TensorFlow for Windows ARM64 (Unofficial Build)

This repository contains unofficial prebuilt TensorFlow wheels for Windows on ARM64.

Important

These wheels are not official TensorFlow releases. They are community-built binaries compiled from the official TensorFlow source code and are not affiliated with, endorsed by, or supported by Google or the TensorFlow project.

Build Information

Item	Value
TensorFlow Version	2.22.0-dev0+selfbuilt
Platform	Windows ARM64 (win_arm64)
Supported Python	Python 3.13 (ARM64)
Compiler	LLVM clang-cl 22
Build System	Bazel 7.7.0

Features

• Native TensorFlow wheel for Windows ARM64
• Native oneDNN v3 support enabled
• Arm Compute Library (ACL) accelerated GEMM through oneDNN
• XLA CPU support enabled

---

Installing Python 3.13 (ARM64)

1. Download the latest Windows ARM64 installer from the official Python website:

   https://www.python.org/downloads/windows/

2. Run the installer.

   • Enable Add Python to PATH.
   • Select Install Now.

3. Verify the installation:

python --version
pip --version

Expected output:

Python 3.13.x

---

Installation

Step 1: Install the TensorFlow wheel

Download the wheel from the repository Releases page.

Install it without dependencies:

pip install tensorflow-2.22.0.dev0+selfbuilt-cp313-cp313-win_arm64.whl --no-deps

---

Step 2: Install the required Python dependencies

pip install absl-py astunparse flatbuffers gast google_pasta keras-nightly libclang ml_dtypes numpy opt_einsum packaging protobuf requests setuptools six termcolor typing_extensions wrapt

---

Step 3: Install h5py

pip install --only-binary=:all: h5py

---

Step 4: Install grpcio

Download the Windows ARM64 wheel from:

https://github.com/khmyznikov/PyEnv-WoA-State/releases/download/grpcio-1.84.0-win_arm64/grpcio-1.84.0.dev0-cp313-cp313-win_arm64.whl

Then install it:

pip install grpcio-1.84.0.dev0-cp313-cp313-win_arm64.whl

---

Verify the Installation

import tensorflow as tf

print(tf.__version__)
print(tf.version.COMPILER_VERSION)
print(tf.config.list_physical_devices())

Expected output:

2.22.0-dev0+selfbuilt
MSVC 195136248
[PhysicalDevice(name='/physical_device:CPU:0', device_type='CPU')]

To verify that oneDNN is enabled, run TensorFlow with:

$env:ONEDNN_VERBOSE=2
$env:TF_ENABLE_ONEDNN_OPTS=1

The output should contain entries similar to:

onednn_verbose,v1,info,oneDNN v3.x
onednn_verbose,v1,primitive,exec,cpu,matmul,gemm:acl

which indicates that oneDNN is using the Arm Compute Library (ACL) backend for optimized CPU operations.

---

Source Modifications

This build is based on the official TensorFlow source code and includes modifications required to enable native Windows ARM64 support.

Major changes include:

• Added native Windows ARM64 (win_arm64) Bazel platform and build configurations.
• Added LLVM clang-cl toolchain support for Windows ARM64.
• Added Windows ARM64 support for TensorFlow's legacy rules_python (pip_parse) dependency resolution.
• Updated platform detection and CPU feature handling for Windows ARM64.
• Added Windows ARM64 support for XLA CPU target machine configuration.
• Updated TensorFlow build rules to recognize and build for Windows ARM64.
• Enabled native oneDNN support on Windows ARM64, including Arm Compute Library (ACL) acceleration.

---

License

TensorFlow is licensed under the Apache License 2.0.

This repository redistributes binaries built from the official TensorFlow source code.

To comply with the Apache License 2.0:

• The original LICENSE file from TensorFlow is included in this repository.
• Existing copyright and license notices have been preserved.
• Source-level modifications made for Windows ARM64 support are documented above.

This repository does not intentionally introduce additional third-party components beyond those already included by TensorFlow. Any future additions will include the appropriate license notices.

---

Disclaimer

This project is an unofficial community build of TensorFlow for Windows ARM64.

It is not affiliated with, endorsed by, or supported by Google or the TensorFlow project. Issues specific to these binaries should be reported in this repository rather than the official TensorFlow project.