Skip to content

Repository files navigation

AWS Logo

AWS Deep Learning Containers

One stop shop for running AI/ML on AWS

Docs · Available Images · Tutorials

Auto Release - PyTorch 2.13 Auto Release - TensorFlow Training 2.21 Auto Release - TensorFlow Inference 2.20 Auto Release - vLLM Auto Release - vLLM-Omni Auto Release - SGLang Auto Release - Ray Auto Release - Base cu130 Auto Release - Base cu132


About

AWS Deep Learning Containers (DLCs) are pre-built Docker images for running AI/ML workloads on AWS. Each image is tested and patched for security vulnerabilities. For more details, visit our documentation.


🔥 What's New

🚀 Release Highlights

  • [2026/08/26] vLLM v0.28.0 (Ubuntu) — EC2: 0.28.0-gpu-py312-ec2 · SageMaker: 0.28.0-gpu-py312 · Kimi-K3 stack-wide optimization (Decode Context Parallel, fused FlashKDA kernels, GEMM-RS sequence parallelism); DeepSeek V4 sparse MLA end-to-end with MTP and DSpark speculative decoding; new models Muse Glimmer, Ling 3.0 Flash, Dots3, Interns2mobius; tiered KV cache offloading to disk; runtime base moves to Ubuntu 24.04 and Transformers 5.15.0; max_num_batched_tokens default 8192 -> 16384.
  • [2026/08/25] vLLM-Omni v1.6 (AL2023) — EC2: omni-cuda-v1.6 · SageMaker: omni-sagemaker-cuda-v1.6 · SageMaker SM_VLLM_* fix — JSON-array env vars (e.g. SM_VLLM_LORA_MODULES) now expand into multiple argv values so multi-value flags parse correctly; no framework bump (still vLLM-Omni 0.26.0).
  • [2026/08/25] vLLM Server v2.4 (AL2023) — EC2: server-cuda-v2.4 · SageMaker: server-sagemaker-cuda-v2.4 · vLLM 0.27.1 (up from 0.27.0); Muse Glimmer model support; SageMaker SM_VLLM_* fix — JSON-array env vars (e.g. SM_VLLM_LORA_MODULES) now expand into multiple argv values so multi-value flags parse correctly.
  • [2026/08/25] SGLang v0.5.18 (Ubuntu) — EC2: 0.5.18-gpu-py312-ec2 · SageMaker: 0.5.18-gpu-py312 · Muse Glimmer and Intern-S2-Mobius, plus diffusion additions SANA-Video, LTX-2.5, Cosmos3 Edge, and LongCat-Image; overlapped checkpoint staging for faster startup (--startup-weight-load-mode overlap); FlashInfer MNNVL standalone allreduce on by default for DeepSeek-V3/V3.2/V4; upstream stack moves to torch 2.13.0 and triton 3.7.1.
  • [2026/08/21] llama.cpp v1.0 (b10433, AL2023) — EC2: server-cpu-v1 · server-cuda-v1 · Graviton: llama-cpp-arm64:server-cpu-v1 · SageMaker: server-sagemaker-cpu-v1 · server-sagemaker-cuda-v1 · llama-cpp-arm64:server-sagemaker-cpu-v1 · Initial release: serve quantized GGUF models with the upstream llama-server OpenAI-compatible API on x86 CPU, NVIDIA GPU (CUDA 13.0.2), and AWS Graviton (ARM64); Python 3.12.
  • [2026/08/20] TensorFlow Serving v2.20.0 (AL2023) — SageMaker: 2.20.0-cpu-py312-amzn2023-sagemaker · 2.20.0-gpu-py312-cu129-amzn2023-sagemaker · TensorFlow Serving 2.20.0 on Amazon Linux 2023 with Python 3.12 and CUDA 12.9.
  • [2026/08/17] SGLang Server v1.3 (AL2023) — EC2: server-cuda-v1.3 · SageMaker: server-sagemaker-cuda-v1.3 · SGLang 0.5.17 (up from 0.5.14); Kimi-K3 (2.8T MoE, MXFP4) support; sgl-kernel 0.4.5, FlashInfer 0.6.15.post1, Mooncake 0.3.12.post1.
  • [2026/08/17] vLLM Server v2.3 (AL2023) — EC2: server-cuda-v2.3 · SageMaker: server-sagemaker-cuda-v2.3 · vLLM 0.27.0 (up from 0.26.0); Kimi K3 (native support + kernels, Rust/Python frontends); FlashInfer 0.6.16.post3; NVIDIA B300 (SM103); new models K-EXAONE-2.0-750B-A37B, jina-embeddings-v5-text-nano, Qwen3.5; dynamic FP8 for Inkling; Baidu Unlimited-OCR smoke test.
  • [2026/08/14] vLLM v0.27.1 (Ubuntu) — EC2: 0.27.1-gpu-py312-ec2 · SageMaker: 0.27.1-gpu-py312 · Kimi K3, Qwen3.5 dense + MoE (EVS video token pruning), K-EXAONE-2.0-750B-A37B, VaultGemma, jina-embeddings-v5-text-nano.
  • [2026/08/14] WhisperX v3.8.6 (AL2023) — EC2: 3.8.6-cu128-amzn2023 · SageMaker: 3.8.6-cu128-amzn2023-sagemaker · Initial release: speech transcription with word-level alignment (wav2vec2) and speaker diarization (pyannote) through an OpenAI-compatible API on CUDA 12.8 / Python 3.12; real-time and asynchronous SageMaker endpoints.
  • [2026/08/12] Ray v1.4 (2.57.0, AL2023) — EC2: serve-ml-cuda-v1.4 · serve-ml-cpu-v1.4 · SageMaker: serve-ml-sagemaker-cuda-v1.4 · serve-ml-sagemaker-cpu-v1.4 · Ray 2.57.0 (up from 2.56.1).
  • [2026/08/08] SGLang v0.5.17 (Ubuntu) — EC2: 0.5.17-gpu-py312-ec2 · SageMaker: 0.5.17-gpu-py312 · Kimi K3, MiniMax H3.
  • [2026/08/07] vLLM-Omni v1.5 (AL2023) — EC2: omni-cuda-v1.5 · SageMaker: omni-sagemaker-cuda-v1.5 · vLLM-Omni 0.26.0 (up from 0.21.0rc1) on vLLM v0.26.0 with the new Rust frontend; SageMaker bidirectional WebSocket streaming (InvokeEndpointWithBidirectionalStream) for low-latency TTS and realtime sessions; FlashInfer 0.6.14; s3tokenizer bundled for CosyVoice3.

📢 Support Updates

  • [2026/04/28] We cannot guarantee security patching on Ubuntu-based vLLM and SGLang images due to the lack of Ubuntu Pro licensing. Customers may continue using these images at their own discretion and risk. We recommend migrating to our Amazon Linux-based images.
  • [2026/02/10] Extended support for PyTorch 2.6 Inference containers until June 30, 2026
    • PyTorch 2.6 Inference images will continue to receive security patches and updates through end of June 2026
    • For complete framework support timelines, see our Support Policy

📝 Blog Posts

🎓 Workshop


License

This project is licensed under the Apache-2.0 License.

About

One stop shop for running AI/ML on AWS.

Topics

Resources

Code of conduct

Contributing

Security policy

Stars

1.2k stars

Watchers

49 watching

Forks

Used by

Contributors

Languages