2.2 Release Notes

GPUaaS Node Reboot functionality/Management -(HAI-2420)

This delivers a controlled and role-aware GPUaaS node reboot capability, ensuring nodes can be safely restarted for maintenance, upgrades, and recovery.

Reduces operational risk and unplanned downtime, improving GPUaaS production reliability.

Unify GPUaaS Instance Provisioning / Management - (HAI-3972)

This delivers a unified, service-based deployment model that brings VM and GPUaaS pod workloads under a single, consistent experience.
It enables reusable GPU service templates, advanced GPU scaling, service exposure, and multi-region support.

All existing GPUaaS subscription APIs, UI flows, scripts and integrations continue to work exactly as they do today. It gives a simpler, standardised way to offer GPU services.

Service functionality now enabled consistently across all workload types including Billing management and Operational capabilities. This improves GPU workload adoption & reduces operational complexity.

Ansible Configuration Support for GPUaaS instances - (HAI-3929)

This extends existing Ansible automation capabilities from VM instances to GPUaaS pod instances, enabling consistent, secure playbook execution across all compute types.

It offers the ability to install and customise Software Stacks onto the Instances during the Workload deployments.

SystemD Support for GPUaaS instances - (HAI-3522)

Enables VM-like systemd support inside GPUaaS containers, allowing users to manage system services in a VM like manner. Now supports nested docker and Kubernetes environments in a GPUaas Pod

Service provider would be able to offer Kubernetes clusters, docker based workloads to their users as a single click usable service. This expands the range of GPU workloads that can run on GPUaaS without requiring full VM’s & with GPU over subscription.

KVM Cluster Redundancy - (HAI-3998)

Enabling Storage Redundancy & Disk Management - It moves redundancy control to the per-VDisk level, allowing each VM disk to independently enable or disable replication based on business need. 

Provides vDisk-level high availability and data protection, reducing data-loss risk while enabling workload-specific redundancy that optimizes storage usage and simplifies operations.

Infiniband Support for Multi-GPU Pods - (HAI-3135)

High-bandwidth, low-latency GPU-to-GPU across nodes with NVIDIA GPUDirect RDMA

Enables faster multi-node distributed training & HPC workloads

HaiDra Migration onto GoLang (HaiDra Re-architecture) - (HAI-4193)

This epic delivers HaiDra v2.0, re-architected as a scalable Golang service with complete API alignment and seamless migration from the existing Python implementation.

It improves performance, reliability and long-term maintainability while enabling HaiDra to support unified pod and VM instance management at scale.

Enable Concurrent Recipe Execution to Improve System Throughput - (HAI-3886)

Enabled concurrent execution of compatible recipes by removing the sequential execution and introducing parallel recipe processing in a centralised, configurable concurrency control.

Improves system throughput and scalability for large installations, prevents long-running operations from blocking other users and provides a robust foundation for safely adding more system and custom recipes over time.

Issue Key

Issue Summary

Business Impact/Value to End users

HAI-4596

High Latency on Accelerators API Endpoints

The Accelerators page now loads significantly faster, reducing wait times from up to 10 seconds to under 2 seconds, providing a much smoother experience when managing GPUaaS subscriptions.

HAI-4555

Performance Degradation on Accelerator Pool Details Page

The Accelerator Pool details page now loads reliably even with large volumes of historical statistical data, ensuring consistent performance for pool administrators at scale.

HAI-4567

Duplicated API Calls on Accelerator Pools Page

Redundant background API requests on the Accelerator Pool details page have been eliminated, reducing unnecessary server load and improving overall page responsiveness.

HAI-4149

Remove Service Port Buttons When Gateway is Unavailable

Service port exposure controls are now hidden in regions where no gateway is configured, preventing users from attempting to set up features that are not available in their region and reducing confusion.

HAI-4265

Remove Remaining Files from sources.list.d During Node Reset

Node resets now perform a complete cleanup of all repository configuration files, ensuring nodes start fresh without leftover artifacts that could interfere with subsequent initialisation or re-provisioning.

Key

Release Notes

HAI-4734

Internal HostedAI system recipes were incorrectly appearing in the Recipe dropdown during service creation. These are now filtered out, ensuring only user-relevant recipes are shown during service setup.

HAI-4731

Service creation was failing with an error when the optional recipe field was left blank. Services can now be created successfully without selecting a recipe, as originally intended.

HAI-4607

Users created with pre_onboard=true were incorrectly assigned "invited" status instead of "active", preventing them from seeing teams or accessing any resources despite successful login. Pre-onboarded users now immediately receive "active" status with full team access.

HAI-4559

Editing or updating an existing service was failing with an "Invalid recipe id" error despite the service configuration being valid in the database. Service updates now complete successfully.

HAI-4557

During instance creation, reselecting a previous step option caused all subsequent steps to disappear, requiring users to start over. The creation flow now correctly maintains all steps regardless of how the user navigates through it.

HAI-4556

Attempting to edit or clone a service caused the UI to revert to the service details tab instead of applying changes. Users can now successfully edit and clone services without interruption.

HAI-4521

The billing comparison widget was displaying the same month name for both the current and previous month (e.g., "Dec v Dec"), making it impossible to distinguish between the two periods. Correct month labels are now displayed.

HAI-4429

Deleting a workspace was performing a hard delete of the database record, causing all historical billing data to be permanently lost. Workspace deletion now performs a soft delete, preserving complete billing history for audit and reporting.

HAI-4424

The individual workspace billing module was displaying $0 for costs even when resources had been actively consumed. Workspace costs are now accurately calculated and reflected in the billing dashboard.

HAI-4422

When instances were deleted, their accumulated billing costs were no longer appearing in the Workspace Billing module, resulting in incomplete and understated cost reports. Deleted instance costs are now correctly included in workspace billing summaries.

HAI-4419

Adding a GPU to a pool was failing when the node's fully qualified hostname exceeded 63 characters (common with cloud provider FQDNs), producing a Kubernetes label validation error. Hostname handling now correctly supports longer hostnames.

HAI-4416

Selecting the "Last 5 min" time window on the Accelerators page was not filtering metrics correctly — the API was always returning all-time data regardless of the selection. All time window filters, including "Last 5 min", now work as expected.

HAI-4397

Several stability issues in the HAISHARE Credit Based Scheduler were identified and resolved, improving the reliability and fairness of GPU resource scheduling across namespaces.

HAI-4395

GPUaaS subscriptions in a "Pending" state — where resources are not yet provisioned or accessible — were being charged for GPU, CPU, RAM, and storage. Billing is now only applied to active, running subscriptions.

HAI-4364

GPUaaS nodes in regions with Compute status enabled were being assigned an incorrect public SSH key from a different node, which could prevent proper node management. The correct SSH key is now applied to GPUaaS nodes in all regions.

HAI-4204

During node initialization, the internal IP address was not being detected on nodes where the internal network interface lacked a default route. IP selection logic has been updated to correctly identify the internal IP across all supported network configurations.

HAI-4036

When creating a GPU pool in low security mode with Spatial Mode selected, the HAISHARE Scheduler was incorrectly starting in normal (temporal) mode. Pool creation now correctly applies the Spatial Mode configuration to the HAISHARE Scheduler.

HAI-3434

Logos uploaded via the SMTP configuration in the admin panel were not being applied to outgoing system emails, which continued to use the default HostedAI branding. Custom SMTP logos are now correctly saved and applied to all outgoing email communications.