How to split workload on Ray QA pipeline?

[Janus150]

I'm trying to run a distributed question-answering pipeline using Ray on Mac. I have set the necessary environment variable to run a cluster, and it executes. I have the yaml configured with multiple replicas for the reader, and I can see in the logs that it deploys them onto multiple machines. When I run the pipeline, it will run one of the replicas, which changes from one run to the next. However, it isn't splitting the workload; all the retrieved documents are sent to one reader.

Is there an argument I'd need to pass to the run method or something additional I'd need to set in the yaml to distribute it?

[silvanocerza]

You should probably take a look at the ray.init() official docs to know about their arguments.

When you're using RayPipeline the ray_args argument is passed to ray.init(), see this line here.
If there is an actual argument that lets you split the workload you should be able to set it like that.