v3.6.0

Features

• Add EFA support in manifest for training jobs (#345)
• Add end-to-end example documentation (#350)
• Add 4 new HyperPod GA regions (ca-central-1, ap-southeast-3, ap-southeast-4, eu-south-2) (#360)

Enhancements

• Update documentation for elastic training arguments (#343)
• Upgrade Inference Operator helm chart (#346)
• Update MIG config for GPU operator (#358)
• Release Health Monitoring Agent 1.0.1249.0_1.0.359.0 with enhanced Nvidia timeout analysis and bug fixes (#361)

Bug Fixes

• Fix canary test failures for GPU quota allocation integration tests (#356)
• Fix region fallback logic for health-monitoring-agent image URIs (#360)
• Remove command flag from init pytorch job integration test (#351)
• Skip expensive integration tests to improve CI performance (#355)

Elastic Training Support for ReInvent Keynote 3

• Adding new command line arguments to the HyperPodTrainingOperator to support elastic training capabailities
  • --elastic-replica-increment-step, --max-node-count, --elastic-graceful-shutdown-timeout-in-seconds, --elastic-scaling-timeout-in-seconds, --elastic-scale-up-snooze-time-in-seconds, --elastic-replica-discrete-values
• Enables dynamic scaling of compute resources during training operations

Hyperpod CLI V2 with Nova recipe support

Hyperpod CLI V2 with Nova recipe support

Parker CLI, Fractional GPU Feature

-Added hp-devspace command set for ML dev environments

-New commands: create, list, get, update, delete, geturl for dev space management
Support for namespace-based auth and resource isolation
Added auth and get-config commands to check permissions and view default settings

Users can request partial GPU resources using MIG profiles instead of full GPUs

Added --accelerator-partition-type, --accelerator-partition-count, accelerator-partition-limit
New list-accelerator-partition-type command to view available GPU partitions for instance types

v3.3.1

Features

• Describe cluster command
  • User can use hyp describe cluster to learn more info about hp clusters
• Jinja template handling logic for inference and training
  • User can modify jinja template to add parameters supported by CRD through init experience of inference and training, for further CLI customization
• Cluster creation template versioning
  • User can choose cloudformation template version through cluster creation expeirence
• KVCache and intelligent routing for HyperPod Inference
  • InferenceEndpointConfig CRD supported is updated to v1
  • KVCache and Intelligent Routing support is added in template version 1.1

init experience Launch

Features

• Init Experience
  • Init, Validate, and Create JumpStart endpoint, Custom endpoint, and PyTorch Training Job with local configuration
• Cluster management
  • Bug fixes for cluster creation

Bug fixes

Features

• Fix for production canary failures caused by bad training job template.
• New version for Health Monitoring Agent (1.0.790.0_1.0.266.0) with minor improvements and bug fixes.

Bug Fixes

• Bug Fixes in cluster creation

v3.2.0 - Cluster Management and Init Experience

Features

Cluster management

    Creation of cluster stack
    Describing and listing a cluster stack
    Updating a cluster

Init Experience

    Init, Validate, Create with local configurations

v3.1.0: Release tg (#209)

v3.1.0 (2025-08-13)

Features

• Task Governance feature for training jobs.