Support New AI Frameworks Faster Using NVIDIA Workload Map

NVIDIA Unifies AI Workload Management on Kubernetes

NVIDIA has released Workload Map, a new open-source component within its DSX AI Factory Platform designed to standardize how distributed AI workloads are described and managed on Kubernetes. The initiative addresses a persistent operational challenge where each AI framework, from JobSet to RayCluster, uses a unique Custom Resource Definition (CRD), forcing developers of schedulers and management platforms to build and maintain separate, framework-specific adapters. By providing a common descriptive layer, Workload Map enables tools to interpret the structure, resource requirements, and status of any supported workload without custom parsing code, significantly accelerating the integration of new AI frameworks into production environments.

At its core, Workload Map functions as a translation layer, using its own Kubernetes CRD (named `Karta`) and a Go library to interpret framework-specific schemas. A developer writes a single YAML definition for a workload type, mapping its unique structure—such as pod templates, replica counts, and status fields—to a shared vocabulary. This allows disparate tools, including the KAI Scheduler for gang scheduling and the NVIDIA Run:ai platform for visibility, to consume the same definition. The project already includes a catalog of pre-built definitions for many popular frameworks.

Key Capabilities of Workload Map

• Provides a common structural view for workloads, resolving components like leader and worker pods.
• Extracts pod templates, replica counts, and status from framework-specific CRDs.
• Enables unified inspection via the Workload Map CLI (`kli`), gang-scheduling via KAI Scheduler, and platform management in NVIDIA Run:ai from a single definition.
• Ships with pre-built definitions for JobSet, PyTorchJob, RayCluster, RayJob, KServe, MPIJob, NVIDIA NIM, and others.

The immediate impact of this standardization is a reduction in development overhead and faster support cycles for the entire AI infrastructure ecosystem. Platform providers like Elotl, which uses Workload Map for multi-cluster scheduling, no longer need to treat support for a new framework as a major roadmap item requiring custom code. For organizations operating large-scale AI factories, this translates to a more unified and manageable environment where heterogeneous workloads from training to inference can be understood and scheduled efficiently, removing a significant layer of operational complexity.

With Workload Map, NVIDIA is abstracting away the complexity of individual AI frameworks at the infrastructure level. This move standardizes the operational layer of the AI stack, making its DSX AI Factory Platform a more cohesive and indispensable foundation for running diverse Kubernetes-native AI workloads.