Qanatpharma/tensorflow-2.21.0-cuda13-aarch64

TensorFlow 2.21.0 — CUDA 13 / aarch64 (DGX Spark GB10)

Pre-built TensorFlow 2.21.0 wheels for aarch64 Linux with CUDA 13 GPU support, targeting the NVIDIA DGX Spark (GB10 Blackwell).

No official TensorFlow wheel exists for this platform. These builds were compiled from source with native GPU kernels.

Quick Install

# Python 3.10
pip install https://huggingface.co/Qanatpharma/tensorflow-2.21.0-cuda13-aarch64/resolve/main/tensorflow-2.21.0-cp310-cp310-linux_aarch64.whl

# Python 3.11
pip install https://huggingface.co/Qanatpharma/tensorflow-2.21.0-cuda13-aarch64/resolve/main/tensorflow-2.21.0-cp311-cp311-linux_aarch64.whl

# Python 3.12
pip install https://huggingface.co/Qanatpharma/tensorflow-2.21.0-cuda13-aarch64/resolve/main/tensorflow-2.21.0-cp312-cp312-linux_aarch64.whl

# Python 3.13
pip install https://huggingface.co/Qanatpharma/tensorflow-2.21.0-cuda13-aarch64/resolve/main/tensorflow-2.21.0-cp313-cp313-linux_aarch64.whl

Available Wheels

Python	Wheel	Size
3.10	tensorflow-2.21.0-cp310-cp310-linux_aarch64.whl	377 MB
3.11	tensorflow-2.21.0-cp311-cp311-linux_aarch64.whl	377 MB
3.12	tensorflow-2.21.0-cp312-cp312-linux_aarch64.whl	377 MB
3.13	tensorflow-2.21.0-cp313-cp313-linux_aarch64.whl	378 MB

Build Specifications

Parameter	Value
TensorFlow	2.21.0
CUDA (hermetic)	13.0.2
cuDNN	9.20.0
Compute capability	SM 10.0 + SM 12.1 (Blackwell)
CPU architecture	aarch64 (armv9-a)
Host compiler	aarch64-linux-gnu-gcc-13
XLA	enabled
OS	Ubuntu 24.04

Target Hardware

• NVIDIA DGX Spark (GB10, 128 GB unified memory)
• Should also work on other aarch64 Linux systems with CUDA 13 and Blackwell GPUs

Verification

import tensorflow as tf
print("TF version:", tf.__version__)
print("GPUs:", tf.config.list_physical_devices("GPU"))
# Expected: [PhysicalDevice(name='/physical_device:GPU:0', device_type='GPU')]

Build Notes

Building TensorFlow for CUDA 13 + aarch64 required several workarounds:

1. Hermetic CUDA 13.0.2 for the toolkit, with local cuDNN 9.20.0 (hermetic cuDNN lacks CUDA 13 aarch64 packages)
2. nvshmem disabled (no CUDA 13 aarch64 support; not needed for single-GPU)
3. SM 12.1 kernels included (GB10 reports compute capability 12.1 under CUDA 13, not 10.0)
4. --config=cuda_wheel required (TF 2.21.0 changed the wheel build config name)

Full build instructions: BUILD.md

License

Apache 2.0 (same as TensorFlow)