adding environment variable for transformers in models #702
Conversation
Documentation preview |
| @@ -77,3 +77,5 @@ RUN if [ "$INSTALL_DISTRIBUTED_EMBEDDINGS" == "true" ]; then \ | |||
| cd /distributed_embeddings && git checkout ${TFDE_VER} && git submodule update --init --recursive && \ | |||
| make pip_pkg && pip install artifacts/*.whl && make clean; \ | |||
| fi | |||
|
|
|||
| ENV FORCE_TF_AVAILABLE=true | |||
There was a problem hiding this comment.
maybe worth adding a comment here providing some background on this - That the env var is for the transformers package to look up. And required because importlib.metadata.version("tensorflow") doesn't work (raises PackageNotFoundError) for our custom version of tensorflow that ships in the container
There was a problem hiding this comment.
Now that I mentioned this, I realised that checking if this also applies in the image where we're getting tensorflow from might be worthwhile. (nvcr.io/nvidia/tensorflow:22.08-tf2-py3). The metadata version import works there. And this is because the dist-info directory for tensorflow is present in the dist-packages directory there.
An alternative (or additional change to this env var) could be to copy over the dist-info directory along with the package perhaps
There was a problem hiding this comment.
Opened a PR here to add dist-info #704
No description provided.