Skip to content

Changelog

Changelog for the Amazon Linux 2023-based PyTorch images (2.12-cu130-amzn2023, 2.12-cpu-amzn2023, and the corresponding *-sagemaker variants).

PyTorch 2.12 — 2026-07-02

Tags: 2.12-cu130-amzn2023 · 2.12-cpu-amzn2023 · 2.12-cu130-amzn2023-sagemaker · 2.12-cpu-amzn2023-sagemaker

Bundled versions: PyTorch 2.12.1 · torchvision 0.27.1 · torchaudio 2.11.0 · CUDA 13.0.2 · Python 3.12 · NCCL 2.26.2 · EFA 1.47.0 · GDRCopy 2.4.4 · flash-attn 2.8.3 · Transformer Engine 2.12.0 · DeepSpeed 0.18.8

Highlights

  • Bumped PyTorch to 2.12.1, with torchvision 0.27.1
  • 2.12.1 fixes a Triton illegal-memory-access in the convolution2d_bwd_weight kernel on B100/B200 (sm100) GPUs (pytorch#187081)

PyTorch 2.11 — 2026-04-30

Tags: 2.11-cu130-amzn2023 · 2.11-cpu-amzn2023 · 2.11-cu130-amzn2023-sagemaker · 2.11-cpu-amzn2023-sagemaker

Highlights

  • Initial release of PyTorch DLC images on Amazon Linux 2023
  • PyTorch 2.11.0 (with torchvision 0.26.0 and torchaudio 2.11.0)
  • CUDA 13.0.2, Python 3.12, NCCL 2.26.2 (GPU variants)
  • EFA 1.47.0 with the AWS NCCL OFI plugin and GDRCopy 2.4.4 for multi-node training
  • flash-attn 2.8.3 and Transformer Engine 2.12.0 for fused attention and FP8 training
  • DeepSpeed 0.18.8 for memory-efficient large-model training
  • NCCL all_reduce_perf binary at /usr/local/bin/all_reduce_perf for verifying EFA connectivity
  • Pre-configured OpenSSH server (port 22) for inter-node MPI/torchrun launches
  • SageMaker variants include the sagemaker-pytorch-training toolkit, MLflow, SHAP, smclarify, and SageMaker-specific data libraries