Master node not working when using distributed queue

[forxyum]

I'm running a 4 GPU cluster, utilizing this node package for load balancing and ease of use, so I don't need to manage 4 instances myself.
When using the Distributed Queue node, the master node never assigns a job to itself, even though it's NOT set to orchestrator-only.
The workaround I've found for the time being is running a 5th instance for cuda0, and setting the master to orchestrator-only, but since both of them are on cuda0, if the master fails to assign the job to a worker and falls back to local execution, I might get into a situation where both instances are trying to work on cuda0 and OOM out.

[robertvoy]

There is a misunderstanding here.

The distributed queue node serves a different purpose. You want to be using the distributed collector. See one of the example workflows for how to use it.

[forxyum]

In a 4 GPU setup, wouldn't that just run the same workflow on all 4 GPUs and collect the results on the master? Basically making a batch size 4 result just 4x faster?
That's not what I'm trying to achieve.
I want to run batch size 1 jobs with different inputs on 1 GPU at a time. The Distributed Queue node allows me to just keep sending jobs to the master instance, and get it distributed between the 4 GPUs.
As of now, even if I have 15 jobs to run, I can just send all of them to the master instance with different inputs, and it will get distributed to one of the available GPUs, the only problem being, the master instance never assigns a job to itself, just the workers, even when not in orchestrator mode.
Isn't this the purpose of the node? If not, could you explain what is?

[robertvoy]

Got it. You are using it correctly then. I’ll be making that change to allow for including the master.