Generic requirements.txt for TensorFlow on both Apple M1 and other devices

Question

I have a new MacBook with the Apple M1 chipset. To install tensorflow, I follow the instructions here, i.e., installing tensorflow-metal and tensorflow-macos instead of the normal tensorflow package.

While this works fine, it means that I can't run the typical pip install -r requirements.txt as long as we have tensorflow in the requirements.txt. If we instead include tensorflow-macos, it'll lead to problems for non-M1 or even non-macOS users.

Our library must work on all platforms. Is there a generic install command that installs the correct TensorFlow version depending on whether the computer is a M1 Mac or not? So that we can use a single requirements.txt for everyone?

Or if that's not possible, can we pass some flag/option, e.g., pip install -r requirements.txt --m1 to install some variation? What's the simplest and most elegant solution here?

score 1 · Accepted Answer · edited May 05 '22 at 11:15

1

According to this post Is there a way to have a conditional requirements.txt file for my Python application based on platform?

You can use conditionals on your requirements.txt, thus

tensorflow==2.8.0; sys_platform != 'darwin' or platform_machine != 'arm64'
tensorflow-macos==2.8.0; sys_platform == 'darwin' and platform_machine == 'arm64'

edited May 05 '22 at 11:15

CGFoX

4,294
5
39
69

answered Apr 14 '22 at 05:32

calmar

1,812
1
13
15

1

Great, I'll try and accept the answer when it works. Thanks! – CGFoX Apr 14 '22 at 07:05
1

Wait, this checks for the OS, not for the architecture, right? Thus, it fails on Intel MacBooks. – CGFoX May 05 '22 at 10:02
1

You also need to check the architecture with `platform_machine`. I updated your answer. – CGFoX May 05 '22 at 11:16

Generic requirements.txt for TensorFlow on both Apple M1 and other devices

1 Answers1