Skip to content

Latest commit

 

History

History
65 lines (48 loc) · 2.85 KB

ROADMAP.md

File metadata and controls

65 lines (48 loc) · 2.85 KB
Raw

Kubeflow Arena Roadmap

Kubeflow Arena 2024 Roadmap

This document defines a high level roadmap for Arena development.

  • Objective：Simplify the user experience by deeply integrating with the Kubeflow Ecosystem.

    • Kubeflow Integration
      • Prepare Arena for release v1.0.0 alongside kubeflow v1.10.
      • Develop a seamless integration with the Training Operator to help simplify model training using command line.
      • Integrate with Kubeflow Pipelines to facilitate model training from a Pipeline.
      • Enable mode serving with KServe.
      • Add documentation to Kubeflow website:
        • Installation, uninstallation, and upgrade processes.
        • Guide for tfjob, mpijob, pytorchJob.

  • Objective：Amplify the Extensibility of the Arena for Different ML frameworks, AIGC models and platforms.

    • Support DeepSpeed Training Job.
    • Support for submitting and managing llm fine-tuning jobs, like DeepSpeed.
    • Enable model serving for an expanded set of models like Baichuan, LLaMA, ChatGLM, Falcon, and more.
    • Extend platform support to include ARM.
    • Integrate Fluid project for efficient data management.
    • Add support for Ray Job management with Kuberay.

  • Objective: Boost Performance and Stability.

    • Regularly publish recommended practices documentation.
    • Enhancements on Arena SDK.
    • Enhance code quality by:
      • Reduce repetitive code.
      • Enhance unit test.
    • Implement automated End-to-End Test:
      • Add integration tests using GitHub Actions.
      • Strive for more than 60% Test Coverage of Supported Features.

Kubeflow Arena 2019 Roadmap

Core CUJs

Objectives: "Make Arena easily to be integrated with External System."

  • Integration
    • Provide the Java, Python and C++ API for system interaction
    • Integrate with Pipelines to support Standalone Job, MPI Job, Estimator Job

Objectives: "Simplify the user experience of the data scientists and provide a low barrier to handle different kind of training jobs and serve different models."

  • High Level Interfaces:
    • Submit and manage other data processing and machine learning jobs, like Spark, Flink, XDL, PyTorch, MXNet
    • Submit and manage Model Serving with KF Serving

Objectives: "Make Arena support the same Operator compatiable with different API version, so the upgrade of operator doesn't impact the existing users' experiences."

  • Compatibility:
    • v1aphla2 and v1 TFJob
    • v1alpha1 and v1aphla2 MPIJob

Objectives: "Enchance the software quality of Arena so it can be in the quick iteration"

  • Refactor the source code

    • Move Training implementation from cmd into pkg

  • Automatic Test Enhancement:

    • Unit test
    • Integration test