kubeflow/trainer

Support Elastic PyTorch in TrainJob

Open

#2,903 opened on Oct 21, 2025

View on GitHub
 (14 comments) (7 reactions) (1 assignee)Go (964 forks)github user discovery
area/apiarea/controllerhelp wantedkind/feature

Repository metrics

Stars
 (2,112 stars)
PR merge metrics
 (PR metrics pending)

Description

What you would like to be added?

As part of Kubeflow Trainer API, we designed the ElasticPolicy API which should allow users to run PyTorch in elastic mode: https://docs.pytorch.org/docs/stable/elastic/run.html

It would be nice if someone could drive KEP for that, and propose API changes to the next release: Trainer v2.2

elasticPolicy:
  minNodes: 2
  maxNodes: 5
  metrics:
    - type: Resource
      resource:
        name: nvidia.com/gpu
        target:
          type: Utilization
          averageUtilization: 75

Potentially, we should support elastic JobSet for that: https://github.com/kubernetes-sigs/jobset/issues/463

cc @kubeflow/kubeflow-trainer-team

/area api /area controller /help

Why is this needed?

We should support Elastic TrainJobs like in Training Operator V1 PyTorchJob.

Love this feature?

Give it a 👍 We prioritize the features with most 👍

Contributor guide

Research direction
Study the existing ElasticPolicy API in the Kubeflow Trainer API, understand PyTorch Elastic training, and review JobSet's elastic support. Then propose a KEP for API changes.
Tech stack
gopython
Domain
backendapidistributed systems
Issue type
Feature
DifficultyEstimated implementation difficulty for a new contributor, from 1 for very small changes to 5 for expert-level work.
4
Estimated timeA rough time range for an experienced contributor to investigate, implement, test, and prepare a pull request.
3-5 days
Activity statusHow available the issue appears right now: fresh, active, stale, blocked, or waiting on maintainer input.
Active
ClarityHow clearly the issue explains the expected change, acceptance criteria, and next step.
Needs investigation
Prerequisites
GoKubernetesKubeflowPyTorch
Newbie friendlinessA 1-100 score estimating how approachable this issue is for first-time contributors.
30