nvcf

A platform, not a single binary

NVIDIA Cloud Functions, or NVCF, is a platform for deploying, managing, and running GPU accelerated workloads at scale. The repository is a monorepo that contains the service code, deployment assets, documentation, examples, CLI code, agent skills, and validation tooling, so the pieces needed to run the platform ship together. The stated goal is to route inference, streaming, and other GPU work to worker clusters so teams can scale demanding jobs with less infrastructure to operate themselves. The docs and roadmap live alongside the code, and the project points to build.nvidia.com as a service powered by NVCF, which shows the platform running in a real product. For someone evaluating it, the monorepo shape means the control plane, the invocation path, and the GPU integration are all present and can be deployed as a self managed system. The README presents NVCF as the layer that handles function lifecycle, routing, and cluster connection rather than as a model server, so it sits above the individual inference services. That distinction is important: a team brings its own model container or Helm chart, and NVCF handles the surrounding concerns of exposing, scaling, and securing it. The inclusion of agent skills and validation tooling in the same repo also suggests the project cares about operating the platform day to day, not just booting it once.

Control plane and invocation plane

The architecture splits into a control plane and an invocation plane. The control plane exposes the NVCF API, manages function and deployment state, handles secret management, and coordinates platform operations. The invocation plane receives HTTP, streaming, and gRPC requests, applies routing and rate limiting, and sends work to the running function workloads. GPU clusters connect through the NVIDIA Cluster Agent, which the README calls NVCA. NVCA registers GPU resources and manages workload execution on GPU nodes. The README includes a diagram of how self managed NVCF can span regions and GPU clusters, which matters for operators who need the platform close to their data or their users. Observability, dashboards, and runbooks are listed as part of the operator story, so monitoring health and debugging workload behavior is part of the design rather than an afterthought. This separation lets the request path stay thin while the management path handles the heavier bookkeeping of deployments and secrets. By naming NVCA explicitly, the README shows that GPU scheduling is delegated to a dedicated agent instead of being baked into the planes, which keeps the cluster integration clean. The multi region diagram also signals that the platform is built for operators who run more than one cluster and need a single control surface across them.

Functions versus tasks

NVCF distinguishes two workload types. A function is a long running, invokable workload used when a client needs an endpoint for inference, streaming, or another service style GPU workflow. Functions can be packaged as a container when the workload is a single service with health and inference endpoints, or as a Helm chart when it needs multiple coordinated containers, sidecars, or other Kubernetes resources. A task is an asynchronous, run to completion workload, used for batch inference, evaluation, fine-tuning, data preparation, or similar jobs that do not need a standing endpoint. This split lets the platform fit both a live API and a background job without forcing one model onto the other. The README keeps the description at the architecture level and points to installation and API reference docs for the concrete steps. The container versus Helm choice for functions is a practical one: a single model server maps cleanly to a container, while a pipeline with sidecars and extra services needs the fuller Kubernetes description that a Helm chart provides. The project is written in Go and is released under the Apache 2.0 license, with its documentation hosted at docs.nvidia.com/nvcf, where the deployment and invocation details are expanded beyond the architecture overview given in the README.

Editorial conclusion

The platform is released under the Apache 2.0 license, written in Go, and documented at docs.nvidia.com/nvcf.