Trying to install torch for python 3.10.14 running on a jetson xavier nx with jet pack 5.1.3-b29 from the various versions available at Index of /compute/redist/jp and getting that “is not a supported wheel on this platform”, or that libcublas.so.*[0-9] not found, or something
If someone could tell me if there even exists, as I tried so many versions and they never seem to work
I’ve tried many times to build it from source and each time is a different error. The instructions at the page you sent me indicate this should be done:
export USE_NCCL=0
export USE_DISTRIBUTED=0
export USE_QNNPACK=0
export USE_PYTORCH_QNNPACK=0
export TORCH_CUDA_ARCH_LIST=“7.2;8.7”
I also put:
export CMAKE_CUDA_COMPILER=/usr/local/cuda-11.4/bin/nvcc
export USE_PRIORITIZED_TEXT_FOR_LD=1
Still, last time the compilation stopped with some error related to caffe.
So, if you could tell me the correct parameters to compile that’d be amazing. Thanks
UPDATE: now the error is /usr/bin/ld: error: linker script file ‘/root/pytorch/cmake/linker_script.ld’ appears multiple times
collect2: error: ld returned 1 exit status
[4615/6678] Building CXX object c10/test/CMakeFiles/c10_DeviceGuard_test.dir/core/DeviceGuard_test.cpp.o
ninja: build stopped: subcommand failed.
Problem is trying to run python 3.10 on a xavier NX. If we try to install the pytorch related to jetpack 5.13 we get the error “not a supported wheel on this platform”
And if we try to install the versions for jetpack 6 we are indeed able to install them, but we get the error “ImportError: /lib/aarch64-linux-gnu/libc.so.6: version `GLIBC_2.34’ not found”
And since updating glibc is the best way to render a system unusable, we are stuck to either using python 3.8 which is useless for stuff such as scikit, or having to buy a jetson orin
Do you want to use scikit?
If yes, we do have some container that has the library installed with Python 3.8 for JetPack 5.
To build PyTorch from the source, we can find the builder and building script in the below link.
But please note that we only provide prebuilt for the default Python version.
So these building script is tested with Python3.8 on JetPack 5 only:
The scikit version requires python 3.10. It actually powers NVIDIA FastPitch, which is the ultimate goal: to have something that speaks in a non-robotic way of the espeak I am using right now