Serve models for parallel benchmarking

[juanmc2005]

Problem

The amount of parallel pipelines that can run in Benchmark is limited because the models need to be copied in each process.

Idea

Serve models in a separate process (e.g. in a Docker container) and make Parallelize send batches to them as needed. This could also have the advantage of running pipelines on low-resource devices by using remotely deployed segmentation and embedding models.

Example

diart.benchmark /wav/dir --reference /rttm/dir --segmentation 192.168.0.20:7007 --num-workers 16

and even:

diart.stream microphone --segmentation 192.168.0.20:7007 --embedding 192.168.0.20:7008