How to install SciPy on Apple Silicon (ARM / M1)

Question

I have successfully installed python 3.9.1 with Numpy and Matplotlib on a new Mac mini with Apple Silicon. However, I cannot install SciPy : I get compilation errors when using

python3 -m pip install scipy

I also tried installing everything from brew, and import scipy works, but using it gives a seg fault. I have installed ARM versions of lapack and openblas, but this does not fix the problem.

Has anyone succeeded? (I am interested in running it natively, not through Rosetta).

Magnate

Answer 1

It's possible to install on regular arm64 brew python, you need to compile it yourself.

If numpy is already installed (from wheels) you'll need to uninstall it:

pip3 uninstall -y numpy pythran

I had to compile numpy, which requires cython and pybind11:

pip3 install cython pybind11

Then numpy can be compiled:

pip3 install --no-binary :all: --no-use-pep517 numpy

Scipy needs pythran (this should happen after installing numpy):

pip3 install pythran

Then we need to compile scipy itself, it depends on fortran and BLAS/LACK:

brew install openblas gfortran

Tell scipy where it can find this library:

export OPENBLAS=/opt/homebrew/opt/openblas/lib/

Then finally compilescipy:

pip3 install --no-binary :all: --no-use-pep517 scipy

Sebaceous

I had to brew install openblas gfortran before numpy... numpy needed openblas at least – Campstool

Unfortunately the SciPy compilation failed with error: Command "ar rcs build/temp.macosx-10.14.6-arm64-3.8/libsuperlu_src.a build/temp.macosx-10.14.6-arm64-3.8/scipy/sparse/linalg/dsolve/SuperLU/SRC/dlaqgs.o [………]" failed with exit status 1 :( – Telega

This answer worked for me on OSX 11.2.3 (20D91) with a MacBook Pro (13-inch, M1, 2020). I did not get any compilation failed like @Telega did – Macronucleus

This worked for me, but subsequently scikit-learn (pip install sklearn) did not. I bring it up because the notes at scikit-learn.org/stable/… and the github link there suggest (to my noob eyes) that sklearn should install, given a successful install of scipy. Before starting a new question, I thought folks who got scipy working might also have advice on this related issue. – Corinnacorinne

I needed pip install pythran – Twoply

Why should we use the --no-use-pep517 forced legacy mode? – Chlorothiazide

It worked after installing python 3.9, wouldn't work with 3.8 – Prudery

I got this when trying to install scipy: libraries mkl_rt not found in [...] – Educationist

I have a M1 pro Mac (Darwin 21.6.0) and I followed the steps above but got this error - ` File "/opt/homebrew/Caskroom/miniforge/base/lib/python3.10/site-packages/Cython/Distutils/old_build_ext.py", line 157, in getattr return _build_ext.build_ext.__getattr__(self, name) File "/opt/homebrew/Caskroom/miniforge/base/lib/python3.10/site-packages/setuptools/_distutils/cmd.py", line 105, in getattr raise AttributeError(attr) AttributeError: fcompiler. Did you mean: 'compiler'? [end of output]` i would really appreciate any help – Mounts

Answer 2

This one worked for me after wasting hours:

pip install --pre -i https://pypi.anaconda.org/scipy-wheels-nightly/simple scipy

Premonition

Yes.. Indeed. After countless hours. This worked on MacBook Air (M1, 2020) with BigSur Version 11.6 (20G165). If I understand correctly, is this basically from anaconda servers providing most stable version of SciPy? Not sure, how it ended up working well with Appl Silicon? Would highly appreciate your answer. – Tennessee

Best answer, works easy! – Refuge

@DhruvPatel This seems to be the "nightly" version of Scipy, which means it's their latest development version. Apparently they do build for ARM for their upcoming 1.8.0 version (which they didn't do for <1.8.0 which is currently released on PyPI). Since it is a development version I would definitely not qualify this as the most stable version, but fortunately this distribution makes it easier for ARM64 users to use. – Rico

This nightly build method also worked for scikit-learn on my Apple M1, but I needed to add --extra-index to the command: pip install --pre --extra-index https://pypi.anaconda.org/scipy-wheels-nightly/simple scikit-learn. Source: scikit-learn.org/stable/developers/advanced_installation.html – Lovmilla

What for people who don't use anaconda distribution? Is there an alternative? – Hemichordate

pip3 instead of pip worked for me. – Koehler

This works but will build against an OpenBlas that is around 5x slower than Accelerate. – Terrellterrena

This command worked for me beautifully for long time, but now I get errors: Looking in indexes: https://pypi.anaconda.org/scipy-wheels-nightly/simple ERROR: Could not find a version that satisfies the requirement scipy (from versions: none) ERROR: No matching distribution found for scipy – Prasad

@Prasad I'll answer to my own comment since I solved the problem. I was running on an old version of OS and support for it in scipy was removed. Updating the OS solved it. – Prasad

