In contrast, the environment variable PIP_CONSTRAINT is visible at the top level and within the isolated build environments, so the numpy constraints are set correctly in the isolated builds for spaCy and its dependencies. The -c constraints.txt setting is only applied to the packages specified at the top level with pip install package, not within the isolated build environments. Note that setting PIP_CONSTRAINT IS NOT equivalent to setting pip install -c constraints.txt. It will take 7-8 minutes to compile and install all the required packages. PIP_CONSTRAINT=constraints.txt OPENBLAS= " $(brew -prefix openblas ) " pip install spacy For conda users who want to install spaCy once without compiling.No compiling required using the experimental miniforge OS X ARM64 conda installer.Option 3: Install binary packages from conda-forge Primarily for developers who plan to recompile spaCy frequently.Option 2: Install with pip without build isolation For users who want to install spaCy once.Install spaCy and its dependencies with pip with the fewest steps.Option 1: Install with pip with build isolation This guide describes three options for installing numpy and spaCy that have been tested as of March 2021: The good news is that it looks like this will be fixed in numpy 1.21.)
(For reference, you need wheel>=0.36.1 for Big Sur. Since numpy and spaCy are configured to use build isolation by default ( PEP 517), a simple pip install spacy does not work: the numpy install fails due to the version of wheel and then the spaCy install fails due to the failed numpy installation. Installing numpy or packages that compile against numpy on an Apple M1 is not straight-forward because the most recent version of numpy (1.20) requires a version of wheel in pyproject.toml ( PEP 518) that doesn't work on Big Sur. Installing numpy 1.20 or earlier and spaCy on an Apple M1 If you run into problems, please search for related questions in the Installation section of the new discussion board, and start a new discussion thread if needed. Note that this situation should improve after the release of numpy 1.21. This guide is cross-posted and updated slightly from the original post for better visibility for spaCy users. Alternatively, you can continue to use the experimental conda-forge packages as