Support Matrix#

Use this matrix to check the prerequisites and boundaries of the documented release. It describes the release configuration and does not certify every host, model provider, or generated workload.

Application Runtime#

The base application uses CPU containers and an external inference provider.

Area

Release requirement or guidance

Container architectures

Linux amd64 and arm64 manifest variants for all three POC Factory images

Local host

Docker Engine on Linux or a compatible Docker Desktop environment; verify daemon and filesystem access

Docker Compose

Plugin version 2.24.0 or later for the bundle’s env_file.required syntax; verify with docker compose version

Starting host capacity

At least 8 GB RAM and 20 GB free disk for the application stack; generated workloads and concurrency can require more

GPU

Not required to execute the POC Factory application containers; external model serving and generated workloads have separate requirements

Kubernetes chart prerequisites

Tagged chart README lists Kubernetes 1.23+ and Helm 3.0+; verify against your organization’s supported cluster versions

Backend scaling

Production chart example uses one backend replica; shared job coordination and storage require additional planning before scaling

Registry

nvcr.io/nvidia; access to the listed poc-factory product containers is required for pulls

The Compose requirement follows Docker’s environment-file reference. Use Installation for the released deployment files.

Inference and Generation#

Generation needs a complete inference profile. The chosen code-generation mode determines the additional Cursor requirement.

Setting

Behavior

MODEL_CODE, MODEL_REASONING, MODEL_BALANCED, MODEL_FAST

OpenAI-compatible chat-completions roles

MODEL_EMBEDDING

Embeddings endpoint for semantic blueprint matching

CODE_GENERATOR_TOOL=auto

Default contract-first Cursor path; requires Cursor CLI and a valid credential; no automatic switch to another generator after starting

CODE_GENERATOR_TOOL=cursor

Explicit full-project Cursor mode

CODE_GENERATOR_TOOL=llm

Explicit inference-backed full-project mode; separate behavior from contract-first generation

CURSOR_API_KEY in deployment

Also set MODEL_CURSOR=auto or a supported explicit Cursor model identifier

User-saved credentials

Available in Settings only when allowed by the deployment’s override policy

The supplied env.example model mappings are starting values. Verify availability and compatible operations with your selected provider rather than treating a model name or a /models response as evidence that the workflow works.

Candidate Validation#

The application records which checks ran. The deployment’s capabilities determine what that evidence can prove.

Environment

Configured candidate checks

Limit

NGC Compose bundle

Static analysis, tests, Docker build, and runtime checks

Requires a working host Docker daemon; the socket grants privileged host access

Helm production example

Static analysis and tests

Docker-dependent deployment and runtime gates are disabled and must remain explicit skips

Optional Capabilities#

These services and credentials are configured separately from the basic application stack.

Capability

Additional dependency

Authenticated GitHub intelligence

Optional read-only GitHub token; internal MCP proxy

Phoenix tracing

full Compose profile or corresponding Helm configuration

Model fine-tuning

Connected NeMo Microservices endpoints, credentials, datasets, and storage

Physical AI and OpenUSD

Workflow-specific services, assets, and downstream runtime requirements

SSH deployment of generated POCs

Authorized target access and a suitable remote Docker environment

Shared browser access

HTTPS, application authentication, stable signing/encryption keys, and managed secret storage

Next Steps#

Read Architecture to understand component boundaries, Installation to deploy, or Troubleshooting when a prerequisite check fails.