Docker memory demands for running demo locally are approaching 16GB about allennlp-demo HOT 4 CLOSED

@schmmd @DeNeutoy FYI.

I should also mention that I'm not 100% sure the errors I was getting today were memory-related, but they went away when I increased memory allocation.

Either way, it seems that on our current trajectory, we will eventually reach a point where machines equipped with 16GB will be unable to run the demo locally via docker.

from allennlp-demo.

@aaronsarnat if you need a short term fix for this, you can modify this line here to only load the relevant model. E.g when I was developing the dependency parser demo, I would add:

    for name, demo_model in MODELS.items():
		if name != "dependency-parser":
			continue
        logger.info(f"loading {name} model")
        predictor = demo_model.predictor()
        app.predictors[name] = predictor

This causes only the relevant model to be loaded, meaning that 1) it's faster and 2) it takes way less memory.

You can see the keys for the models here.

from allennlp-demo.

@DeNeutoy thanks for the follow-up. I will definitely try this if I run into any memory blockers.

I suppose another solution is we each get one of these:

from allennlp-demo.

We added better support for running a subset of the models.

from allennlp-demo.