how do I set the Cuda architecture to 86

[Rotoslider]

I managed to get this working on another computer last month but can not remember how to get it to select the proper type of GPU. "Failed to detect a default CUDA architecture." The instructions say " set the TCNN_CUDA_ARCHITECTURES environment variable for the GPU you would like to use." I have windows 10. I set an environment variable name: TCNN_CUDA_ARCHITECTURES and value: 86 Rebooted and still get the error. I tried to set it in the CMakeLists.txt file. This caused a different error. Where do I set the value for the 3060ti card I have in the computer. Cmake version 3.23.2 Cuda 11.7 MSVC 19.29.30145.0 Using the Developer Command prompt for VS 2019 and running: cmake . -B build I get the following error: -- Building for: Ninja -- The C compiler identification is MSVC 19.29.30145.0 -- The CXX compiler identification is MSVC 19.29.30145.0 -- The CUDA compiler identification is unknown CMake Error at C:/msys64/mingw64/share/cmake/Modules/CMakeDetermineCUDACompiler.cmake:633 (message): Failed to detect a default CUDA architecture. Compiler output: Call Stack (most recent call first): CMakeLists.txt:11 (PROJECT) -- Configuring incomplete, errors occurred!

I know this has been asked a hundred times but after reading all of the incomplete answers I still have no idea.

[unitycoder]

did you delete old build folder after setting variable? then build again.

also i had to downgrade cmake to 3.22.5 to get rid of some similar errors.. (although still havent managed to fully compile, due to other errors later)

[Rotoslider]

Yes I delete the build folder between builds. If I place the following into the CMakeLists.txt file: find_package(CUDAToolkit REQUIRED) if(NOT DEFINED CMAKE_CUDA_ARCHITECTURES) set(CMAKE_CUDA_ARCHITECTURES 86) endif() The CMAKE_CUDA_ARCHITECTURES: 86 do not all work with this compiler. If I type nvcc --version I get nvcc: NVIDIA (R) Cuda compiler driver Copyright (c) 2005-2022 NVIDIA Corporation Built on Tue_May__3_19:00:59_Pacific_Daylight_Time_2022 Cuda compilation tools, release 11.7, V11.7.64 Build cuda_11.7.r11.7/compiler.31294372_0

Seems it does not like the new nvidia compiler or the cmake compiler. I need to look into down grading cmake then I guess.

The old computer that instant-ngp built on had cmake version 3.23.2 which is the same version as the computer I cannot get to buid. The Cuda compiler that built correctly before was 11.6.124. Where now it is and 11.7.64. Seems it does not like the new toolkit. I went back to Cuda 11.6.124 and I still cannot get it to build. Compiler error as above. The C Compiler ,CMake and Cuda Compilers are the same versions as the computer before that was able to be built. Any suggestions?

[Rolliboy]

Hi! Same problem here. Have you managed to solve this error? I solved it with this: cmake . -D TCNN_CUDA_ARCHITECTURES=86 -B build

[Rotoslider]

sadly I still get the same error even with cmake . -D TCNN_CUDA_ARCHITECTURES=86 -B build Not sure whats going on. I deleted the build folder before trying it. results: `c:\Users\me\Documents\instant-ngp>cmake . -D TCNN_CUDA_ARCHITECTURES=86 -B build -- Building for: Ninja -- Found CUDAToolkit: C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v11.6/include (found version "11.6.124") -- The C compiler identification is MSVC 19.29.30145.0 -- The CXX compiler identification is MSVC 19.29.30145.0 -- The CUDA compiler identification is unknown CMake Error at C:/msys64/mingw64/share/cmake/Modules/CMakeDetermineCUDACompiler.cmake:654 (message): The CMAKE_CUDA_ARCHITECTURES:

86

do not all work with this compiler.`

[unitycoder]

i added these into top of CMakeLists.txt (note custom folder for toolkit, use our own or remove that line)

set(TCNN_CUDA_ARCHITECTURES "86")
set(CUDA_ARCHITECTURES "86")
set(CMAKE_CUDA_ARCHITECTURES "86")
set(CudaToolkitDir "D:/sdk/CUDA/")

but in the end it brought new errors and then stuck at compiling "error : identifier "cublasComputeType_t" is undefined". (might be also relevant that i have vs, cuda, cmake and other tools in non-default install locations, not in c: drive)

[unitycoder]

so finally got it to run, here's my notes.. https://gist.github.com/unitycoder/7e59247ad0f36f77bc8402cd9a22d4fb

but i think my solution was to install everything on default locations.. (had to get bigger c: drive, reinstalled windows, visual studio 2019, cuda 11.4, cmake 3.22.5 all in default c: locations)

[jundaf2]

I encountered a similar issue and solved by setting CUDACXX to the correct executable. which is ...\cuda\bin\nvcc.

[alandevkota]

I encountered a similar issue and solved by setting CUDACXX to the correct executable. which is ...\cuda\bin\nvcc.

Hi! I am new to cmake, don't have much knowledge about it. How do I set set the executable like you said??

[shoeychi]

set ~/.bashrc

export CUDACXX=/usr/local/cuda/bin/nvcc

[jundaf2]

Hi, @shoeychi gives the solution, sorry for not replying, I missed the message.

[alandevkota]

Thanks!!

[baughmann]

Ensure that the CUDA bins are in your path. For me, they were installed to /usr/local/cuda-12/bin/ which was not in my path, so CMake couldn't find it while building llama.cpp.

Simply add a line like the following to your .zshrc or .bashrc, or add /usr/local/cuda-12/bin: to the end before $PATH if you already have a line like this.

export PATH=/usr/local/cuda-12/bin:$PATH