Onnxruntime 1.17.3 compilation error on Jetpack 6.0


I am getting compilation errors trying to compile onnxruntime 1.17.3 against cuda 12.2 and tensorrt on our agx orin 64gb devkit. I am using the c++ bindings for onnxruntime as well as the cuda and tensorrt executionproviders, so I have no option but to compile from source. I have tried asking on the onnxruntime git repo as well, but a similar issue has been open for over a month now and has no reply, so I don’t have that much hope of it getting solved over there.


The build seems to fail halfway through the build on the cuda provider, giving the message nvcc fatal : 'ON': expected a number
I am trying to compile onnxruntime 1.17.3 against cuda 12.2 (preinstalled version on the jetpack 6.0.0 release, as well as tensorrt).

For my application I need to be able to use tensorrt, due to the optimizations available.


To me: relatively high. :)
Our development would get stuck on this issue.

Target system

Linux ubuntu 5.15.136-tegra #1 SMP PREEMPT Mon May 6 09:56:39 PDT 2024 aarch64 aarch64 aarch64 GNU/Linux
Jetpack 6.0.0

# R36 (release), REVISION: 3.0, GCID: 36106755, BOARD: generic, EABI: aarch64, DATE: Thu Apr 25 03:14:05 UTC 2024 \# KERNEL_VARIANT: oot

build script

cmake \
  -Donnxruntime_ENABLE_LANGUAGE_INTEROP_OPS=true \
  -Donnxruntime_ENABLE_CUDA_PROFILING=true \
  -Donnxruntime_USE_CUDA=true \
  -Donnxruntime_BUILD_FOR_NATIVE_MACHINE=true \
  -Donnxruntime_USE_TENSORRT=true \
  -Donnxruntime_BUILD_SHARED_LIB=true \
  -S /opt/onnxruntime/onnxruntime-1.17.3/cmake \
  -B /opt/onnxruntime/onnxruntime-1.17.3/build

compiler version

nvcc --version

Cuda compilation tools, release 12.2, V12.2.140
Build cuda_12.2.r12.2/compiler.33191640_0

gcc --version

gcc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0


Terminal output:

[ 51%] Building CUDA object CMakeFiles/onnxruntime_providers_cuda.dir/opt/onnxruntime/onnxruntime-1.17.3/onnxruntime/core/providers/cuda/activation/activations_impl.cu.o
nvcc fatal   : 'ON': expected a number
gmake[2]: *** [CMakeFiles/onnxruntime_providers_cuda.dir/build.make:1365: CMakeFiles/onnxruntime_providers_cuda.dir/opt/onnxruntime/onnxruntime-1.17.3/onnxruntime/core/providers/cuda/activation/activations_impl.cu.o] Error 1
gmake[1]: *** [CMakeFiles/Makefile2:2117: CMakeFiles/onnxruntime_providers_cuda.dir/all] Error 2
gmake: *** [Makefile:166: all] Error 2

makefile code at the location the crash seems to happen:

/usr/local/cuda-12.2/bin/nvcc -forward-unknown-to-host-compiler $(CUDA_DEFINES) $(CUDA_INCLUDES) $(CUDA_FLAGS) -MD -MT CMakeFiles/onnxruntime_providers_cuda.dir/opt/onnxruntime/onnxruntime-1.17.3/onnxruntime/core/providers/cuda/activation/activations_impl.cu.o -MF CMakeFiles/onnxruntime_providers_cuda.dir/opt/onnxruntime/onnxruntime-1.17.3/onnxruntime/core/providers/cuda/activation/activations_impl.cu.o.d -x cu -c /opt/onnxruntime/onnxruntime-1.17.3/onnxruntime/core/providers/cuda/activation/activations_impl.cu -o CMakeFiles/onnxruntime_providers_cuda.dir/opt/onnxruntime/onnxruntime-1.17.3/onnxruntime/core/providers/cuda/activation/activations_impl.cu.o

@j.heuver I pretty regularly build onnxruntime from source in container, you can see how I do it in build script here:

I haven’t specifically built onnxruntime 1.17.3, but have compiled 1.17.0 on JetPack 6 this way.

Thanks for the script, I will try to get that set up this afternoon! I’ll post whether it does work for me or not. I have tried building the pythonwheel(suggested on the onnxruntime website for installation) in the meantime too, but ran into other compilation errors there.