aws · sirutBuasai · Jul 8, 2026 · Jul 7, 2026 · Jul 7, 2026
diff --git a/README.md b/README.md
@@ -28,6 +28,7 @@ ______________________________________________________________________
 
 ### 🚀 Release Highlights
 
+- **[2026/07/06]** [SGLang Server v1.2 (AL2023)](https://gallery.ecr.aws/deep-learning-containers/sglang) — EC2: `server-cuda-v1.2` · SageMaker: `server-sagemaker-cuda-v1.2` · SGLang `0.5.14`; adds support for the NVIDIA LocateAnything-3B vision-grounding model and upgrades sgl-kernel 0.4.4, FlashInfer 0.6.12, Mooncake 0.3.11.post1, NCCL 2.30.4, and EFA 1.49.0.
 - **[2026/07/02]** [vLLM Server v2.1 (AL2023)](https://gallery.ecr.aws/deep-learning-containers/vllm) — EC2: `server-cuda-v2.1` · SageMaker: `server-sagemaker-cuda-v2.1` · vLLM `0.24.0`; adds JetBrains Mellum2-12B-A2.5B-Thinking (`MellumForCausalLM`); FlashInfer 0.6.12; drops the `transformers<5.10` pin (now requires transformers ≥ 5.5.3).
 - **[2026/07/02]** [vLLM-Omni v1.4 (AL2023)](https://gallery.ecr.aws/deep-learning-containers/vllm) — EC2: `omni-cuda-v1.4` · SageMaker: `omni-sagemaker-cuda-v1.4` · SageMaker `/v1/videos` and `/v1/videos/sync` accept `application/json` again (JSON→multipart conversion restored); no framework bump (vLLM-Omni 0.21.0rc1).
 - **[2026/07/02]** [PyTorch v2.12.1](https://gallery.ecr.aws/deep-learning-containers/pytorch) — EC2: `2.12.1-cu130-amzn2023` · SageMaker: `2.12.1-cu130-amzn2023-sagemaker` · Amazon Linux 2023 with EFA, flash-attn, and Transformer Engine; PyTorch 2.12.1 bug-fix release (B200/Triton correctness fixes).

diff --git a/docs/sglang/changelog/index.md b/docs/sglang/changelog/index.md
@@ -4,6 +4,30 @@ Changelog for the Amazon Linux 2023-based SGLang images (`server-cuda`, `server-
 
 * * *
 
+## v1.2.0 — 2026-07-06
+
+**Tags:** `server-cuda-v1.2` · `server-sagemaker-cuda-v1.2`
+
+**SGLang source:** [bc8b3ab](https://github.com/sgl-project/sglang/commit/bc8b3ab1f55c60951b586d84d4aec773a6f654df) (`0.5.14+amzn2023.bc8b3ab`)
+
+**Bundled versions:** CUDA 13.0.3 · Python 3.12 · PyTorch 2.11.0 · sgl-kernel 0.4.4 · FlashInfer 0.6.12 · Mooncake 0.3.11.post1 · NCCL 2.30.4
+
+### Highlights
+
+- Bumped SGLang to `0.5.14` (upstream commit `bc8b3ab`)
+- Added support for **[NVIDIA LocateAnything-3B](https://huggingface.co/nvidia/LocateAnything-3B)** — a multimodal vision-grounding model that returns
+  bounding boxes (`<box>…</box>`) for objects matching a text description. Bundled the `decord`, `lmdb`, and `peft` runtime dependencies required by
+  the model's custom Hugging Face processors
+- Upgraded EFA to `1.49.0`
+- Upgraded stack: sgl-kernel `0.4.4`, FlashInfer `0.6.12`, Mooncake `0.3.11.post1`, NCCL `2.30.4`, gdrcopy `2.6`, Rust `1.96.1`
+
+### Security
+
+- Allowlisted CVE-2026-27145 (Go stdlib `x509` `VerifyHostname` CPU exhaustion) — embedded in Mooncake's `libetcd_wrapper.so`; cannot be patched
+  without an upstream Mooncake rebuild with Go 1.26.4+
+
+* * *
+
 ## v1.1.0 — 2026-06-29
 
 **Tags:** `server-cuda-v1.1` · `server-sagemaker-cuda-v1.1`

diff --git a/docs/sglang/configuration.md b/docs/sglang/configuration.md
@@ -27,6 +27,9 @@ docker run --gpus all -p 30000:30000 \
 
 For gated models (Llama, Gemma, etc.), pass `-e HF_TOKEN=<your_hf_token>`.
 
