Is your feature request related to a problem? Please describe.
Installing Transformer Engine on a Linux aarch64 GB200 environment with Python 3.12, PyTorch 2.11, CUDA 13.0, and the C++11 ABI finds no compatible prebuilt transformer-engine-torch wheel.
The PyPI package intentionally ships as an sdist and attempts to download a matching framework-extension wheel from the Transformer Engine GitHub release. v2.17 currently provides only x86_64 extension assets, so ARM64 installation falls back to a native extension build. That adds substantial setup time, build dependencies, and failure modes for GB200 users.
The related CUDA 13 core-wheel regression is tracked in #3309.
Describe the solution you'd like
Publish a prebuilt transformer_engine_torch release asset for this compatibility tuple:
- CPython 3.12
- Linux aarch64
- PyTorch 2.11
- CUDA 13.0
- C++11 ABI enabled
- GB200 / SM100
The asset should follow the existing get_wheel_url() naming convention so installing transformer-engine-torch can retrieve it automatically instead of compiling locally.
Release automation should also:
- Build with the required Blackwell architecture target.
- Verify wheel platform and ABI tags.
- Run an install/import smoke check in a matching ARM64 environment.
- Preserve complete compiler output on failure.
Describe alternatives you've considered
Building the framework extension from source works in a correctly provisioned environment, but requires a full CUDA development toolchain and materially increases installation friction. Increasing build parallelism does not remove those requirements and can increase memory or disk pressure.
Additional context
Current release automation already includes ubuntu-22.04-arm, but the release matrix remains fixed to Python 3.12, PyTorch 2.8.0, and CUDA 12.9.1:
https://github.com/NVIDIA/TransformerEngine/blob/main/.github/workflows/attach-wheels-to-release.yml
Recent ARM64 release jobs have not produced assets:
The composite action currently captures docker run output through tail -n 1, which hides the underlying compiler failure from the job log.
Is your feature request related to a problem? Please describe.
Installing Transformer Engine on a Linux aarch64 GB200 environment with Python 3.12, PyTorch 2.11, CUDA 13.0, and the C++11 ABI finds no compatible prebuilt
transformer-engine-torchwheel.The PyPI package intentionally ships as an sdist and attempts to download a matching framework-extension wheel from the Transformer Engine GitHub release. v2.17 currently provides only x86_64 extension assets, so ARM64 installation falls back to a native extension build. That adds substantial setup time, build dependencies, and failure modes for GB200 users.
The related CUDA 13 core-wheel regression is tracked in #3309.
Describe the solution you'd like
Publish a prebuilt
transformer_engine_torchrelease asset for this compatibility tuple:The asset should follow the existing
get_wheel_url()naming convention so installingtransformer-engine-torchcan retrieve it automatically instead of compiling locally.Release automation should also:
Describe alternatives you've considered
Building the framework extension from source works in a correctly provisioned environment, but requires a full CUDA development toolchain and materially increases installation friction. Increasing build parallelism does not remove those requirements and can increase memory or disk pressure.
Additional context
Current release automation already includes
ubuntu-22.04-arm, but the release matrix remains fixed to Python 3.12, PyTorch 2.8.0, and CUDA 12.9.1:https://github.com/NVIDIA/TransformerEngine/blob/main/.github/workflows/attach-wheels-to-release.yml
Recent ARM64 release jobs have not produced assets:
The composite action currently captures
docker runoutput throughtail -n 1, which hides the underlying compiler failure from the job log.