Compatibility issues with torchaudio on NVIDIA Blackwell (GB10) / aarch64 using NGC PyTorch 25.12

[chenqy2]

Hello,
I am working on the new NVIDIA DGX Spark system (equipped with Blackwell GB10 GPU and aarch64 CPU). I am using the latest NVIDIA NGC container: nvcr.io/nvidia/pytorch:25.12-py3, which comes with a custom PyTorch 2.10.0a0 build.
I am facing critical issues when trying to use torchaudio in this environment:
PIP Binary Issue: A standard pip install torchaudio fetches the +cpu version because there are no pre-compiled +cuXXX wheels available for the aarch64 architecture on PyPI.
ABI Mismatch: If I attempt to force-install a GPU-enabled wheel from other sources, I encounter OSError: undefined symbol (e.g., _ZNK3c106SymInt22maybe_as_int_slow_pathEv), likely due to an ABI mismatch between NVIDIA's custom PyTorch build and the community torchaudio builds.
Source Build Failure: When trying to build torchaudio (main branch) from source inside the container to match the local PyTorch ABI, the build fails because it cannot find certain headers. It seems the header layout in the NGC PyTorch 2.10.0a0 build has changed (e.g., missing torch/csrc/stable/device.h or torch/headeronly/core/TensorAccessor.h).
Question:
Is there a plan to provide official aarch64 wheels with CUDA support for the Blackwell architecture? Alternatively, is there a recommended way to build torchaudio from source against these specific NVIDIA NGC alpha builds?
Any guidance on how to get a functional GPU-accelerated torchaudio on GB10/aarch64 would be greatly appreciated.