+For multimodal vision-grounding models (e.g. LocateAnything-3B), pass `--trust-remote-code` and send image inputs via the OpenAI chat `image_url`
+content type — see the [EC2 vision-grounding example](deployment/ec2.md#vision-grounding-multimodal).
+
 ## Amazon SageMaker AI (`server-sagemaker-cuda`)
 
 The SageMaker image serves on **port 8080** and accepts SGLang flags via `SM_SGLANG_*` environment variables. Each variable is converted to the

diff --git a/docs/sglang/deployment/ec2.md b/docs/sglang/deployment/ec2.md
@@ -41,6 +41,42 @@ docker run --gpus all --ipc=host -p 30000:30000 \
 
 `--ipc=host` enables shared memory between GPU processes.
 
+## Vision Grounding (Multimodal)
+
+The images support multimodal vision-grounding models such as [NVIDIA LocateAnything-3B](https://huggingface.co/nvidia/LocateAnything-3B), which
+return bounding boxes for objects matching a text description. These models ship a custom Hugging Face processor, so pass `--trust-remote-code`:
+
+```bash
+docker run --gpus all -p 30000:30000 \
+  public.ecr.aws/deep-learning-containers/sglang:server-cuda \
+  --model-path nvidia/LocateAnything-3B \
+  --trust-remote-code \
+  --context-length 16384 \
+  --host 0.0.0.0 --port 30000
+```
+
+Send an image with a grounding prompt using the OpenAI chat format with an `image_url` content part:
+
+```bash
+curl http://localhost:30000/v1/chat/completions \
+  -H "Content-Type: application/json" \
+  -d '{
+    "model": "nvidia/LocateAnything-3B",
+    "messages": [{
+      "role": "user",
+      "content": [
+        {"type": "image_url", "image_url": {"url": "https://example.com/car.jpg"}},
+        {"type": "text", "text": "Locate all the instances that matches the following description: car"}
+      ]
+    }],
+    "max_tokens": 2048,
+    "skip_special_tokens": false
+  }'
+```
+
+The response content contains `<box>x1, y1, x2, y2</box>` coordinates for each detected instance. Pass `"skip_special_tokens": false` so the box
+tokens are preserved in the output.
+
 ## Model-Specific Tuning
 
 For recommended serving flags, hardware configurations, and quantization options per model, see the

diff --git a/docs/sglang/index.md b/docs/sglang/index.md
@@ -29,6 +29,10 @@ In addition to SGLang and its core stack (PyTorch 2.11, CUDA 13.0, NCCL, Python
 
 The images are built from SGLang source against the H100 (sm_90) and Blackwell (sm_100, sm_103) CUDA architectures.
 
+The runtime also bundles `decord`, `lmdb`, and `peft` to support multimodal vision-grounding models such as
+[NVIDIA LocateAnything-3B](https://huggingface.co/nvidia/LocateAnything-3B), which return bounding boxes for objects matching a text prompt. See
+[Supported Models](models/index.md) for the tested set.
+
 ## API Endpoints
 
 The container runs SGLang's [OpenAI-compatible API server](https://docs.sglang.ai/basic_usage/openai_api.html). Common endpoints:

diff --git a/docs/sglang/models/index.md b/docs/sglang/models/index.md
@@ -18,6 +18,11 @@ release. A *Smoke + Benchmark* tag means both apply.
 |  | [Qwen/Qwen3-Coder-Next-FP8](https://huggingface.co/Qwen/Qwen3-Coder-Next-FP8) | Benchmark |
 | **GPT-OSS** | [openai/gpt-oss-20b](https://huggingface.co/openai/gpt-oss-20b) | Benchmark |
 | **DeepSeek** | [deepseek-ai/DeepSeek-V4-Flash](https://huggingface.co/deepseek-ai) | Benchmark |
+| **LocateAnything** | [nvidia/LocateAnything-3B](https://huggingface.co/nvidia/LocateAnything-3B) | Smoke |
+
+**LocateAnything-3B** is a multimodal vision-grounding model: given an image and a text description, it returns bounding boxes (`<box>…</box>`) for
+the matching objects. Serve it with `--trust-remote-code` and send image inputs via the OpenAI chat `image_url` content type — see the
+[EC2 vision-grounding example](../deployment/ec2.md#vision-grounding-multimodal).
 
 ## Custom Models