Skip to content

LLM Serving using SGLang DLC

Production-ready Docker images for serving large language models with SGLang on AWS. Built on Amazon Linux 2023 with ongoing security patching.

Images

Platform Image Default Port
EC2 / EKS public.ecr.aws/deep-learning-containers/sglang:server-cuda 30000
Amazon SageMaker AI public.ecr.aws/deep-learning-containers/sglang:server-sagemaker-cuda 8080

All images are also available on the ECR Public Gallery. For private ECR URIs, see Image Access.

What's Included

In addition to SGLang and its core stack (PyTorch 2.11, CUDA 13.0, NCCL, Python 3.12), the images bundle:

  • FlashInfer — fused attention kernels with precompiled cubins for fast cold start
  • DeepEP — expert-parallel kernels for large MoE models (DeepSeek, Qwen MoE)
  • Mooncake — KV-cache transfer engine for disaggregated prefill/decode
  • sgl-kernel — SGLang's custom CUDA kernels, built from source for the bundled CUDA arch list
  • EFA and OpenMPI — high-throughput multi-node networking on supported instances

The images are built from SGLang source against the H100 (sm_90) and Blackwell (sm_100, sm_103) CUDA architectures.

API Endpoints

The container runs SGLang's OpenAI-compatible API server. Common endpoints:

Endpoint Purpose
POST /v1/chat/completions Chat-style generation
POST /v1/completions Legacy text completion
POST /v1/embeddings Generate embeddings (embedding models)
POST /generate SGLang-native generation API
GET /v1/models List loaded model(s)
GET /get_model_info Model metadata
GET /health, /health_generate Liveness probe
POST /flush_cache Flush the radix attention cache
GET /metrics Prometheus metrics

Refer to SGLang's API documentation for request/response schemas and the full endpoint list.

How We Build

These images are curated builds, not simple repackages of upstream releases:

  • Built from upstream source — images build SGLang from a pinned upstream commit, each gated by our regression test suite before publication.
  • Regression-tested — validated against a suite of models on every release. See Supported Models for the full list.
  • Security-patched — continuously maintained with security patches from AWS on an Amazon Linux 2023 base.