Issue with torchaudio Undefined Symbol (_ZNK5torch8autograd4Node4nameEv) when using GPU-enabled PyTorch on Jetson Orin nano super for STT

[1.samuelmiklos]

Hi all,

I’m working on a Speech-to-Text project on a Jetson Orin Nano Super running JetPack 6.2. My goal is to run a custom Whisper-based STT model that has been fine-tuned. The project involves real-time transcription of incoming audio (aviation communication) and leveraging GPU acceleration for low-latency inference.

My Setup:

Device: NVIDIA Jetson Orin nano super
JetPack Version: 6.2
Python Version: 3.10.12
PyTorch: I installed a GPU-enabled PyTorch wheel provided by NVIDIA for JetPack 6.1/6.2: torch-2.5.0a0+872d972e41.nv24.08-cp310-cp310-linux_aarch64.whl -This version correctly enables CUDA (verified via torch.cuda.is_available()).
torchaudio: I need torchaudio to handle audio input in my STT pipeline. However, when I try to install torchaudio (using the pre-built wheels from PyPI or otherwise), it pulls in a CPU-only version of torch or builds against a different ABI. I then get the error when I run my code:

OSError: /home/Orin/.local/lib/python3.10/site-packages/torchaudio/lib/libtorchaudio.so: undefined symbol: _ZNK5torch8autograd4Node4nameEv

This error indicates that the torchaudio wheel I installed is not ABI-compatible with my GPU-enabled PyTorch build. I’ve attempted the following:

1. Installed the specific GPU-enabled torch (2.5.0a0+872d972e41.nv24.08) manually.
2. Tried installing torchaudio from available wheels (e.g., torchaudio-2.5.0-cp310-cp310-linux_aarch64.whl) with the --no-deps flag, but the resulting torchaudio still fails to load due to the undefined symbol error.

What I Need Help With:

• Is there an official GPU-enabled torchaudio wheel (or recommended build) that matches my GPU-enabled torch (2.5.0a0+872d972e41.nv24.08) for Jetson Orin running JP6.2?
• If not, what is the best practice for building torchaudio from source on this platform so that it is ABI-compatible with the torch version I’m using?
• Any additional tips for setting up a stable environment for running my custom Whisper STT model on Jetson Orin would be greatly appreciated.

I have verified that torch runs with CUDA support (using the GPU-enabled torch wheel) when torchaudio is not present. The issue arises only when torchaudio is installed and imported.

Thanks in advance for your assistance!

Best regards,
Samuel

[AastaLLL]

Hi,

Could you try the packages in the below links:

https://pypi.jetson-ai-lab.dev/jp6/cu126

Not sure if the TorchAudio has the function you need or not.
If it doesn’t, could you try to build it with the command mentioned in the below link?

packages/pytorch/torchaudio/build.sh

Thanks.