docs(vllm): vLLM Server v2.1 (0.24.0) + vLLM-Omni v1.4 AL2023 releases

[Jyothirmaikottu]

Summary

Documents the four AL2023 auto-releases that shipped on 2026-07-02 but whose docs were not fully updated by the release pipeline. The pipeline's step4-docs-pr only regenerates the Ubuntu vLLM gallery data (docs/src/data/vllm/) and skips vLLM-Omni entirely (vllm_omni: Not defined in tracker config), so the AL2023 server-cuda / omni-cuda gallery entries and changelogs were missing.

vLLM Server (AL2023) — v2.1 / vLLM 0.24.0

• Gallery data: add docs/src/data/vllm-server/0.24.0+amzn2023.7b3d595e-gpu-{ec2,sagemaker}.yml → tags server-cuda-v2.1, server-sagemaker-cuda-v2.1.
• Changelog: new v2.1.0 entry — vLLM 0.24.0, FlashInfer 0.6.12, transformers<5.10 pin dropped (now requires transformers ≥ 5.5.3), mistral_common now optional.
• Models: add JetBrains Mellum2-12B-A2.5B-Thinking (MellumForCausalLM, Smoke + Benchmark).

vLLM-Omni (AL2023) — v1.4

• Changelog: new v1.4.0 entry — SageMaker /v1/videos and /v1/videos/sync accept application/json again (JSON→multipart conversion restored). No framework bump (still vLLM-Omni 0.21.0rc1).
• SageMaker deployment doc: updated the video-route note — JSON works again as of omni-sagemaker-cuda-v1.4; the multipart path remains supported.

README

• Add both AL2023 releases to Release Highlights; drop the two oldest entries (SGLang v0.5.13, SGLang Server v1.0) to keep the list bounded.

Testing

• python docs/src/main.py --verbose — generation passes; available_images.md renders the new server-cuda-v2.1 / server-sagemaker-cuda-v2.1 rows.
• mkdocs build --strict — passes (exit 0).
• Docs prose wrapped at 150 cols to match the flowmark hook; no trailing whitespace; YAML validates.