docs(vllm-omni): refresh for v1.7.0 (vLLM-Omni 0.28.0) (#6771)

* docs(vllm-omni): refresh for v1.7.0 (vLLM-Omni 0.28.0)

- drift: correct s3tokenizer version reference 0.26.0 -> 0.28.0 in index.md
- Add changelog entry for v1.7.0 (vLLM-Omni 0.28.0, DLC PR #6680)

* docs(vllm-omni): focus v1.7.0 changelog Highlights on container changes

- Replace upstream-PR Highlights with the container-side change: PEFT LoRA validated on SageMaker (new endpoint test, PR #6680)
- Preserve the upstream v0.26.0...v0.28.0 compare link on the source line

* docs(vllm-omni): drop upstream-PR frontend bullet from v1.7.0 Changes

* docs(vllm-omni): add v1.7 (0.28.0) entry to README What's New

* docs(vllm-omni): add v1.7 (0.28.0) to available_images data

---------

Co-authored-by: Yadan Wei <weiyadan@amazon.com>

Author: Yadan-Wei

README.md

@@ -31,6 +31,7 @@ ______________________________________________________________________
 
 ### 🚀 Release Highlights
 
+- **[2026/09/21]** [vLLM-Omni v1.7 (AL2023)](https://gallery.ecr.aws/deep-learning-containers/vllm) — EC2: `omni-cuda-v1.7` · SageMaker: `omni-sagemaker-cuda-v1.7` · vLLM-Omni `0.28.0` (up from 0.26.0); PEFT LoRA adapters validated on SageMaker — per-request `lora` selection on `/v1/images/generations`; FlashInfer 0.6.16.post3, NCCL pinned 2.30.7, CUDA-13 mooncake wheel.
 - **[2026/09/11]** [vLLM v0.29.0 (Ubuntu)](https://gallery.ecr.aws/deep-learning-containers/vllm) — EC2: `0.29.0-gpu-py312-ec2` · SageMaker: `0.29.0-gpu-py312` · New models: Hy4-preview (Tencent 770B/49B-active MoE with Gated DeepSeek Sparse Attention and native MTP), Qwen3.8-Flash-Next (BF16/FP8/NVFP4, MTP), GraniteSWA, GraniteMoeSWA, NemotronH_Omni_Reasoning_V3 (MTP), and Kimi K3 NVFP4 checkpoints.
 - **[2026/09/07]** [SGLang v0.5.19 (Ubuntu)](https://gallery.ecr.aws/deep-learning-containers/sglang) — EC2: `0.5.19-gpu-py312-ec2` · SageMaker: `0.5.19-gpu-py312` · Qwen3.8, Ling-3.0, Spark2.5, Granite 4.2; beam search; DeepEP v2 MoE all-to-all.
 - **[2026/09/01]** [Ray LLM v1.0 (2.58.0, AL2023)](https://gallery.ecr.aws/deep-learning-containers/ray) — EC2/EKS: `serve-llm-cuda-v1.0` · Initial release: OpenAI-compatible LLM serving with Ray Serve and vLLM 0.26.0 on PyTorch 2.11.0 / CUDA 13.0.2 / Python 3.13; `ray[llm]`'s `build_openai_app` runs vLLM behind Ray Serve — single GPU on EC2, and multi-node serving on EKS via KubeRay.

docs/src/data/vllm-omni/0.28.0-gpu-ec2.yml

@@ -0,0 +1,12 @@
+framework: vLLM-Omni
+version: "0.28.0"
+ecr_repository: vllm
+accelerator: gpu
+python: py312
+cuda: cu130
+os: amzn2023
+platform: default
+public_registry: true
+
+tags:
+  - "omni-cuda-v1.7"

docs/src/data/vllm-omni/0.28.0-gpu-sagemaker.yml

@@ -0,0 +1,12 @@
+framework: vLLM-Omni
+version: "0.28.0"
+ecr_repository: vllm
+accelerator: gpu
+python: py312
+cuda: cu130
+os: amzn2023
+platform: sagemaker
+public_registry: true
+
+tags:
+  - "omni-sagemaker-cuda-v1.7"

docs/vllm-omni/changelog/index.md

@@ -4,6 +4,35 @@ Changelog for the Amazon Linux 2023-based vLLM-Omni images (`omni-cuda`, `omni-s
 
 * * *
 
+## v1.7.0 — 2026-09-21
+
+**Tags:** `omni-cuda-v1.7` · `omni-sagemaker-cuda-v1.7`
+
+**vLLM-Omni source:** [v0.28.0](https://github.com/vllm-project/vllm-omni/releases/tag/v0.28.0) (upstream vLLM v0.28.0) —
+[full upstream diff](https://github.com/vllm-project/vllm-omni/compare/v0.26.0...v0.28.0)
+
+**DLC PR:** [#6680](https://github.com/aws/deep-learning-containers/pull/6680)
+
+### Highlights
+
+- **PEFT LoRA validated on SageMaker.** A new endpoint test deploys an SD-3.5-medium image model with a baked-in PEFT LoRA adapter registered at
+  startup (`SM_VLLM_ENABLE_LORA`, `SM_VLLM_LORA_MODULES`, `SM_VLLM_MAX_LORA_RANK`) and confirms that per-request adapter selection via the `lora`
+  field on `/v1/images/generations` changes the generated image.
+
+### Changes
+
+- **Bundled-library bumps.** FlashInfer 0.6.14 → 0.6.16.post3; DeepEP repinned; NCCL pinned to 2.30.7 because DeepEPv2's GIN backend requires NCCL ≥
+  2.30.4.
+- **CUDA 13 mooncake wheel.** The KV-connector install now swaps `mooncake-transfer-engine` for its `-cuda13` variant to match the CUDA 13 base.
+- **`transformers` requirement** widened to `>= 5.10.1, < 5.15` (5.15 introduces a breaking change).
+
+### Notes
+
+- No CUDA change — still CUDA 13.0.2 and Python 3.12. This is a framework bump (v1.6 → v1.7); the EC2 and SageMaker images are rebuilt in lockstep to
+  keep the paired tags aligned.
+
+* * *
+
 ## v1.6.0 — 2026-08-25
 
 **Tags:** `omni-cuda-v1.6` · `omni-sagemaker-cuda-v1.6`

docs/vllm-omni/index.md

@@ -37,7 +37,7 @@ In addition to vLLM-Omni and its core stack (PyTorch, CUDA 13.0, NCCL, Python 3.
 - **[EFA](https://aws.amazon.com/hpc/efa/) and [OpenMPI](https://www.open-mpi.org/)** — high-throughput multi-node networking on supported instances
 - **espeak-ng and ffmpeg** — system-level dependencies for TTS phonemizer and audio/video encoding
 - **[s3tokenizer](https://github.com/xingchensong/S3Tokenizer)** — speech tokenizer required by CosyVoice3 (hard-imported at model load in vLLM-Omni
-  0.26.0)
+  0.28.0)
 
 The {{ sm_short }} image additionally includes a routing middleware that dispatches `/invocations` to omni-specific routes (TTS, image, video, etc.)
 via the `CustomAttributes` header. See [{{ sagemaker }} Deployment](deployment/sagemaker.md).