Apple Silicon M1 Macbook air for 01_intro

Part 1 (2020)
1

Hello

I am running Mac OS X Big Sur on Apple macbook air m1
with python 3.8 (the python provided by Apple on BigSur)

learn.fine_tune(1)

can not be done with Apple Silicon on 01_intro jupyter notebook

i have this log message

[W NNPACK.cpp:80] Could not initialize NNPACK! Reason: Unsupported hardware.
[W ParallelNative.cpp:206] Warning: Cannot set number of intraop threads after parallel work has started or after set_num_threads call when using native parallel backend (function set_num_threads)

the only solution i found is to

export OMP_NUM_THREADS=1

just before starting

jupyter notebook

but this is not a good solution because the CPU is only < 100%

Capture d’écran 2020-12-27 à 21.54.49
Capture d’écran 2020-12-27 à 21.54.491058×784 162 KB

I hope fastai will upgrade it’s code as tensorflow did :wink:

thank you

3

the same monitoring with an intel processor

Capture d’écran 2020-12-27 à 22.07.35
Capture d’écran 2020-12-27 à 22.07.351092×589 131 KB

4

fastai cannot do anything regarding M1 Macbook Air support. PyTorch will have to include support:

4 Likes
5

how are you able to run fastai / fastbook without pytorch being compatible with apple silicon?

6

@vincentwi : I managed to install pytorch without any problem

with pip

python3 -m venv ./env

source env/bin/activate

pip install --upgrade pip

pip install torch torchvision
7

@ilovescience : it is actually running on my M1 Macbook Air … slowly, without any GPU support, but it runs

8

Yeah it’s just running on the CPU. You aren’t going to get the speedups you see with TensorFlow because TensorFlow has GPU support for M1 Mac. If you were to run both PyTorch/fastai and TensorFlow on just CPU, they would likely be comparable.

1 Like
9

@ilovescience : this is not the reason why i wrote this post.
On my Macbook Pro, INTEL version, the fast.ai script runs well without any GPU support, with 6 processes.
On my Macbook Air, M1 version, without Rosetta, the same script can not be run with 2 processes or more, I had to force the export OMP_NUM_THREADS=1

10

Again this is likely some optimization that PyTorch hasn’t supported. Have you tried running some experiments with pure PyTorch to compare?

11

ok, i will give it a try

1 Like
12

hi @ilovescience

I just ran https://pytorch.org/tutorials/beginner/blitz/cifar10_tutorial.html#sphx-glr-beginner-blitz-cifar10-tutorial-py
where some scripts use 2 workers … and i have to say that everything went well
Core 5 and Core 6 were able to train the classifier multithreads

Capture d’écran 2020-12-29 à 19.43.43
Capture d’écran 2020-12-29 à 19.43.431034×630 63.1 KB

13

if you have pytorch are you then able to install the entire fastbook package? Meanining, can you fully do the fast.ai bootcamp on Apple Silicon (M1)?

14

@vincentwi

i was able to import the whole fastbook package, but in order to run and test “part 1” of the book,

until the part where you can upload a picture, and guess if this is a cat or not

before starting “jupyter notebook”, I had to force the env variable
OMP_NUM_THREADS=1
the “learning” part of the python script runs, very very slowly.

In result, i could train the model. yes.
With the same script, my “intel” macbookpro is fastest, by far.

I did not try the remaining code of the script

15

I was just able to share the whole model from lesson 1 on my local MBP M1 Max with 24 GPU Cores.

2 Likes