infineon/tensorflow-gpu

TensorFlow 2.20 GPU wheel for linux_aarch64 (CUDA 12.8 / cuDNN 9.8)

Self-built tensorflow wheel for the platforms PyPI does not ship a GPU build for. Produced by scripts/build_tf_gpu_aarch64.sh in the LPWWD pipeline repo on an NVIDIA Spark / GB10 host.

Why this exists

PyPI ships a CPU-only tensorflow wheel for linux_aarch64. There is no pip-installable GPU TensorFlow on this platform/Python combo, so to get GPU acceleration without Docker the wheel has to be built from source. A cold from-source build is 2–4 h and ~50–80 GB of bazel artifacts; this repo lets every other aarch64 host skip that.

Contents

File	Size	sha256
tensorflow-2.20.0.dev0+selfbuilt-cp312-cp312-linux_aarch64.whl	~495 MiB	6c63ce87206ac1485b5858a100f098674943098da946837b77d8d6c07a7ec35b
tensorflow-2.20.0.dev0+selfbuilt-cp312-cp312-linux_aarch64.whl.sha256	—	sidecar

Build configuration

Setting	Value
TensorFlow	v2.20.0
Python	3.12 (cp312)
Platform tag	linux_aarch64 (ARM 64-bit)
CUDA	12.8 (hermetic)
cuDNN	9.8 (hermetic)
Compute capabilities	9.0 (Hopper) + 12.0 (Blackwell / GB10 sm_120)
Device compiler	nvcc
Host compiler	clang-17 (via --config=nvcc_clang)
Bazel	7.4.1
Build host	NVIDIA Spark (GB10, aarch64, 20 cores, 121.7 GiB unified memory)

This wheel will run on any linux_aarch64 host with a CUDA-12.x driver and a GPU of compute capability 9.0 or 12.0 (e.g. H100/H200/Hopper and Blackwell/GB10). Other compute capabilities are not embedded — if your device has e.g. sm_80 you need a rebuild.

Install

pip download \
    --no-deps \
    --dest . \
    "https://huggingface.co/datasets/infineon/tensorflow-gpu/resolve/main/tensorflow-2.20.0.dev0+selfbuilt-cp312-cp312-linux_aarch64.whl"
sha256sum -c <(echo "6c63ce87206ac1485b5858a100f098674943098da946837b77d8d6c07a7ec35b  tensorflow-2.20.0.dev0+selfbuilt-cp312-cp312-linux_aarch64.whl")
pip install --upgrade "./tensorflow-2.20.0.dev0+selfbuilt-cp312-cp312-linux_aarch64.whl"

Or with huggingface_hub:

from huggingface_hub import hf_hub_download
whl = hf_hub_download(
    repo_id="infineon/tensorflow-gpu",
    repo_type="dataset",
    filename="tensorflow-2.20.0.dev0+selfbuilt-cp312-cp312-linux_aarch64.whl",
)

Then verify the GPU is visible:

import tensorflow as tf
print(tf.__version__, tf.config.list_physical_devices("GPU"))

Compatibility matrix

Host arch	CUDA driver	GPU SM	Status
linux_aarch64	12.8+	sm_90 (Hopper)	OK
linux_aarch64	12.8+	sm_120 (Blackwell / GB10)	OK
linux_aarch64	12.8+	other SM	rebuild required
linux_x86_64	—	—	wrong arch; use upstream PyPI
macOS / Windows	—	—	not supported

Provenance

Built from the upstream tensorflow/tensorflow repo at tag v2.20.0 (no patches) using scripts/build_tf_gpu_aarch64.sh. The build script pins all toolchain versions (Bazel, CUDA, cuDNN, clang) and is the single source of truth — re-running it on a fresh aarch64 host with TF_VERSION=v2.20.0 reproduces this wheel bit-identically modulo timestamps.

License

TensorFlow itself is Apache-2.0. This dataset card is also Apache-2.0.