Apple M1 - How to build 2.6 or Master (Fixed)

[0xA1B2]

The genius command "pip install tensorflow" does not even work in Debian 11, so we have to compile from source. The instructions on the TF homepage do not work for the M1.

Has anyone compiled TF 2.6+ on the M1 and can share a guideline please ?

[ymodak]

This issue better raised on https://github.com/apple/tensorflow_macos/issues
Please refer a previous thread which discusses similar issue. Thanks!

[0xA1B2]

Current Progress on MacOS 12 Beta 5 & default Python 3.9.6

Install from Developer site:
Xcode Command Line Tools 13.x
Xcode Gui 13.x # reboot

conda deactivate #exit Conda environment

pip3 install --upgrade pip

brew install bazel
nano .bazelversion # change to 4.2.0
cat .bazelversion # == 4.2.0
bazel --version # == 4.2.0

git clone https://github.com/tensorflow/tensorflow.git
cd tensorflow
#git checkout master
#git pull --rebase

./configure

bazel build --config=macos_arm64 tensorflow/tools/pip_package:build_pip_package

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

Result in:
/System/Volumes/Data/private/tmp/tensorflow_pkg/

tensorflow-2.7.0-cp39-cp39-macosx_12_0_arm64.whl

pip install *.whl

---

To get GPU support, we need to install the new Pluggable Device:
https://pypi.org/project/tensorflow-metal/#files

python3 -m pip install tensorflow-metal # 0.1.2

Those wheels are for TF 2.5, it works with TF 2.7 but slower. Apple needs to regularly update tensorflow-metal to match new TF main versions, and also release sources.

[ymodak]

Thanks for sharing your analysis. I will close this issue now.

[asapsmc]

Any way to make this work for a M1 inside a Conda environment?

[ymodak]

@MR-T77 TF build install issues in Conda environment are supported on https://github.com/ContinuumIO/anaconda-issues/issues
Thanks!

[asapsmc]

@ymodak sorry, I tried here because I can't find any way to have tensorflow working in my M1.

[google-ml-butler]

Are you satisfied with the resolution of your issue?
Yes
No

[MuhTawfik]

I was having same problem what was missing is after installing miniforge you have to change the directory to it as below

cd /Users/<your_username>/miniforgeX. "don't forget to change<your_username> with your user name and to change minifrogX with version like miniforge3
example: cd /Users/mohamed/miniforge3

• after that create env
  conda create -n data-science python=3.9
• activate
  conda activate data-science
• add apple chnnel
  conda config --prepend channels apple
• install dependency
  conda install -c apple tensorflow-deps
• then normal install
  pip install tensorflow-macos
  pip install tensorflow-metal

hope that help

[BenjaminDEMAILLE]

Current Progress on MacOS 12 Beta 5 & default Python 3.9.6

Install from Developer site: Xcode Command Line Tools 13.x Xcode Gui 13.x # reboot

conda deactivate #exit Conda environment

pip3 install --upgrade pip

brew install bazel nano .bazelversion # change to 4.2.0 cat .bazelversion # == 4.2.0 bazel --version # == 4.2.0

git clone https://github.com/tensorflow/tensorflow.git cd tensorflow #git checkout master #git pull --rebase

./configure

bazel build --config=macos_arm64 tensorflow/tools/pip_package:build_pip_package

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

Result in: /System/Volumes/Data/private/tmp/tensorflow_pkg/

tensorflow-2.7.0-cp39-cp39-macosx_12_0_arm64.whl

pip install *.whl

To get GPU support, we need to install the new Pluggable Device: https://pypi.org/project/tensorflow-metal/#files

python3 -m pip install tensorflow-metal # 0.1.2

Those wheels are for TF 2.5, it works with TF 2.7 but slower. Apple needs to regularly update tensorflow-metal to match new TF main versions, and also release sources.

