How dsx-ai-factory/nvcf works
NVCF is a platform for deploying, managing, and invoking GPU workloads across clusters. A representative user workflow configures the CLI, creates and deploys a function, then sends an invocation through the invocation plane to a running workload. The broader architecture also includes control-plane lifecycle management, NVCA cluster integration, artifact registries, asynchronous tasks, and observability. The README supports the platform-level links; sampled source confirms selected LLM and registry behaviors, but most service internals were not inspected.
Main entry points: nvcf-core Control APIs · cmd NVCF CLI · FunctionLlmService.java Function management · app.rs HTTP invocation
Components 22
Control plane
- Control APIs
src/control-plane-services/cloud-functions/nvcf-core - Function management
src/control-plane-services/cloud-functions/nvcf-core/src/main/java/com/nvidia/nvcf/service/function/FunctionLlmService.java - Function records
src/control-plane-services/cloud-functions/nvcf-core/src/main/java/com/nvidia/nvcf/persistence/function/entity/FunctionEntity.java - LLM auth and routing
src/control-plane-services/cloud-functions/nvcf-core/src/main/java/com/nvidia/nvcf/grpc/GrpcLlmService.java - Cluster management
src/control-plane-services/instance-cluster-management
Invocation plane
- HTTP invocation
src/invocation-plane-services/http-invocation/crates/server/src/app.rs - gRPC proxy
src/invocation-plane-services/grpc-proxy/main.go - LLM gateway
src/invocation-plane-services/llm-api-gateway/server/server.go - Rate limiting
src/invocation-plane-services/ratelimiter/cmd/main.go
Compute and observability
- GPU cluster agent
src/compute-plane-services/nvca - Shared storage
src/compute-plane-services/nvca/pkg/storage/sharedstorage.go - Secrets agent
src/compute-plane-services/ess-agent - Telemetry collector
src/compute-plane-services/byoo-otel-collector/internal/otelcollector/otelcollector.go - Running workloads
- GPU clusters
Platform services
- Artifact registry access
src/libraries/java/nv-boot-parent/nv-boot-starter-registries/src/main/java/com/nvidia/boot/registries/service/registry/client/volcengine/VolcengineArtifactValidationService.java - Task services
src/control-plane-services/cloud-tasks - Artifact registries
Other components
- Platform user
- Invocation clients
Connections 23
- Platform user → NVCF CLI: configures and operates
README.md - Platform user → NVCF UI: manages workloads
README.md - NVCF CLI → Control APIs: manages functions
README.md - NVCF UI → Control APIs: manages functions
README.md - Control APIs → Function management: manages lifecycle
README.md - Function management → Function records: persists state
README.md - LLM auth and routing → Function management: uses LLM config
src/control-plane-services/cloud-functions/nvcf-core/src/main/java/com/nvidia/nvcf/grpc/GrpcLlmService.java - Control APIs → Cluster management: coordinates clusters
README.md - Invocation clients → HTTP invocation: sends requests
README.md - Invocation clients → gRPC proxy: sends gRPC
README.md - HTTP invocation → Running workloads: routes invocation
README.md - gRPC proxy → Running workloads: routes invocation
README.md - LLM gateway → LLM auth and routing: checks LLM auth
README.md - Rate limiting → HTTP invocation: limits requests
README.md - GPU cluster agent → GPU clusters: registers and manages
README.md - Running workloads → GPU clusters: runs on
README.md - GPU cluster agent → Shared storage: reconciles storage
src/compute-plane-services/nvca/pkg/storage/sharedstorage.go - Secrets agent → Running workloads: provides secrets
README.md - Running workloads → Telemetry collector: emits telemetry
README.md - Control APIs → Artifact registry access: validates artifacts
src/libraries/java/nv-boot-parent/nv-boot-starter-registries/src/main/java/com/nvidia/boot/registries/service/registry/client/volcengine/VolcengineArtifactValidationService.java - Artifact registry access → Artifact registries: checks artifacts
README.md - NVCF CLI → Task services: manages tasks
README.md - Control APIs → Task services: supports workloads
README.md