NVIDIA details Workload Map for Kubernetes AI frameworks

News

NVIDIA says Workload Map provides a shared way to describe Kubernetes workloads across frameworks, with support for JobSet, KAI Scheduler and Run:ai.

NVIDIA has outlined Workload Map, an open-source component in its DSX AI Factory Platform, as a way to describe Kubernetes workloads once and reuse that structure across frameworks and tools. The company said the approach is designed to reduce the need for separate framework-specific parsers and adapters when managing distributed AI jobs.

Shared workload descriptions for distributed AI jobs

According to NVIDIA, operating an AI factory on Kubernetes requires a common understanding of distributed workloads, including pod templates, replica counts and status. The company said that workload types such as JobSet and LeaderWorkerSet often require separate custom adapters before schedulers, controllers or dashboards can interpret them. Workload Map is intended to provide a reusable structural description that can be consumed across those systems.

NVIDIA used a JobSet example to show the issue. In the manifest described in the post, a distributed training job includes one leader job and two worker jobs, with each pod requesting four GPUs. The company said the manifest represents a total request of 12 GPUs. It also said kubectl shows the JobSet, its three Jobs and their pods as separate resource entries, while a platform still needs a definition to understand the roles and aggregate the resource request.

How the definition works

NVIDIA said Workload Map, formerly called Karta, ships as a Kubernetes CRD and a Go library that interprets workload definitions. Each definition specifies where to find pod templates, replica counts and status fields for a particular workload type. The company said that lets schedulers and platform tools interpret supported workload types without maintaining separate parsing code for each framework.

For JobSet, NVIDIA said the definition maps native status fields into a shared vocabulary. In the example, the post said JobSet does not report a plain Running condition, so the Workload Map definition derives it using a jq expression over replicatedJobsStatus and maps Completed and Failed conditions to shared categories. The post also described child component paths that identify the JobSet’s replicated Jobs, pod templates, replica scaling and instance IDs.

The company said Workload Map includes prebuilt definitions for JobSet, PyTorchJob, RayCluster, RayJob, RayService, KServe, Knative, MPIJob, NVIDIA NIM, LeaderWorkerSet, Milvus, Grove PodCliqueSet and NVIDIA Dynamo. Additional workload types can be added by writing a YAML definition.

Use across inspection, scheduling and visibility

NVIDIA said the same JobSet definition can be used in multiple workflows. For inspection, the Workload Map CLI reads workloads from a cluster and renders them through their workload definitions. In the example shown in the post, the CLI output identified the leader and worker roles and reported a total of 12 GPUs for the workload. NVIDIA also showed an earlier Karta CLI view that resolved the JobSet into leader and worker roles and listed node placement.

For scheduling, the post said Workload Map can include gang-scheduling instructions. NVIDIA said KAI Scheduler uses Workload Map as a fallback in its pod grouper for workload types without a native plugin, while a matching definition can provide the structure and grouping information needed for gang admission and placement decisions.

For platform visibility, NVIDIA said NVIDIA Run:ai can register the workload type with a single API call using the catalog definition and then manage the JobSet as one workload. The post said the same definition supports the CLI, KAI Scheduler and NVIDIA Run:ai for inspection, scheduling and platform visibility.

Adoption and availability

NVIDIA said Workload Map is available under the Apache 2.0 license and is used in production within NVIDIA Run:ai. The company also said Elotl’s Nova uses Workload Map to interpret workload structure and resource requirements when selecting a cluster for distributed workloads.

The post said Workload Map is available through the NVIDIA DSX AI Factory GitHub organization. NVIDIA suggested that users running AI workloads on Kubernetes can start with the workload catalog to find a definition for a framework they already use, or adapt one for a custom workload type. The announcement did not specify a new product release date beyond the post’s publication date of Nov. 2, 2026.