Mac High Sierra 10.13.4（AMD Radeon Pro 450）TensorFlow GPU配置求助

Hey there! Let's tackle this for your MacBook Pro setup—since you're rocking an AMD Radeon Pro 450 and macOS High Sierra 10.13.4, let's cut straight to what works (and what's a dead end) for GPU-accelerated TensorFlow.

• CUDA is exclusive to NVIDIA GPUs, so your AMD Radeon Pro 450 can't use it at all. Don't waste time going down that path—it won't work for your hardware.

TensorFlow used to have experimental OpenCL support for AMD GPUs, but it's not included in official pre-built packages. You'll need to compile TensorFlow from source with OpenCL enabled, and here's how to do it for your High Sierra setup:

1. Prep Your Dependencies First #

• Install Xcode 9.4.1 (the last version fully compatible with High Sierra 10.13.4) and its command line tools:

  xcode-select --install
  

• If you don't have Homebrew installed, grab it first, then use it to install these required tools:

  brew install python3 bazel@0.15.0 opencl-headers
  

  Note: Bazel 0.15.0 is critical here—newer Bazel versions don't play nice with High Sierra or the older TensorFlow branch we'll use.

2. Clone the Right TensorFlow Version #

You'll need a TensorFlow release that still includes OpenCL support—TensorFlow 1.12.x is the last major version with experimental OpenCL support. Clone that specific branch:

git clone https://github.com/tensorflow/tensorflow.git
cd tensorflow
git checkout r1.12

3. Configure TensorFlow for OpenCL #

Run the configuration script to set up build options:

./configure

When prompted:

• Accept most defaults, but when asked about OpenCL support, select y (yes).
• Point it to your OpenCL headers (usually /usr/local/include if you installed via Homebrew).
• Make sure to set Python 3 as your default interpreter.

4. Compile TensorFlow with OpenCL #

Compile the pip package using Bazel—this will take 30-60 minutes depending on your CPU, so grab a coffee:

bazel build --config=opt --config=opencl //tensorflow/tools/pip_package:build_pip_package

Once compilation finishes, build the actual pip package:

./bazel-bin/tensorflow/tools/pip_package/build_pip_package /tmp/tensorflow_pkg

5. Install Your Custom TensorFlow Build #

Install the compiled package with pip3:

pip3 install /tmp/tensorflow_pkg/tensorflow-1.12.*.whl

6. Verify GPU Acceleration is Working #

Run this quick test script to check if TensorFlow detects your AMD GPU:

import tensorflow as tf
# Run a simple computation
with tf.Session() as sess:
    a = tf.constant([1.0, 2.0, 3.0], shape=[3], name='a')
    b = tf.constant([1.0, 2.0, 3.0], shape=[3], name='b')
    c = tf.add(a, b)
    print(sess.run(c))
# Check detected devices
from tensorflow.python.client import device_lib
print(device_lib.list_local_devices())

Look for a device entry labeled device_type: "GPU" that references your Radeon Pro 450—this confirms GPU acceleration is active.

• OpenCL support in TensorFlow 1.x is experimental, so you might hit minor bugs or see less performance than you would with NVIDIA's CUDA. But it's the only game in town for AMD GPUs on older macOS versions.
• High Sierra 10.13.4 is pretty outdated, so stick strictly to the dependency versions we listed—mismatches are the #1 cause of compilation failures here.
• If you run into errors during compilation, double-check that you're on the r1.12 TensorFlow branch and using Bazel 0.15.0.

内容的提问来源于stack exchange，提问作者Rochan