Apple M1 chip - illegal hardware instruction

[SaschaHeyer]

System information

• OS Platform and Distribution (e.g., Linux Ubuntu 16.04): MacOS Big Sur 11.0.1
• TensorFlow installed from (source or binary): https://www.tensorflow.org/install/pip
• TensorFlow version: Latest stable (pip)
• Python version: Python 3.8.5
• Installed using virtualenv? pip? conda?: exactly like in the instruction https://www.tensorflow.org/install/pip
• Bazel version (if compiling from source): /
• GCC/Compiler version (if compiling from source): /
• CUDA/cuDNN version: /
• GPU model and memory: /

Describe the problem

After running the verification steps

python -c "import tensorflow as tf;print(tf.reduce_sum(tf.random.normal([1000, 1000])))"

The following error appear

illegal hardware instruction

Any other info / logs
This issue only appear on the Mac with the Apple M1 chip. The same setup procedure https://www.tensorflow.org/install/pip works fine on my other Mac`s.

[mihaimaruseac]

We don't support the M1 chip. This comes from an Apple fork.

[SaschaHeyer]

Hello Mihai,

thanks for your quick response.
This information rather surprises me especially since Apple and TensorFlow did quite some advertising about it.

• This means we will have to work on some kind of private Apple fork rather then on the official TensorFlow version?
• Are there any plans to integrate this directly into TensorFlow?
• This also means we are bounded to TensorFlow 2.4 on Apple M1 Macs?

It seems somehow odd to use a fork just because it is another chip.

[mihaimaruseac]

The new chip is on ARM, afaik. We only release x86_64 official pips and leave everything else to the community.

there is #45404 to make this repo build on the ARM chips but will take a while until it lands and we still don't have this supported during release time. We have less than 1 engineer left for releasing.

[SaschaHeyer]

In the near future, we’ll be making updates like this even easier for users 
to get these performance numbers by integrating the forked version into
the TensorFlow master branch.

https://blog.tensorflow.org/2020/11/accelerating-tensorflow-performance-on-mac.html

Can you provide us an ETA for that?

Great to see the community is working on that #45404. But it usually takes a lot of time till builds are implemented and integrated. That means I have to rely on Apple and their support. This is a big issue in my opinion.

[mihaimaruseac]

Afaik Apple has not communicated at all with the team that does the releases. Given that there is only 1 person (actually less since releases are not prioritized), there cannot be an ETA.

[Young1783]

#46044 (comment)
I hope this comment will help you before merge the Apple fork, which seems it will be a longgggg time.
:)

[ben-xD]

Shouldn't it work when using Rosetta 2 though? Unfortunately it doesn't when running arch -x86_64 python -c "import tensorflow as tf;print(tf.reduce_sum(tf.random.normal([1000, 1000])))"

Edit: Looks like this goes over how to fix it: Compile TensorFlow from source using Bazel on Rosetta 2 #46044 (comment)

[mihaimaruseac]

For Rosetta2: see #46044 (master branch should be able to compile, 2.4 branch waiting for a patch release)
For M1 chip, we landed #45404 so now if you compile this repo from source you should also get a working TF.

We are currently looking at how we can provide ongoing support.

I think we can close this now and leave #46044 open.

[OneCricketeer]

The blog post specifically links to a separate tensorflow repo that needs to be built, not the version that comes from PyPi

https://blog.tensorflow.org/2020/11/accelerating-tensorflow-performance-on-mac.html

[ymodak]

Closing this issue in the light of #45404. Thanks!

[google-ml-butler]

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

[nijatmursali]

Is there any update on this issue? Now it is June 12th and I'm still getting this error.

[Genius3435]

I wanted to add that I get this not just with tensorflow, but with gcc/g++ (more often if I have some optimization/target pragmas), and with lots of things that use gcc/g++, like a rust crate I just tried to run.

[shanu12joshi]

Hey, It's July I am still facing the issue, Is there a solution yet?

[msk-psp]

You guys must see this post
https://developer.apple.com/metal/tensorflow-plugin/

this post describes how to import TensorFlow with there plugins

[YeungYathin]

I am M1 Pro user. I met this problem after activating the virtual environment Anaconda and installing tensorflow. When I run my program and it failed. I'm perplexed about it and hope that Tim Cook can help me.