infineon/tensorflow-gpu

1---2license: apache-2.03language:4  - en5tags:6  - tensorflow7  - tensorflow-gpu8  - aarch649  - arm6410  - linux-aarch6411  - cuda12  - cuda-1213  - cudnn-914  - gpu15  - blackwell16  - gb1017  - sm_9018  - sm_12019  - python-3.1220  - wheel21  - selfbuilt22pretty_name: TensorFlow GPU wheels for linux_aarch64 (CUDA 12.8 / cuDNN 9.8)23size_categories:24  - n<1K25---26 27# TensorFlow 2.20 GPU wheel for linux_aarch64 (CUDA 12.8 / cuDNN 9.8)28 29Self-built `tensorflow` wheel for the platforms PyPI does **not** ship a30GPU build for. Produced by31[`scripts/build_tf_gpu_aarch64.sh`](https://github.com/Infineon/lpwwd/blob/main/scripts/build_tf_gpu_aarch64.sh)32in the LPWWD pipeline repo on an NVIDIA Spark / GB10 host.33 34## Why this exists35 36PyPI ships a CPU-only `tensorflow` wheel for `linux_aarch64`. There is no37pip-installable GPU TensorFlow on this platform/Python combo, so to get38GPU acceleration without Docker the wheel has to be built from source.39A cold from-source build is 2–4 h and ~50–80 GB of bazel artifacts; this40repo lets every other aarch64 host skip that.41 42## Contents43 44| File | Size | sha256 |45|---|---:|---|46| `tensorflow-2.20.0.dev0+selfbuilt-cp312-cp312-linux_aarch64.whl` | ~495 MiB | `6c63ce87206ac1485b5858a100f098674943098da946837b77d8d6c07a7ec35b` |47| `tensorflow-2.20.0.dev0+selfbuilt-cp312-cp312-linux_aarch64.whl.sha256` | — | sidecar |48 49## Build configuration50 51| Setting | Value |52|---|---|53| TensorFlow | `v2.20.0` |54| Python | 3.12 (cp312) |55| Platform tag | `linux_aarch64` (ARM 64-bit) |56| CUDA | 12.8 (hermetic) |57| cuDNN | 9.8 (hermetic) |58| Compute capabilities | `9.0` (Hopper) + `12.0` (Blackwell / GB10 `sm_120`) |59| Device compiler | `nvcc` |60| Host compiler | `clang-17` (via `--config=nvcc_clang`) |61| Bazel | 7.4.1 |62| Build host | NVIDIA Spark (GB10, aarch64, 20 cores, 121.7 GiB unified memory) |63 64This wheel will run on any `linux_aarch64` host with a CUDA-12.x driver65and a GPU of compute capability 9.0 or 12.0 (e.g. H100/H200/Hopper and66Blackwell/GB10). Other compute capabilities are not embedded — if your67device has e.g. `sm_80` you need a rebuild.68 69## Install70 71```bash72pip download \73    --no-deps \74    --dest . \75    "https://huggingface.co/datasets/infineon/tensorflow-gpu/resolve/main/tensorflow-2.20.0.dev0+selfbuilt-cp312-cp312-linux_aarch64.whl"76sha256sum -c <(echo "6c63ce87206ac1485b5858a100f098674943098da946837b77d8d6c07a7ec35b  tensorflow-2.20.0.dev0+selfbuilt-cp312-cp312-linux_aarch64.whl")77pip install --upgrade "./tensorflow-2.20.0.dev0+selfbuilt-cp312-cp312-linux_aarch64.whl"78```79 80Or with `huggingface_hub`:81 82```python83from huggingface_hub import hf_hub_download84whl = hf_hub_download(85    repo_id="infineon/tensorflow-gpu",86    repo_type="dataset",87    filename="tensorflow-2.20.0.dev0+selfbuilt-cp312-cp312-linux_aarch64.whl",88)89```90 91Then verify the GPU is visible:92 93```python94import tensorflow as tf95print(tf.__version__, tf.config.list_physical_devices("GPU"))96```97 98## Compatibility matrix99 100| Host arch | CUDA driver | GPU SM | Status |101|---|---|---|---|102| `linux_aarch64` | 12.8+ | `sm_90` (Hopper) | OK |103| `linux_aarch64` | 12.8+ | `sm_120` (Blackwell / GB10) | OK |104| `linux_aarch64` | 12.8+ | other SM | rebuild required |105| `linux_x86_64` | — | — | wrong arch; use upstream PyPI |106| `macOS` / Windows | — | — | not supported |107 108## Provenance109 110Built from the upstream `tensorflow/tensorflow` repo at tag `v2.20.0`111(no patches) using112[`scripts/build_tf_gpu_aarch64.sh`](https://github.com/Infineon/lpwwd/blob/main/scripts/build_tf_gpu_aarch64.sh).113The build script pins all toolchain versions (Bazel, CUDA, cuDNN, clang)114and is the single source of truth — re-running it on a fresh aarch64115host with `TF_VERSION=v2.20.0` reproduces this wheel bit-identically116modulo timestamps.117 118## License119 120TensorFlow itself is Apache-2.0. This dataset card is also Apache-2.0.121