Closed seanpmorgan closed 9 months ago
We have the same issue when I've introduced this in the release PR. It seems that pytest is not installed in the same python env where we are executing tests with bazel.
Fixes #2849
@angerson we (and other SIGs / downstream consumers) have had issues with the the hermetic python bazel rule. Can you or someone in SIG build take a look at this failing test when time allows:
install_deps
call does not properly install the 3 requirements files that we have specified, since we error on ModuleNotFoundError: No module named 'pytest'
py_binary()
for configure.py
using the hermetic python after the install deps are available instead of using the system python install as is down in the test above. We tried a few things on the last release... but we want to get ahead of this for upcoming releases when the new workspace strategy may be required (I see that we're still publishing multiple py docker images for future tf releases though).
@seanpmorgan Why do you pull tensorflow via http_archive and also via pip? You started doing it not so long ago - https://github.com/tensorflow/addons/commit/1c3a2b2b0c5ca902e056427ec943e326cf6b70c1 Looking at the code, you should install it only via pip.
@seanpmorgan Why do you pull tensorflow via http_archive and also via pip? You started doing it not so long ago - 1c3a2b2 Looking at the code, you should install it only via pip.
https://github.com/tensorflow/addons/blob/master/WORKSPACE#L18-L32
Because we load the 4 bazel workspace files that are provided since we compile custom CPU and CUDA ops. IIRC when on TF2.9 there was rev in CUDA version and our previous local toolchain broke. We started loading the workspace files from tensorflow to build the same way.
@seanpmorgan Why do you pull tensorflow via http_archive and also via pip? You started doing it not so long ago - 1c3a2b2 Looking at the code, you should install it only via pip.
https://github.com/tensorflow/addons/blob/master/WORKSPACE#L18-L32
Because we load the 4 bazel workspace files that are provided since we compile custom CPU and CUDA ops. IIRC when on TF2.9 there was rev in CUDA version and our previous local toolchain broke. We started loading the workspace files from tensorflow to build the same way.
There are 3 ways to fix it:
@georgiyekkert Can you send a PR on this repo directly or over this PR's source branch? Thanks
Description