Answer 3

This solution worked on my M1 machine with pyenv:

brew install openblas
OPENBLAS="$(brew --prefix openblas)" pip install numpy scipy

Corse

This worked for me. I was also able to install scikit-learn using the same approach: OPENBLAS="$(brew --prefix openblas)" pip install numpy scikit-learn – Anderson

Answer 4

You can install miniforge from https://github.com/conda-forge/miniforge#miniforge3 and then install those packages with,

conda install numpy scipy matplotlib

Polypus

I did. This produces a segfault whenever I import a submodule of scipy. Numpy and matplotliib are fine, though. – Tearful

You can also install miniforge using brew: brew install --cask miniforge – Laughing

Thank you! Installation worked for me, however I got stuck with importing the installed packages in virtual environments, where my local packages were ignored. The solution is documented here: #66768759 – Telega

Thank you. I think this is the best solution for someone that uses tensorflow and miniforge3 on MacBook m1 – Hostetler

Answer 5

For me the easiest solutions:

brew install scipy

Probably good idea to edit the PATH, so the homebrew version will be the default.

Torrlow

Note that if you do this, brew installs python3.9 at this location /opt/homebrew/Cellar/[email protected]/3.9.6/bin/python3, and you will only be able to import scipy if you run this python executable – Heather

Do NOT install a Python package with brew. Use an env to isolate things. This is a recipe for disaster. See my answer for how it should be done for SciPy. – Lomond

seconding @Lomond on this – Brillatsavarin

Answer 6

I managed to get scipy installed on Apple Silicon. I mostly followed the instructions by lutzroeder here: https://github.com/scipy/scipy/issues/13409

Those instructions weren't successful for me, but running 'pip3 install scipy' worked afterwards. I think this fixed the problem for me:

/opt/homebrew/bin/brew install openblas

export OPENBLAS=$(/opt/homebrew/bin/brew --prefix openblas)

export CFLAGS="-falign-functions=8 ${CFLAGS}"

Decurved

worked for me on m1 macOS 13 – Lidalidah

Answer 7

For those who need it for short-term purposes and don't want too much hustle - it seems to work with python 3.6.4 and scipy 1.5.4 out of the box (Big Sur 11.5.2, M1 chip).

Hist

This does not provide an answer to the question. Once you have sufficient reputation you will be able to comment on any post; instead, provide answers that don't require clarification from the asker. - From Review – Rhinencephalon

@Rhinencephalon the implied answer is "use/specify an earlier version of Python and/or Scipy, such that it will work out of box". In the Python ecosystem, this is actually a sane (if not desirable) answer, because of how packaging works: software distributors offer pre-built "wheels" for environments where they can, and expect users to build locally otherwise. This is trivial for pure-Python projects, but can be nightmarish for code that relies on extensions written in other languages. Sometimes, there aren't wheels yet for newer versions, for every platform. – Healion

Answer 8

What version of scipy you're trying to install?

To me running on Macbook air M1, I needed to increase the version from scipy==1.5.1 to scipy==1.7.3, so I guess you should use 1.7.3 version or above it and all will be fine...

 pip install -Iv scipy==1.7.3

Or just add in your file requirements.txt this line:

scipy==1.7.3

Seamstress

Answer 9

SciPy maintainer here.

SciPy is now distributed for M1, on both conda and pip. You need at least macOS 12.0 (Monterey) and Python >= 3.8. Our support for macOS is good, meaning you should not need to do anything special or compile yourself unless you want to use a development version or older macOS versions.

If you need to compile, only use our guides (e.g. https://scipy.github.io/devdocs/dev/contributor/building.html#building-from-sources). There are tons of moving pieces and doing something else might work today, but not tomorrow.

Avoid using pip to compile SciPy. Use our developer script python dev.py build after having followed the guides I linked to setup system dependencies for your platform.

Do NOT use brew to install a Python package, use an environment. brew is great to install system dependencies or other dev tools, but should not be used to install Python packages. Down the road this is a recipe for trouble.

Lomond

from which version it's supported? I was able to install 1.7.3 without any extra work. – Hildebrandt

I don’t recall. Just don’t play the wizard with trying to compile yourself, my recommendations are still valid today ;) – Lomond

Answer 10

In addition, if someone has this error message>

########### CLIB COMPILER OPTIMIZATION ###########
Platform      :
  Architecture: aarch64
  Compiler    : clang

CPU baseline  :
  Requested   : 'min'
  Enabled     : none
  Flags       : none
  Extra checks: none

CPU dispatch  :
  Requested   : 'max -xop -fma4'
  Enabled     : none
  Generated   : none
CCompilerOpt.cache_flush[809] : write cache to path 

I found this solution before compile numpy and scipy

Analysis of reasons: From the above error message, you can see that the last error shows that clang has an error, so it is speculated that it should be an error caused by the compiler, because the new version of the xcode command tool uses the arm version of the compilation method by default, and if we want to use For x86 architecture, we need to manually set the specific architecture through environment variables.

