External contributors cannot build compute-plane images — base images pinned to NVIDIA-internal registries

[sureshkrishnan-ai]

Problem

Building the platform images from source fails for anyone
outside NVIDIA, because several services pin base images
from registries that are not externally accessible:

1. urm.nvidia.com (internal JFrog Artifactory) — used as
   the distroless Go base in:
   • src/compute-plane-services/worker-init/MODULE.bazel
   • src/compute-plane-services/worker-utils/MODULE.bazel
   • src/compute-plane-services/worker-llm-credentials/MODULE.bazel
2. Private nvcr.io dev orgs — e.g. nvsnap Dockerfiles
   default to:
   ARG BASE_IMAGE=nvcr.io/0651155215864979/ncp-dev/nvsnap-agent-base:...
3. The published platform charts/images also live in a
   private nvcr.io org, so a catalog-only NGC key gets 403.

Impact

The source is open, but the build chain isn't reproducible
externally. Combined with entitlement-gated published
images, external users have no supported path to run the
self-hosted control plane.

Ask

• Repin internal bases to public equivalents (e.g.
  gcr.io/distroless, public CUDA images), or document
  supported substitutions
• Clarify whether external source builds are an intended
  use case, and what the intended path to registry
  access/entitlement is for self-hosted evaluation

Happy to test a patched build or contribute the repinning
if maintainers agree on target base images.

[kristinapathak]

This was partially fixed (bazel files) in 7b85129. More work is still needed here. @balajinvda, can you fix the rest?

[balaji-g]

Thanks for the detailed report. Triaging against current main:

1. worker-init / worker-utils / worker-llm-credentials distroless base — already resolved. On main these pin the public nvcr.io/nvidia/distroless/go base (by digest), not urm.nvidia.com. There are no urm.nvidia.com base-image pins left in any MODULE.bazel, so these build from a clean external checkout.

2. nvsnap default base — accurate. nvsnap/docker/agent/Dockerfile.app defaults BASE_IMAGE to a private image because nvsnap's base is its own agent image, which is not publicly pullable. nvsnap is a self-contained Bazel module and is intentionally excluded from the public build matrix; it is built only by internal CI. We're making that intent explicit (a note in the GHA build matrix and in nvsnap's CONTRIBUTING.md) so it no longer looks like an oversight. A publicly buildable base for nvsnap is follow-up work tracked by this issue.

3. Published images in a private NGC org — those are publish destinations, entitlement-gated by design, and separate from source-build inputs. Building the rest of the platform from source does not require them.

Net: the general compute-plane services build externally today; nvsnap is the one component that does not, by design, pending a public base. Documentation change is up in review now.

[balaji-g]

Resolved. The documentation change has merged.

Summary of the outcome:

• worker-init / worker-utils / worker-llm-credentials already build from a clean external checkout on main (public nvcr.io/nvidia/distroless/go base; no urm.nvidia.com base pins remain).
• nvsnap is now explicitly documented as internal-only: its OCI base is nvsnap's own agent image, which is not publicly pullable. It is a self-contained Bazel module already excluded from the public build matrix, and that intent is now recorded in the GHA build matrix and nvsnap's CONTRIBUTING.md.
• The private NGC org for published images is entitlement-gated by design and is separate from source builds.

Net: the general compute-plane services are externally buildable today; nvsnap is intentionally internal-only for now. A publicly buildable nvsnap base would be a separate enhancement; please open a new issue if you'd like to pursue that. Thanks again for the report.

[balaji-g]

Resolved. The documentation change has merged.

Summary of the outcome:

• worker-init / worker-utils / worker-llm-credentials already build from a clean external checkout on main (public nvcr.io/nvidia/distroless/go base; no urm.nvidia.com base pins remain).
• nvsnap is now explicitly documented as internal-only: its OCI base is nvsnap's own agent image, which is not publicly pullable. It is a self-contained Bazel module already excluded from the public build matrix, and that intent is now recorded in the GHA build matrix and nvsnap's CONTRIBUTING.md.
• The private NGC org for published images is entitlement-gated by design and is separate from source builds.

Net: the general compute-plane services are externally buildable today; nvsnap is intentionally internal-only for now. A publicly buildable nvsnap base would be a separate enhancement; please open a new issue if you'd like to pursue that. Thanks again for the report.