Why Was Installing tiny-cuda-nn & Nerfstudio on My RTX 3060 Such a Nightmare? (And How I Finally Fixed It)

[JohnsonManuel]

Installing tiny-cuda-nn & Nerfstudio on Windows (RTX 3060 + CUDA 11.8)

After spending over 48 hours troubleshooting, I finally got tiny-cuda-nn and Nerfstudio installed successfully. Here's what worked for me.

---

System Specs & Versions

• GPU: NVIDIA RTX 3060
• CUDA Version: 11.8
• Visual Studio Version: VS 2019 v17.8 (Worked!)
• Python Version: 3.8

---

Issues Faced & Fixes

1️⃣ TCNN_CUDA_ARCHITECTURES Error

Issue: Compilation failure due to missing CUDA architectures.

Fix: Set the CUDA architecture to 86:

set TCNN_CUDA_ARCHITECTURES=86

2️⃣ VS 2022 v17.10+ Not Working

Issue: CUDA (11.8) requires _MSC_VER < 1910 || _MSC_VER >= 1940, but VS 2022 sets _MSC_VER = 1940.

Fix: Removed VS 2022 v17.10+ and installed VS 2022 Fall 2023 LTSC (v17.8).

Either during installation of tinycudann or nerfstudio using pip, if it fails because of an error message mentioning Rust/Cargo, download and execute the installer from rustup.rs and try again.

3️⃣ CMake and cl Not Recognized

Issue: Running cmake or cl returned "not recognized as a command."

Fix: Manually added the following paths to the system PATH:

set PATH=C:\Program Files\Microsoft Visual Studio\2022\Enterprise\Common7\IDE\CommonExtensions\Microsoft\CMake\CMake\bin;%PATH%
set PATH=C:\Program Files (x86)\Microsoft Visual Studio\2022\BuildTools\VC\Tools\MSVC\14.38.33130\bin\Hostx64\x86;%PATH%

Then activated the MSVC environment:

call "C:\Program Files\Microsoft Visual Studio\2022\BuildTools\VC\Auxiliary\Build\vcvars64.bat"

4️⃣ Conda Environment Variables Not Set

Issue: CUDA was not detected properly inside Conda.

Fix: Manually set the following environment variables before installation:

set CUDAVER=11.8
set CUDA_HOME=%CONDA_PREFIX%
set CUDA_ROOT=%CONDA_PREFIX%
set PATH=%CONDA_PREFIX%\Library\bin;%PATH%
set LD_LIBRARY_PATH=%CONDA_PREFIX%\Library\lib;%LD_LIBRARY_PATH%
set LD_LIBRARY_PATH=%CONDA_PREFIX%\Library\lib64;%LD_LIBRARY_PATH%
set CUDA_HOST_COMPILER=C:\Program Files\Microsoft Visual Studio\2019\Community\VC\Tools\MSVC\14.3X.X\bin\Hostx64\x64\cl.exe

💡 To make these settings permanent, add them to your System Environment Variables.

5️⃣ Created a Conda Environment for Nerfstudio

conda create --name nerfstudio -y python=3.8
conda activate nerfstudio
python -m pip install --upgrade pip

6️⃣ Installed PyTorch with CUDA 11.8

pip install torch==1.12.1+cu113 torchvision==0.13.1+cu113 -f https://download.pytorch.org/whl/torch_stable.html

7️⃣ Installed tiny-cuda-nn via GitHub

Initially, the installation failed. However, after running:

pip install git+https://github.com/NVlabs/tiny-cuda-nn/#subdirectory=bindings/torch

it finally worked.

8️⃣ Installed Nerfstudio

pip install nerfstudio

---

🎉 Success!

Now, both tiny-cuda-nn and Nerfstudio are working perfectly! Hopefully, this helps someone else facing similar issues. 🚀