export ARCHFLAGS="-arch x86_64"

example:

3c790c45799ec8c598753ebb22/build/temp.macosx-10.14.6-arm64-3.8/ccompiler_opt_cache_clib.py
    ----------------------------------------
ERROR: Command errored out with exit status 1: /Users/daniel_edu/Projects/PERSONAL/great_expectation_demo/.env/bin/python3 -u -c 'import io, os, sys, setuptools, tokenize; sys.argv[0] = '"'"'/private/var/folders/zb/c_b9kh2x1px7vl5683rwz8fr0000gn/T/pip-install-y8alaej_/numpy_3d813a3c790c45799ec8c598753ebb22/setup.py'"'"'; __file__='"'"'/private/var/folders/zb/c_b9kh2x1px7vl5683rwz8fr0000gn/T/pip-install-y8alaej_/numpy_3d813a3c790c45799ec8c598753ebb22/setup.py'"'"';f = getattr(tokenize, '"'"'open'"'"', open)(__file__) if os.path.exists(__file__) else io.StringIO('"'"'from setuptools import setup; setup()'"'"');code = f.read().replace('"'"'\r\n'"'"', '"'"'\n'"'"');f.close();exec(compile(code, __file__, '"'"'exec'"'"'))' install --record /private/var/folders/zb/c_b9kh2x1px7vl5683rwz8fr0000gn/T/pip-record-q9vraevr/install-record.txt --single-version-externally-managed --compile --install-headers /Users/daniel_edu/Projects/PERSONAL/great_expectation_demo/.env/include/site/python3.8/numpy Check the logs for full command output.
(.env) ➜  great_expectation_demo git:(master) ✗ export ARCHFLAGS="-arch x86_64"
(.env) ➜  great_expectation_demo git:(master) ✗ pip install --no-binary :all: --no-use-pep517 numpy
Collecting numpy
  Using cached numpy-1.21.5.zip (10.7 MB)
  Preparing metadata (setup.py) ... done
Skipping wheel build for numpy, due to binaries being disabled for it.
Installing collected packages: numpy
    Running setup.py install for numpy ... done
Successfully installed numpy-1.21.5

Quote

Answer 11

The following worked for me.

I'm currently using Python 3.10.8, installed using brew.

And currently, when installing numpy==1.23.4, setuptools < 60.0.0 is required.

I'm using (brew --prefix)/bin/python3 -m pip for explicitly calling the pip from python 3.10 installed by brew.

Here are the versions I've just installed.

# python         3.10.8

# pip            22.3
# setuptools     59.8.0
# wheel          0.37.1

# numpy          1.23.4
# scipy          1.9.3
# pandas         1.5.1
# scikit-learn   1.1.3
# seaborn        0.12.1
# statsmodels    0.13.2

# gcc            12.2.0
# openblas       0.3.21
# gfortran       12
# pybind11       2.10.0
# Cython         0.29.32
# pythran        0.12.0

Here are the steps I followed:

# setuptools < 60.0.0 is required for numpy==1.23.4 in Python 3.10.8
$(brew --prefix)/bin/python3 -m pip install --upgrade pip==22.3 setuptools==59.8.0 wheel==0.37.1

# uninstall numpy and pythran first
$(brew --prefix)/bin/python3 -m pip uninstall -y numpy pythran

# uninstall scipy
$(brew --prefix)/bin/python3 -m pip uninstall -y scipy

# install prerequisites (with brew)
brew install gcc
brew install openblas
brew install gfortran

# set environment variables for compilers to find openblas
export LDFLAGS="-L/opt/homebrew/opt/openblas/lib"
export CPPFLAGS="-I/opt/homebrew/opt/openblas/include"

# install the prerequisites (with pip)
$(brew --prefix)/bin/python3 -m pip install pybind11
$(brew --prefix)/bin/python3 -m pip install Cython

# install numpy
$(brew --prefix)/bin/python3 -m pip install --no-binary :all: numpy

# install pythran after installing numpy, before installing scipy
$(brew --prefix)/bin/python3 -m pip install pythran

# install scipy
export OPENBLAS="$(brew --prefix)/opt/openblas/lib/"
$(brew --prefix)/bin/python3 -m pip install scipy

# install pandas
$(brew --prefix)/bin/python3 -m pip install pandas

# install scikit-learn
$(brew --prefix)/bin/python3 -m pip install scikit-learn

# install seaborn
$(brew --prefix)/bin/python3 -m pip install seaborn

# install statsmodels
$(brew --prefix)/bin/python3 -m pip install statsmodels

Gessner

Answer 12

According to this Github issue, Scipy doesn't work on MacOS 11 (Big Sur). If none of these solutions are working for you I'd suggest updating your OS.

Susian

Answer 13

I use conda install scipy to resolve this problem. Conda have a custom version of scipy for Apple M1. Update macOS to 12 if you don't want to use Conda.

Cowherb