Don't work for me. macOS 12.6

[thealy]

Above procedure died after compiling over 15K files:
ERROR: /Users/thealy/Downloads/tensorflow/tensorflow/python/BUILD:3250:24: Linking tensorflow/python/_pywrap_tensorflow_internal.so failed: (Aborted): cc_wrapper.sh failed: error executing command external/local_config_cc/cc_wrapper.sh @bazel-out/darwin_arm64-opt/bin/tensorflow/python/_pywrap_tensorflow_internal.so-2.params

ld: malformed trie, node past end file 'bazel-out/darwin_arm64-opt/bin/_solib_darwin_arm64/libtensorflow_Slibtensorflow_Ucc.2.12.0.dylib'
clang: error: linker command failed with exit code 1 (use -v to see invocation)
Error in child process '/usr/bin/xcrun'. 1
external/local_config_cc/cc_wrapper.sh: line 69: 8045 Abort trap: 6 "$(/usr/bin/dirname "$0")"/wrapped_clang "$@"

Target //tensorflow/tools/pip_package:build_pip_package failed to build
Use --verbose_failures to see the command lines of failed build steps.

INFO: Elapsed time: 2578.594s, Critical Path: 245.43s
INFO: 18477 processes: 4603 internal, 13874 local.
FAILED: Build did NOT complete successfully

[dorothykiz1]

Running this bazel build --config=macos_arm64 tensorflow/tools/pip_package:build_pip_package compiles for a very long time.

[karllessard]

I'm also hitting the malformed trie, node past end file error.

This issue should not have been closed in the first place, we are still not able to build sources on an Apple Silicon, which is the main topic here. The tensorflow_macos binaries distributed by Apple only helps users using TensorFlow in Python, it is not a viable solution for those using other language bindings, such as Java, C++ or Go. We need at least to be able to compile it from sources, which does not seem to be possible at the moment.

[pooya-jannaty]

Having a similar malformed trie issue both on an M1 mac and Intel-based.

[karllessard]

Downgrading Xcode CL Tools to 13.x did the trick for me, probably related to this: #58368

[Yanyan-Wang]

@0xA1B2, Thanks! With many additional tips, it works for me on Mac Big Sur 11 with M1 and python 3.

I am a fresher of tensorflow and python and I came across several problems during the process, hope it helps buddies who have the same problems:

1. Version of XCode should match your system, e.g. my Macos 11 does not support XCode 13 and higher versions.

2. When configuring with ./configure , ROCm and Cuda should be both disabled. I have tried ROCm enabled and Cuda disabled (since I have no cuda on my mac), the next build command results in:

Error in fail: ERROR: Building with --config=rocm but TensorFlow is not configured to build with GPU support. Please re-run ./configure and enter 'Y' at the prompt to build with GPU support

3. After installing the correct version of XCode, The building command bazel build --config=macos_arm64 tensorflow/tools/pip_package:build_pip_package resulted in:

Xcode version must be specified to use an Apple CROSSTOOL

And I found the solution here as:

bazel clean --expunge 
sudo xcode-select -s /Applications/Xcode.app/Contents/Developer
sudo xcodebuild -license
bazel clean --expunge 
bazel build --config=opt //tensorflow/tools/pip_package:build_pip_package

4. The building command takes about one hour to finish………… be patient and do not terminate it

5. After suffering so many difficulties, the last step pip install *.whl(before it, remember cd the directory /System/Volumes/Data/private/tmp/tensorflow_pkg/ where your .whl file at) came across:

ERROR: tensorflow-2.15.0-cp310-cp310-macosx_11_0_arm64.whl is not a supported wheel on this platform.

It is caused by wrong python version in my case. Checkout the file name of .whl: tensorflow-2.15.0-cp310-cp310-macosx_11_0_arm64, which means it should be ran with python 3.10 (cp310 in file name) and maxos 11 (maxcos_11) and arm64 architecture (arm64, which matches the M1 Mac).

And the command pip install *.whl will run with default python version in your computer, which is python 2 in mine. I also have python 3, to specify the python version, I ran python3 -m pip install *.whl instead. But it also failed, I found that python3 --version in my terminal resulted in 3.11 which does not match cp10. So I ran python3.10 -m pip install *.whl instead.

6. Then, python3.10 -m pip install *.whl resulted in:

ERROR: No matching distribution found for tensorflow-io-gcs-filesystem>=0.23.1

I followed the answer here as:

• clone the tensorflow/io repository
• cd into it's directory
• run: python3 setup.py -q bdist_wheel
• and go back to /System/Volumes/Data/private/tmp/tensorflow_pkg/ (where your .whl file at)
• run python3.10 -m pip install *.whl again

Finally works, it takes whole day!!