docs: restructure site, add changelogs, improve content

[junpuf]

Summary

Major documentation site restructure for clarity, navigation, and content quality.

Site Structure

• Nav bar: Home → User Guide → Blog Posts → Resources
• User Guide: vLLM, vLLM-Omni, Ray, PyTorch, Base — each multi-page with sidebar nav (Overview, Supported Models, Deployment, Configuration, Changelog)
• Resources: Reference (Image Access, Available Images, Region Availability, Support Policy, Release Notifications) + Security

Home Page

• New landing page: DLC intro, quick start example, use-case cards (3 + 2 grid), footer links
• Replaces previous README.md copy
• New "Build Your Own Image" card pointing to Base guide

Framework Pages

• vLLM, vLLM-Omni, Ray: split monolithic pages into focused sub-pages with changelogs
• PyTorch: new guide for the AL2023 PyTorch training image (overview + EC2/SageMaker deployment + changelog)
• Base: new guide for the lightweight CUDA + Python base images (overview + changelog, covers both v1 / CUDA 12.9 and v2 / CUDA 13.0)
• Updated SageMaker deployment docs with standard-supervisor features (PR feat: wire vLLM SageMaker entrypoint to standard-supervisor #6044)
• Added "What's Included" sections, default port columns, model coverage labels (Smoke / Benchmark / Smoke + Benchmark)
• Documented SageMaker model resolution order, S3 model loading via runai-streamer, OpenAI-compatible API endpoints

Release-Notes Pipeline Removal

The auto-generated docs/releasenotes/ pages were not wired into the site nav and had no inbound links — effectively dead surface. The manually-maintained User Guide changelogs now own image history.

• Deleted docs/releasenotes/ tree
• Removed generate_release_notes() and helpers from docs/src/generate.py
• Removed release-notes templates, table config, and tests
• Stripped announcements: and packages: from 91 data YAMLs (no longer consumed)
• Trimmed scripts/autocurrency/docs-pr.sh to drop docker-image introspection and the dead announcements:/packages: emit
• Removed docs_packages: from .github/config/autocurrency-tracker.yml
• Updated test_generate.py to drop the release-notes mock

Reference

• Separated Region Availability into its own page
• Trimmed Image Access to essentials
• Simplified Release Notifications

Fact-Check Findings Fixed

• RAYSERVE_NUM_GPUS was a phantom variable in Ray deployment docs (no script reads it) — removed
• CodeArtifact (CA_REPOSITORY_ARN) was incorrectly implied to work on Ray EC2 image — corrected to SageMaker-only

CI Hardening

• mkdocs build → mkdocs build --strict in docs-test.yml. PRs that introduce broken internal links, missing nav targets, broken anchors, or page conflicts will now fail CI (existing test_links.py covered .md link targets but not anchors).
• Added /README.md, /tutorials/README.md, /DEVELOPMENT.md to mkdocs.yaml exclude_docs so contributor docs don't conflict with the published site.
• Sidebar nav section titles darkened for readability.

Test plan

• mkdocs build --strict passes locally
• All 74 docs tests pass (pytest test/docs/)
• pre-commit run --all-files passes (except actionlint which needs network in the local env)
• All pages return 200 (verified locally)
• Autocurrency unit tests: same pre-existing baseline (no new regressions)

[Eren-Jeager123]

autocurrency/docspr part lgtm

[sirutBuasai]

A small thing is that release notes used to have the main package version such as python version, cuda version, pytorch version, etc. Now the new changelogs doesn't display which version or commit it comes from but rather only the main package (eg: vllm) source commit.

Another small nit is that we should try to use the variables such as EC2/ECS/EKS/SageMaker variables from global.yml as much as possible to keep everything in the docs standard and any future changes are easy to make

[junpuf]

@sirutBuasai thanks for the review. Both points addressed in 6fcc8a0e:

1. Per-release version info on changelogs. PyTorch and Ray changelogs already include per-release framework versions; only the vLLM changelog was light. Added a **Bundled versions:** CUDA · Python · FlashInfer · DeepEP line to each vLLM entry (v1.0, v1.1, v1.2, v1.3), plus the wheel-version tag (e.g., 0.20.0.dev361+amzn2023.3f5bd482) inline with the source-commit link. Sourced from docker/vllm/versions.env at each release commit.

2. Use global.yml variables for service names. Converted hardcoded EC2, EKS, SageMaker, Amazon SageMaker AI literals to {{ ec2_short }} / {{ eks_short }} / {{ sm_short }} / {{ sagemaker }} across:

• docs/{vllm,vllm-omni,ray,pytorch}/index.md
• docs/index.md (landing page cards + walkthrough line)
• docs/user_guide/index.md

Image tag URLs (*-sagemaker-cuda), file paths, and code blocks left literal since they're identifiers, not service names.

One follow-up thought worth raising: it's worth weighing whether this level of templating is worth the effort across the doc tree. Writing EC2 is meaningfully easier to read and write than {{ ec2_short }}, well-known acronyms like EC2 / EKS / SageMaker are unlikely to be mistyped by humans or coding agents. Variables make a clear difference for complex / evolving strings (full product names, version-pinned identifiers, paths) where a future rename should propagate automatically. For two-letter service abbreviations the cost-benefit is closer — the consistency win is real but the maintenance / readability cost is non-trivial.