Standalone deployment of individual modules

The ASKCOS backend consists of the API gateway and the individual prediction services. Each of these prediction services has its own repository. When the prediction service does not depend on other services, it is possible to start and use the service by itself, without starting the API gateway. Simply follow the README of the module repo to serve the module as a standalone service, which will be much more lightweight. In addition, some of these modules have experimental features such as singularity and/or GPU-based serving, which could be useful for certain scenarios (e.g., when deploying without root access).

Currently, the modules that can be served standalone are

• atom_map/indigo
• atom_map/rxnmapper
• atom_map/wln
• context_recommender
• fast_filter
• forward_predictor/augmented_transformer
• forward_predictor/graph2smiles
• forward_predictor/wldn5
• reaction_classification
• retro/augmented_transformer
• retro/graph2smiles
• retro/template_relevance
• scscore
• site_selectivity

Modules not listed here generally have dependencies. For example, the tree_search/mcts service is dependent on several other modules and most notably, on the retro/* modules. The dependency calls are done via API calls which are centrally managed by the API gateway. These modules therefore require the API gateway to be started and cannot be served in a standalone manner.