Dependency for numpy 1.17 #977
Conversation
|
Hi, @xaleryb, is this only for testing purposes? What is the rationale for this? I think we should not restrict numpy version for SDC at all, since by itself it doesn't need specific numpy. |
|
hi @kozlov-alexey this is not for testing only we plan to merge and include to 2021.3 release. This is not a restriction because we do not set run dependency for numpy 1.17 (this is inly for building) and we expect sdc build (based on numpy 1.17) working with latest numpy which will be a part of 2021.3. This is for compatibility on AIKIt level |
Then this has to merged to development branch first, as per our agreements. But:
this has impact not only on AIKit but anyone who will build SDC with conda. There should probably be other way to restrict numpy version that is used when building SDC for AIKit, shouldn't it? |
There was a problem hiding this comment.
@tomashek, @xaleryb I would prefer not to submit this as there's no requirement for SDC package itself to be built with specific numpy version. Also there are no description in the commit that explains why this change is needed and when it can be reverted (if it's some kind of workaround, at least some GitHub reference is needed). Also there should generally be more proper ways of fixing numpy version during specific conda builds (e.g. CONDA_NPY env variable may be set).
| @@ -25,7 +25,8 @@ requirements: | |||
| - python | |||
| - setuptools | |||
| - numba {{ NUMBA_VERSION }} | |||
| - numpy | |||
| - numpy # [win or osx or py==38] | |||
There was a problem hiding this comment.
Why numba build version is restricted only on linux and python=3.7?
How it's going to work on win or linux+python=3.8 (i.e. will the core issue remain unsolved there)?
There was a problem hiding this comment.
The initial problem we are trying to resolve by this change is applicable only for linux.
Please add description with motivation for this change and add the reference to related Github/Jira issues. |
@tomashek could you please share this information if any. |
@kozlov-alexey , I do not think there is an Issue in this github to reference. |
This PR is to follow Anaconda BKM to build using the oldest numpy with a stable API, which is 1.16. Intel's oldest numpy supporting python 3.7 is numpy 1.17, so we use that only at build time so ensure broadest compatibility with newer numpy versions.
Relates to tensorflow/tensorflow#47691 (TensorFlow does not support numpy 1.20.x)