[lfx-mentorship-2025-Sept-Nov] Support TFJob and PyTorchJob in Karmada via Default Resource Interpreters

[RainbowMango]

CNCF LFX mentorship
https://github.com/cncf/mentoring/tree/main/programs/lfx-mentorship/2025/03-Sep-Nov

LFX URL: https://mentorship.lfx.linuxfoundation.org/project/5b276804-9268-44c8-b339-1699c04397eb

Time Line: https://github.com/cncf/mentoring/tree/main/programs/lfx-mentorship/2025/03-Sep-Nov#timeline

Mentors: @lfbear @RainbowMango

What would you like to be added:
Implement default resource interpreters for TFJob and PyTorchJob workloads in Karmada. This includes writing the interpreter logic in Golang and Lua, ensuring proper scheduling and propagation of these workloads across member clusters. Additionally, create documentation on how to run TFJob and PyTorchJob on Karmada, and introduce the integration and best practices to the Kubeflow community.

Why is this needed:
Refer to #5115, and #6535

Iterated Tasks(draft, will update with details):

• Implement default(aka, built-in) resource interpreter in third-party resourcecustomizations
  • For TFJob (Add resource interpreter for TFJob. #6946)
  • For PyTorchJob (Add resource interpreter for PytorchJob #6826)
• Make a presentation at the Karmada community meeting.
  • Please feel free to add an agenda to the meeting notes.
  • By joining the Google groups, you will be able to edit the meeting notes.
• Write documentation on the Karmada website to demonstrate the use case. (Will add a overall documents to cover all kinds of multi-components jobs once them get on board.)
  • For TFJob
  • For PyTorchJob

[muskan-rathor]

Hi @RainbowMango 👋

I’m interested in this mentorship and have been trying to understand the project. From what I get, the goal is to make Karmada work with TFJob and PyTorchJob by writing resource interpreters in Go (and maybe Lua) so they can run smoothly across clusters.

I’m really excited about the mix of Kubernetes, ML workloads, and multi-cluster scheduling. Looking forward to learning more and contributing!

Thanks!

[muskan-rathor]

Hi @RainbowMango 👋

I'm starting to get familiar with the codebase and looking into how custom resource interpreters are implemented. Here’s what I’ve understood so far:

• pkg/util/interpreter/: Contains utilities for defining matchers and extracting/interpreting sub-resources from custom workloads.
• pkg/resourceinterpreter/: Appears to be the right place to add support for custom resources like TFJob and PyTorchJob.

I searched the codebase and couldn’t find existing interpreters for either TFJob or PyTorchJob, so I believe it’s valid to explore adding support for them.

Next, I plan to:

• Draft a basic matcher for TFJob
• Explore how to register and test the interpreter logic

Would appreciate any tips on how to best test custom interpreters!

Thanks!

[sriharan0804]

hi @RainbowMango
i am applied for this lfx's CNCF - Karmada: TFJob and PyTorchJob Resource Interpreters (2025 Term 3) and
I’m interested in contributing to this project. I have experience with Python, ML/DL, and Docker, and I’m eager to learn more about Karmada and its integration with ML workloads. Could you guide me on how I can get started?

[RainbowMango]

Maybe we can join the PyTorchJob ecosystem after this program.
https://pytorch.org/join-ecosystem/