h5py: Fails to build from source on MacOS 11 + M1

Operating System: MacOS 11.2
Python Version: 3.9.1
h5py version: 3.1.0
HDF5 version: 1.12.0

Attempting to install h5py via pip fails for ARM Macs; there is no prebuilt wheel for this platform, so pip attempts to build h5py from source. However, this fails because it attempts to use Numpy 1.19.3 (when on Python 3.9), but Numpy 1.19.3 does not build on M1 Macs out of the box (build log: h5py_build_fail.log).

If you modify setup.py so that h5py builds against Numpy 1.20.1 instead (which works on M1 Macs), h5py builds successfully.

Dumb example patch that fixes the build but isn’t intended to be actually merged:

diff --git a/setup.py b/setup.py
index cc583bc9..9c0400cc 100755
--- a/setup.py
+++ b/setup.py
@@ -29,7 +29,7 @@ NUMPY_MIN_VERSIONS = [
     # Numpy    Python
     ('1.14.5', "=='3.7'"),
     ('1.17.5', "=='3.8'"),
-    ('1.19.3', ">='3.9'"),
+    ('1.20.1', ">='3.9'"),
 ]

 # these are required to use h5py

About this issue

Original URL
State: closed
Created 3 years ago
Reactions: 5
Comments: 17 (7 by maintainers)

Most upvoted comments

I found an alternative solution:

HDF5_DIR=/opt/homebrew/opt/hdf5 pip3 install --no-build-isolation h5py

+15

jiegec on Jun 29, 2021

I have found that pip install only works if you have installed hdf5 separately first on the Mac M1. This was possible through brew but not pip.

Steps:

brew install hdf5
export HDF5_DIR=<path/to/homebrew/Cellar/hdf5/version>
pip install h5py

craigvanden on Apr 6, 2022

@WillTrojak Your fix worked! I also needed to do export HDF5_DIR=/opt/homebrew/Cellar/hdf5/1.12.1/ and then follow the steps from here to build it

RubensGitHub on Aug 27, 2021

Sorry to reopen an old issue, but I had this same problem trying to install h5py and I think I came up with a bit of dirty fix. More than happy for someone to point out a more elegant fix.

The issue for me at least seemed to be with numpy 1.19.3 failing to build. With homebrew numpy 1.20.3 is available, but even with that install the build for h5py will try and build 1.19.3.

By cloning the current master, then bumping the numpy version to 1.20.3 in setup.py, this seemed to get round the issue.

This may have some side effects, but at least it works for now.

WillTrojak on Jun 17, 2021