Separate BSO as a server

[show981111]

Pull BSO out as a server and implement the client on the training side.

Components

• BSO server
• BSO client

Currently, working on simple unit tests.

TODO:

• [x] Integrate examples/trace_driven/run_single.py
• [x] Add delete /job for servers
• [x] ORM and connect DB
• [x] Dockerfile
• [x] Kubeflow
• [x] clean up old codes

Thoughts

• Any additional endpoints we have to support from the server?

Schema

[jaywonchung]

Just noting here for when we integrate the client-side BatchSizeOptimizer with HuggingFace's Trainer, I think we can fetch the latest validation metric with TrainerState.best_metric on the call to on_evaluate.

[jaywonchung]

@show981111 I had to unpin Pydantic to get Zeus working with something else that I'm working on. Please take a look at #37 and rebase.

[jaywonchung]

@show981111 Can the server be used without K8s or KubeFlow as well?

[show981111]

@show981111 Can the server be used without K8s or KubeFlow as well?

Yes. can just run using uvicorn just like any fastapi app.

[vercel[bot]]

@show981111 is attempting to deploy a commit to the jaywonchung's projects Team on Vercel.

A member of the Team first needs to authorize it.

[vercel[bot]]

The latest updates on your projects. Learn more about Vercel for Git ↗︎

Name	Status	Preview	Comments	Updated (UTC)
zeus	❌ Failed (Inspect)			Apr 4, 2024 3:42pm

[show981111]

I finished the implementation. I haven't changed the /examples section yet since our implementation is still in review.

[jaywonchung]

Happy to see that you're making progress on documentation.

1. I pushed to master to (finally) enable local documentation building. Please rebase to / merge master, and after that, you can locally build & preview the documentation:

   $ pip install -r docs/requirements.txt
   $ mkdocs serve -a localhost:7777

2. Please create a Mermaid sequence diagram to describe how the server and client talks to each other.