IssueTracker

What you would like to accomplish:

It would be desired to be able to define environment variables during the configuration of a model version so they can be available before the code import, an analogy of this behavior can be found in Cloud Functions [1], where Cloud Functions allows you to specify arbitrary key/value pairs at the time of deployment which will be surfaced as literal environment variables accessible by your code at runtime. Environment variables are stored in the Cloud Functions backend, are bound to a single function, and adopt the same lifecycle as the function to which they are bound. Currently, the TF_CONFIG environment variable [2] is used to provide configuration strings.

[1] https://cloud.google.com/functions/docs/env-var

[2] https://www.tensorflow.org/guide/distributed_training#setting_up_tf_config_environment_variable