Skip to content

Latest commit

 

History

History
191 lines (127 loc) · 8.05 KB

NVIDIA-win.md

File metadata and controls

191 lines (127 loc) · 8.05 KB
Raw

install NVIDIA’s deep learning stack on Windows:
CUDA toolkit + cuDNN + TensorRT

✍️ inspired from https://david-littlefield.medium.com/how-to-install-the-nvidia-cuda-driver-toolkit-cudnn-and-tensorrt-on-windows-10-3fcf97e54522

i wrote this guide coz nvidia docs suck

📑 in case u still somehow need official docs:

📑 in case u need info about cuda compute capability:

prepare at least 15gb disk space (10gb msvc + 5gb nvidia)

tested combination: Visual Studio v17 (2022) + CUDA v12.6 + cuDNN v9.4 + TensorRT v10.4

🔖 easy 1st steps with graphical interface 📱

0️⃣ NVIDIA driver

⏬ download: https://www.nvidia.com/download/index.aspx

alternatively install companion software:

each cuda version requires a minimum driver version, see: https://docs.nvidia.com/cuda/cuda-toolkit-release-notes/index.html

1️⃣ C++ tools set (CMake, ninja, cl, MSBuild, vcpkg)

if need compile binaries with cuda

⏬ download either:

⚠️ Visual Studio (VS) not visual studio code (vscode), the latter has nothing to do in all my guide

👉 when install, select checkbox “Desktop Development with C++”

🔎 verify after install: in Start menu, there’re 2 new items “Developer Command Prompt for VS 2022” & “Developer PowerShell for VS 2022”

if u open any of those 2 and run any of these following commands, it should return messages instead of errors

  • cmake --version returns 3.2█.█-msvc1
  • cl returns 19.██.███
  • msbuild -ver returns 17.█.█
  • ninja --version returns 1.11.█

see where to find header files: echo %INCLUDE% (cmd) or echo $env:INCLUDE (pwsh)

2️⃣ CUDA toolkit

⏬ download:

👉 when install, select “Advanced” → select at least “Development” + “Runtime”, and if Visual Studio installed “VS integration”

🔎 verify after install: open System Properties > tab Advanced > Environment Variables > system

  • should have %CUDA_PATH% and %CUDA_PATH_V12_6% set to C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.6
  • %PATH% should contain %CUDA_PATH%\bin & %CUDA_PATH%\libnvvp but not other things from %CUDA_PATH%
  • optional: if not exist, set %CUDA_HOME% & %CUDA_ROOT% same as %CUDA_PATH%
  • optional: if exist %LD_LIBRARY_PATH% add %CUDA_PATH%\lib\x64

🔎 also verify if nvcc --version return correct cuda version

N.B. nvidia-smi show max cuda version of driver, even if cuda not installed

CMake CUDA fix

only needed if install Visual C++ Build Tools (instead of the full Visual Studio)

copy all 4 files from C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.6\extras\visual_studio_integration\MSBuildExtensions to the directory: C:\Program Files (x86)\Microsoft Visual Studio\2022\BuildTools\MSBuild\Microsoft\VC\v170\BuildCustomizations

🔖 convoluted steps without installer 🗿

⚠️ need NVIDIA acc + join Developer Program

⚠️ open an admin cmd console to use throughout following steps

if u prefer pwsh instead, in following steps replace COPY with Copy-Item and %CUDA_PATH% with $env:CUDA_PATH

N.B. native Windows consoles are case-insensitive (msys2/cygwin is not native)

⚠️ if multiple cuda versions co-exist, in following steps replace CUDA_PATH with the corresponding CUDA_PATH_V1█_█

N.B. call CUDA_PATH in double quotes coz path contain whitespaces

📑 core principle of following steps:

files copy to
.dll %CUDA_PATH%\bin
.lib %CUDA_PATH%\lib
.h %CUDA_PATH%\include

3️⃣ cuDNN

update: new graphical installer starting from v9

⏬ download either:

👉 extract zip to a location, navigate console to the extracted folder (in some old cudnn version need to descend into .\cuda) then run

COPY bin\cudnn*.dll     "%CUDA_PATH%\bin"
COPY lib\x64\cudnn*.lib "%CUDA_PATH%\lib\x64"
COPY include\cudnn*.h   "%CUDA_PATH%\include"

🐍 boost PyTorch with cudnn fix

pytorch embedded with a minimal cuda runtime but usually not shipped with most recent cudnn, so users can replace pytorch linked .dll files to get more boost from recent update

get pytorch cudnn version: python -c "import torch; print(torch.backends.cudnn.version())"

get pytorch lib directory: python -c "import torch; print(torch.__path__[0] + r'\lib')"

copy 7 cudnn .dll files to pytorch lib directory

🚩 zlib

N.B. this library is in official docs but poorly explained, should be improved in future version of cudnn

update: no need this step starting from cudnn v9

⏬ download: https://www.winimage.com/zLibDll/zlib123dllx64.zip

👉 extract zip to a location, navigate console to in the extracted folder then run

COPY dll_x64\zlibwapi.dll "%CUDA_PATH%\bin"
COPY dll_x64\zlibwapi.lib "%CUDA_PATH%\lib\x64"

N.B. official code to build from source: https://zlib.net/ (must rename as zlibwapi)

4️⃣ TensorRT

⏬ download: https://developer.nvidia.com/tensorrt-download

if latest version has multiple choices: select “GA” (general availability) not “EA” (early access) nor “RC” (release candidate)

👉 extract zip to a location, navigate console to in the extracted folder then run

COPY bin\trtexec.exe "%CUDA_PATH%\bin"
COPY lib\nv*.dll     "%CUDA_PATH%\bin"
COPY lib\nv*.lib     "%CUDA_PATH%\lib\x64"
COPY include\Nv*.h   "%CUDA_PATH%\include"

🐍 python packages for tensorrt (optional)

only if u need those python pkgs

no need admin console, remember to activate if u use conda/venv

pip install graphsurgeon\graphsurgeon-….whl
pip install uff\uff-….whl
pip install onnx_graphsurgeon\onnx_graphsurgeon-….whl
pip install python\tensorrt-….whl

alternatively pip install onnx-graphsurgeon polygraphy uff --extra-index-url=https://pypi.ngc.nvidia.com
(or https://developer.download.nvidia.com/compute/redist)

N.B. pip install tensorrt only available on Linux; for Windows need version ≥9

other possible libraries with cuda

see full list

for e.g.

cuda header files: https://gitlab.com/nvidia/headers

appendix 1: deliverable links:

appendix 2: other hardware toolkits: