Skip to content

README.md

Latest commit

 

History

History
93 lines (68 loc) · 4.15 KB

README.md

File metadata and controls

93 lines (68 loc) · 4.15 KB
Raw

Base Images

Foundation layer for ODH (Open Data Hub) workbenches and runtimes. All images are CentOS Stream 9 (c9s) with Python 3.12.

RHOAI (Red Hat OpenShift AI) workbenches and runtimes use RHEL 9.6 EUS base images from AIPCC, not these c9s images.

OS version policy (EL9 only)

This repository builds on Enterprise Linux 9 only. ODH base-images/ and ODH workbench/runtime builds use CentOS Stream 9 (stream9, c9s) via quay.io/opendatahub/odh-base-image-*-py312-c9s. RHOAI/Konflux builds use RHEL 9.6 EUS (AIPCC) via quay.io/aipcc/base-images/*-el9.6. The ubi9-python-* directory names denote EL9 compatibility; they are not FROM ubi9 bases. Accidental upgrades to stream10, ubi10, or rhel10 are out of scope until explicitly planned.

Pin OS FROM references with an explicit tag and digest — for example quay.io/centos/centos:stream9@sha256:…. Digest-only refs (without a tag) let Renovate/MintMaker follow the registry default tag, which can jump to stream10. Tag bumps for quay.io/centos/centos are blocked by allowedVersions in .github/renovate.json5; digest refreshes within stream9 remain allowed.

Layout

Directory Purpose
cpu/ CPU-only base image
cuda/ NVIDIA CUDA (12.9, 13.0)
rocm/ AMD ROCm (6.4, 7.1)
build-args/ .conf files with INDEX_URL per variant
utils/ Shared scripts: aipcc.sh, dnf-helper.sh, fix-permissions, pip.conf.in, uv.toml.in
copr/ Tool to rebuild Fedora SRPMs for EL9 (README)

How Base Images Are Consumed

Downstream Dockerfiles reference base images via build arg:

ARG BASE_IMAGE
FROM ${BASE_IMAGE}

Resolved at build time by Tekton / Konflux pipelines.

Building Locally

Base images are not in the Makefile -- they are built by Tekton/Konflux pipelines only. To build locally, invoke podman build directly. Build context is the repo root (Dockerfiles use COPY base-images/utils/...):

podman build \
  --build-arg-file=base-images/build-args/cpu.conf \
  -f base-images/cpu/c9s-python-3.12/Dockerfile.cpu .

Python Versions and RHEL Lifecycle

Red Hat ships 2 out of every 3 CPython releases in AppStream, skipping the version released ~6 months before the next RHEL major (3.10 was skipped for RHEL 9, 3.13 for RHEL 10).

RHEL Default (full-life) AppStream Skipped
9 3.9 3.11 (9.2, retires May 2026), 3.12 (9.4, retires Apr 2027) 3.10
10 3.12 3.14 arriving via z-stream 3.13
  • Python 3.13 is EPEL-only, not in RHEL AppStream
  • Python 3.14 is in CentOS Stream 9/10 and shipping to RHEL 9.8 / 10.2

This repo ships Python 3.12, aligned with both RHEL 9 AppStream (supported until Apr 2027) and RHEL 10 full-life. The next version to consider is 3.14 (3.13 was skipped). The AIPCC wheel builder already has CI for 3.14 (AIPCC-8168).

References:

Adding a New Accelerator Version

The CUDA and ROCm Dockerfiles are adapted from upstream vendor references:

Steps:

  1. Copy an existing version directory (e.g. cuda/12.9/ -> cuda/13.1/)
  2. Update the Dockerfile stages from the upstream vendor Dockerfiles for the new SDK version
  3. Update cuda-repos/ repo files if needed (GPG keys, baseurls)
  4. Create build-args/<variant>.conf with the correct INDEX_URL
  5. Add Tekton pipeline YAMLs in .tekton/