Request: Provide +cpu wheels for ARM64 (aarch64) in official TorchVision releases

[DHolmanCoding]

🚀 The feature

Please publish ARM64 (aarch64) CPU-only +cpu wheels for TorchVision to the official PyTorch CPU index at https://download.pytorch.org/whl/torchvision/ so they mirror the existing x86_64 +cpu artifacts (and match the torch CPU-only availability, which already includes aarch64).

Concretely:

Build and upload torchvision==+cpu wheels for manylinux aarch64 (and macOS arm64 if you use the +cpu channel there).

Keep version/ABI pairing aligned with the corresponding torch==+cpu wheels.

Document the install matrix (x86_64 & aarch64 parity) in the install guide.

This makes the Torch/TorchVision packaging matrix symmetric across architectures and channels, and avoids special-casing ARM in dependency managers.

Motivation, pitch

I’m building lightweight CPU-only deployments on ARM64 devices (aarch64). The goal is to ship small images without CUDA/NVIDIA baggage. Today there’s an asymmetry:

torch provides aarch64 +cpu wheels on the CPU index.

torchvision does not provide aarch64 +cpu wheels on that index; ARM64 wheels exist only as plain (no +cpu) on PyPI.

Alternatives

No response

Additional context

No response

[atalman]

@DHolmanCoding Please try https://download.pytorch.org/whl/test/cpu/torchvision/index.html These are all CPU aarch64 builds

[DHolmanCoding]

@DHolmanCoding Please try https://download.pytorch.org/whl/test/cpu/torchvision/index.html These are all CPU aarch64 builds

I plugged this into my build multi-architecture build and observe that on linux/amd it finds a compatible version:

#19 16.06 [urllib3:urllib3.connectionpool] [https://download.pytorch.org:443](https://download.pytorch.org/) "GET /whl/test/cpu/torchvision-0.23.0%2Bcpu-cp311-cp311-manylinux_2_28_x86_64.whl HTTP/1.1" 200 1911930

Then on the linux/arm build it appears reject all the versions that are available:

#23 152.6 Skipping wheel torchvision-0.23.0+cpu-cp310-cp310-manylinux_2_28_x86_64.whl as this is not supported by the current environment
#23 152.6 Skipping wheel torchvision-0.23.0+cpu-cp310-cp310-win_amd64.whl as this is not supported by the current environment
#23 152.7 Skipping wheel torchvision-0.23.0+cpu-cp311-cp311-manylinux_2_28_x86_64.whl as this is not supported by the current environment
#23 152.7 Skipping wheel torchvision-0.23.0+cpu-cp311-cp311-win_amd64.whl as this is not supported by the current environment
#23 152.7 Skipping wheel torchvision-0.23.0+cpu-cp312-cp312-manylinux_2_28_x86_64.whl as this is not supported by the current environment
#23 152.7 Skipping wheel torchvision-0.23.0+cpu-cp312-cp312-win_amd64.whl as this is not supported by the current environment
#23 152.7 Skipping wheel torchvision-0.23.0+cpu-cp313-cp313-manylinux_2_28_x86_64.whl as this is not supported by the current environment
#23 152.7 Skipping wheel torchvision-0.23.0+cpu-cp313-cp313-win_amd64.whl as this is not supported by the current environment
#23 152.7 Skipping wheel torchvision-0.23.0+cpu-cp313-cp313t-manylinux_2_28_x86_64.whl as this is not supported by the current environment
#23 152.7 Skipping wheel torchvision-0.23.0+cpu-cp313-cp313t-win_amd64.whl as this is not supported by the current environment
#23 152.7 Skipping wheel torchvision-0.23.0+cpu-cp39-cp39-manylinux_2_28_x86_64.whl as this is not supported by the current environment
#23 152.7 Skipping wheel torchvision-0.23.0+cpu-cp39-cp39-win_amd64.whl as this is not supported by the current environment

and reports Unable to find installation candidates for torchvision (0.23.0+cpu)