display Driver Model)Device supports Unified Addressing(UVA): Yes Device supports Compute Preemption: Yes Supports Cooperative Kernel Launch: Yes Supports MultiDevice Co-op Kernel Launch: No Device PCI Domain ID / Bus ID / location ID:0/1/0Compute Mode:<Default(multiplehostthreads can use ::cudaSetDevice()with device simultaneously)>deviceQuery, CUDA Driver=CUDART, CUDA Driver Version=13.0, CUDA Runtime Version=12.8, NumDevs=1, Device0=NVIDIA GeForce RTX5060Ti Result=PASS PS C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.8\extras\demo_suite>

安装pytorch

下载 Pytorch 之前，建议创建一个独立的虚拟环境，避免后续下载和更新包时发生冲突。

这里 python 版本根据自己本地的版本进行设置，可以通过 python --version 查看。需要特别注意：Pytorch 稳定版目前最高只支持 3.12，因此下载 Pytorch 稳定版时，创建虚拟环境时尽量设置 3.12 版本的 python。当然，下面就会提到 RTX 50 系列的 GPU 不用选择稳定版，具体是否能支持 3.13 不太清楚，但是 3.12 是完全没问题的。

conda create-ncu28-py133python=3.13

cu28-py133 是我自己取的虚拟环境名。然后激活虚拟环境。

conda activate cu28-py133

接下来要在该环境下下载 Pytorch。在 Pytorch 官网上选择好要下载版本后，会在 Run this Command 一栏中显示下载命令。在已激活的虚拟环境中执行此命令即可下载 Pytorch。对于非 RTX 50 系列的伙伴们，可以选择下载稳定版 Stable (2.9.0)，然后选择对应的 CUDA 版本。

如果使用的是 RTX 50 系列的 GPU，由于目前 Pytorch 的稳定版 Stable (2.9.0) 只支持到 RTX 40 系列，务必改为下载 Preview (Nightly)，并选择 CUDA 12.8 版本以支持 GPU 的计算架构版本，否则后续使用 GPU 运行代码的时候会出现 Pytorch 编译与 GPU 计算能力不兼容的问题。因为像我的 RTX 5060 计算架构版本是 sm_120，但是 stable 版本只支持到 sm_90，所以会有冲突。

(cu128-py133)C:\Users\HP>pip3install--pretorch torchvision --index-url https://download.pytorch.org/whl/nightly/cu128 Lookinginindexes: https://download.pytorch.org/whl/nightly/cu128 Requirement already satisfied: torchin.\AppData\Roaming\Python\Python313\site-packages(2.12.0.dev20260316+cu128)Requirement already satisfied: torchvisionin.\AppData\Roaming\Python\Python313\site-packages(0.26.0.dev20260316+cu128)Collecting filelock(from torch)Downloading filelock-3.25.2-py3-none-any.whl.metadata(2.0kB)Collecting typing-extensions>=4.10.0(from torch)Downloading https://download.pytorch.org/whl/nightly/typing_extensions-4.15.0-py3-none-any.whl.metadata(3.3kB)Requirement already satisfied: setuptools<82inD:\anaconda3\envs\cu128-py133\Lib\site-packages(from torch)(80.10.2)Collecting sympy>=1.13.3(from torch)Downloading sympy-1.14.0-py3-none-any.whl.metadata(12kB)Collecting networkx>=2.5.1(from torch)Downloading networkx-3.6.1-py3-none-any.whl.metadata(6.8kB)Collecting jinja2(from torch)Downloading https://download.pytorch.org/whl/nightly/jinja2-3.1.6-py3-none-any.whl.metadata(2.9kB)Collecting fsspec>=0.8.5(from torch)Downloading fsspec-2026.2.0-py3-none-any.whl.metadata(10kB)Collecting numpy(from torchvision)Downloading numpy-2.4.3-cp313-cp313-win_amd64.whl.metadata(6.6kB)Collecting pillow!=8.3.*,>=5.3.0(from torchvision)Downloading pillow-12.1.1-cp313-cp313-win_amd64.whl.metadata(9.0kB)Collecting mpmath<1.4,>=1.1.0(from sympy>=1.13.3->torch)Downloading mpmath-1.3.0-py3-none-any.whl.metadata(8.6kB)Collecting MarkupSafe>=2.0(from jinja2->torch)Downloading https://download.pytorch.org/whl/nightly/MarkupSafe-3.0.2-cp313-cp313-win_amd64.whl.metadata(4.1kB)Downloading fsspec-2026.2.0-py3-none-any.whl(202kB)Downloading networkx-3.6.1-py3-none-any.whl(2.1MB)━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━2.1/2.1 MB172.1kB/s0:00:10 Downloading pillow-12.1.1-cp313-cp313-win_amd64.whl(7.0MB)━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━7.0/7.0 MB86.1kB/s0:01:21 Downloading sympy-1.14.0-py3-none-any.whl(6.3MB)━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━6.3/6.3 MB83.5kB/s0:01:30 Downloading mpmath-1.3.0-py3-none-any.whl(536kB)━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━536.2/536.2 kB83.8kB/s0:00:05 Downloading https://download.pytorch.org/whl/nightly/typing_extensions-4.15.0-py3-none-any.whl(44kB)Downloading filelock-3.25.2-py3-none-any.whl(26kB)Downloading https://download.pytorch.org/whl/nightly/jinja2-3.1.6-py3-none-any.whl(134kB)Downloading https://download.pytorch.org/whl/nightly/MarkupSafe-3.0.2-cp313-cp313-win_amd64.whl(15kB)Downloading numpy-2.4.3-cp313-cp313-win_amd64.whl(12.3MB)━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━12.3/12.3 MB95.4kB/s0:02:37 Installing collected packages: mpmath, typing-extensions, sympy, pillow, numpy, networkx, MarkupSafe, fsspec, filelock, jinja2 Successfully installed MarkupSafe-3.0.2 filelock-3.25.2 fsspec-2026.2.0 jinja2-3.1.6 mpmath-1.3.0 networkx-3.6.1 numpy-2.4.3 pillow-12.1.1 sympy-1.14.0 typing-extensions-4.15.0

验证torch是否下载成功

开启 python，在 python 中输入以下命令。

(cu128-py133)C:\Users\HP>import torch'import'不是内部或外部命令，也不是可运行的程序 或批处理文件。(cu128-py133)C:\Users\HP>python Python3.13.12|packaged by Anaconda, Inc.|(main, Feb242026,16:05:56)[MSC v.194264bit(AMD64)]on win32 Type"help","copyright","credits"or"license"formoreinformation.>>>importtorch>>>print(torch.cuda.is_available())True>>>

若显示 True 说明 CUDA 和 Pytorch 安装成功。

备注：如果之前 RTX 50 系列的伙伴没有选择 Preview (Nightly) CUDA 12.8，在这里也会显示 True，只有在实际运行代码的时候才会发现报错。