Reference

Components, versions and licenses

This repository is licensed under the Apache License 2.0: see LICENSE and NOTICE. That license covers only the files in this repository: scripts, configuration examples and documentation. The repository does not contain or redistribute container images, model weights or third-party software. The operator downloads those separately, and they come under their own licenses and terms, which the operator must review and comply with.

Component

Pinned version

License

Where it runs

Codendum (this repository)

0.1.0

Apache-2.0

GB10 host, workstations

vLLM container image

vllm/vllm-openai:v0.30.0@sha256:8a69ffad015f138d7170c4ddc429e230a3bc1c1719f67e14324749df200a4b90 (index; the linux/arm64 manifest is sha256:4864d466…), CUDA 13.0.2

vLLM: Apache-2.0; the image bundles third-party software (for example the CUDA runtime and PyTorch) under their own licenses

GB10 host

Model

Qwen/Qwen3-Coder-30B-A3B-Instruct-FP8 at revision dcaee4d4dfc5ee71ad501f01f530e5652438fde0

Apache-2.0, according to the model card; check the model’s license file

GB10 host

OpenCode

V2 (2.0.x); 1.x with the V1 example

MIT

Workstations

nginx container image

nginx:1.30.5-alpine@sha256:0985e772fb9f729e6fa0980da05fca5d9c468e870eed43071545afa9d2e27d94

nginx: BSD-2-Clause; the Alpine base includes packages under their own licenses

GB10 host

Docker Engine, NVIDIA Container Toolkit

Preinstalled with DGX OS

Apache-2.0

GB10 host

actions/checkout

v7.0.1 @ 3d3c42e5aac5ba805825da76410c181273ba90b1

MIT

CI only

Code of conduct

Contributor Covenant 3.0

CC BY-SA 4.0

Documentation

Versioning and compatibility

Codendum follows Semantic Versioning. The public interface is:

  • script names and options;

  • the CODENDUM_* variable names;

  • the profile names;

  • the layout of the configuration examples.

Changes are recorded in CHANGELOG.md.

Area

Status

Offline tests (CI)

Ubuntu 24.04 on x64 and ARM64: lint, script tests against a mock vLLM, proxy tests (nginx container and distribution package), documentation and secret checks

Target host

DGX OS 7 on NVIDIA GB10. Run on a Lenovo ThinkStation PGX (DGX OS 7.2.3) on 2026-09-29: startup, smoke tests, proxy, a real OpenCode session and classroom simulations

vLLM

Built for v0.30.0. qwen3_coder requires ≥ 0.25 (earlier releases had a different parser implementation); CODENDUM_MAX_NUM_QUEUED_REQS requires ≥ 0.29

OpenCode

V2 native configuration; V1 format provided separately

Sources

Author

Codendum is created and maintained by Stefano Noferi (GitHub: stefanoferi).