AWS published an engineering guide for scaling MoE reinforcement-learning post-training (RLHF/GRPO) on Amazon EKS, combining EFA networking, DeepEP expert-parallel communication, and EC2 Spot instances for rollout generation. Amazon's libfabric contributions to DeepEP enable native EFA support, delivering 40% higher aggregate rollout throughput on 48 P5en instances. The work targets the communication bottleneck that increasingly dominates sparse MoE training as expert parallelism spans more nodes.
AWS introduced the ability to create and manage Amazon SageMaker Spaces on SageMaker HyperPod EKS clusters directly from the SageMaker Studio UI. Data scientists can launch JupyterLab and Code Editor environments via a new IDE and Notebooks tab, replacing the previous CLI/kubectl-only workflow. The Spaces add-on itself remains free; users pay for underlying cluster compute and SSH-over-SSM connectivity.
AWS introduced the ability to create and manage Amazon SageMaker Spaces on SageMaker HyperPod EKS clusters directly from the SageMaker Studio UI. Data scientists can launch JupyterLab and Code Editor environments via a new IDE and Notebooks tab, replacing the previous CLI/kubectl-only workflow. The Spaces add-on itself remains free; users pay for underlying cluster compute and SSH-over-SSM connectivity.
AWS published a walkthrough showing how to run the open-source SkyRL RL framework on SageMaker HyperPod to post-train Qwen3-VL-8B for multi-turn visual maze navigation. GRPO training raises maze solve rate from 43.75% to over 95% on a fixed 64-maze eval set. It matters to practitioners because it shows a reproducible recipe for multi-node RL post-training with colocation, checkpointing, and managed monitoring.
AWS published a walkthrough showing how to run the open-source SkyRL RL framework on SageMaker HyperPod to post-train Qwen3-VL-8B for multi-turn visual maze navigation. GRPO training raises maze solve rate from 43.75% to over 95% on a fixed 64-maze eval set. It matters to practitioners because it shows a reproducible recipe for multi-node RL post-training with colocation, checkpointing, and managed monitoring.
A new paper introduces Best-Practice Critic Optimization (BPCO), a recipe for training a critic in reinforcement learning for large language models. It combines DPPO, bounded value predictions, Monte Carlo targets, unnormalized advantages, and length-adaptive GAE to achieve stability and efficiency, potentially replacing GRPO's multi-sample approach.
A new paper introduces Best-Practice Critic Optimization (BPCO), a recipe for training a critic in reinforcement learning for large language models. It combines DPPO, bounded value predictions, Monte Carlo targets, unnormalized advantages, and length-adaptive GAE to achieve stability and efficiency, potentially replacing GRPO's multi-sample approach.
An AWS engineering post details integrating the NVIDIA Resiliency Extension (NVRx) with PyTorch FSDP training on Amazon EKS, combining async checkpointing with two fault-recovery layers: in-process restart and ft_launcher in-job restart. Benchmarks on H100 clusters show async checkpointing sustains 99%+ training efficiency versus 57-61% for synchronous saves, while recovery layers restore training in seconds instead of full job restarts.
An AWS engineering post details integrating the NVIDIA Resiliency Extension (NVRx) with PyTorch FSDP training on Amazon EKS, combining async checkpointing with two fault-recovery layers: in-process restart and ft_launcher in-job restart. Benchmarks on H100 clusters show async checkpointing sustains 99%+ training efficiency versus 57-61% for synchronous saves, while recovery layers restore training in seconds instead of full job restarts.
AWS published a walkthrough building an AI product tagging system using Amazon SageMaker serverless model customization. It fine-tunes Qwen3-8B with supervised fine-tuning (SFT) and then reinforcement learning with verifiable rewards (RLVR) via GRPO on an Amazon sales dataset, raising the overall tagging score from 0.354 to 0.6941. The approach targets high-volume retail catalog enrichment where a stable schema and programmatically scorable outputs favor a small custom model over a frontier LLM.
AWS published a walkthrough building an AI product tagging system using Amazon SageMaker serverless model customization. It fine-tunes Qwen3-8B with supervised fine-tuning (SFT) and then reinforcement learning with verifiable rewards (RLVR) via GRPO on an Amazon sales dataset, raising the overall tagging score from 0.354 to 0.6941. The approach targets high-volume retail catalog enrichment where a stable schema and programmatically scorable outputs favor a small custom model over a frontier LLM.