Arch Linux User Repository

You have a commented out part in the PKGBUILD in which you compile the python bindings.

# python bindings (fails to build with python 3.11) #local _pyver ...

Since my GPU (1060) is no longer supported by tensorrt 10, I had to get the 9 version to work and compile the python bindings for python 3.12 myself.

All I had to do was set another ENV variable and adjust the install command.

in build()...{ ... local -x TENSORRT_MODULE="tensorrt" ... }

the generated whl is located somewhere else, therefore adapt the install command

in package_python-tensorrt() { ...

python -m installer --destdir="$pkgdir" "TensorRT/python/build/bindings_wheel/dist/"*.whl

... }

I cannot say whether everything is correct. But everything compiles and the models also run (rife+upscale via trt).

In case someone might need it.

@milianw I could compile your 'binsim.cu' source file using cuda 12.5.1 by running the exact same nvcc command that you posted in the mentioned nvidia thread. No errors, no warnings, and the 'binsimCUDA' executable builds fine. I cannot answer why you are getting these errors, and further discussing this here will be out of the scope of this AUR web page.

@dbermond: if gcc is not an issue, then why did I see the compile errors from the linked forum thread? I have gcc13 installed, but only

$ ls /usr/include/c++/
14.1.1  v1

So gcc13 will still end up using libstc++ headers from gcc14 which are incompatible. How is this supposed to work?

@milianw Sure, I will be happy to update the package if you provide the fix for this issue that I reported upstream in the same day 10.2 was released. And there is also this another one which I reported, but I could fix it myself. Please notice that if you can fix the first issue, other ones may arise later in the compilation, or even in the python modules, so make sure to check everything. Regarding the gcc usage in cuda, each cuda version uses a specific gcc version. cuda 12.5 uses gcc13 (and not gcc14), so the gcc version is not a problem for us, since the cuda package is already using the correct one.

meh, just updating the versions won't be sufficient since cuda (even 12.5 apparently) is not compatible with gcc 14 system includes: https://forums.developer.nvidia.com/t/cuda-12-4-nvcc-and-gcc-14-1-incompatibility/293295

there's now tensorrt 10.2.0.19 with cuda 12.5 support, this package could be updated accordingly

Thank you for great and speedy answer. I should've probably searched for the issue a bit, apologies for being lazy

@FuzzyAtish TLDR: it fails to build against cuda 12.5, downgrade cuda to 12.4.1 and it will work. The long story: there are two issues regarding this. Firstly, tensorrt 10.0.1 apparently does not support cuda 12.5, as you can see in the upstream documentation. Secondly, cuda 12.5 is not supported by the current version of the nvidia drivers that are available in the official repositories (cuda 12.5 was pushed without a driver that supports it), and you can read more details about this in the Arch Linux cuda package issues page at this link.

I'm getting the following error when building

#17 2311.3 [ 36%] Building CUDA object 
plugin/CMakeFiles/nvinfer_plugin.dir/roiAlignPlugin/roiAlignKernel.cu.o
#17 2320.2 home/builder/.cache/yay/tensorrt/src/TensorRT/plugin/common/common.cuh(321): 
error: identifier "FLT_MAX" is undefined
#17 2320.2       float threadData(-FLT_MAX);
#17 2320.2                         ^
#17 2320.2
#17 2320.2 home/builder/.cache/yay/tensorrt/src/TensorRT/plugin/common/common.cuh(373): 
error: identifier "FLT_MAX" is undefined
#17 2320.2       float threadData(-FLT_MAX);
#17 2320.2                         ^
#17 2320.2
#17 2320.3 2 errors detected in the compilation of "/home/builder/.cache/yay/tensorrt/src/TensorRT/plugin/roiAlignPlugin/roiAlignKernel.cu".