All news with #amazon sagemaker ai tag

🔧Amazon SageMaker HyperPod now supports AMI-based node lifecycle configuration for Slurm clusters, provisioning nodes with the software and configurations needed for production-ready AI/ML training environments. The AMI includes required components such as Docker, Enroot, and Pyxis, plus Slurm accounting, SSH key generation, log rotation, and user home setup. To enable it, omit the LifeCycleConfig block when creating clusters or select "None" under Lifecycle scripts in the console; you can still supply an extension script for additional customization or continue using full custom lifecycle scripts if you need complete control. This feature is available in all AWS Regions where SageMaker HyperPod is offered.

🔍 Amazon Quick now supports Amazon S3 table buckets as a direct data source, enabling dashboards, conversational analytics, and exploration of Apache Iceberg tables stored in S3 without intermediate warehouses or OLAP layers. Paired with Zero-ETL ingestion from systems like Salesforce, SAP, and Amazon Kinesis Data Firehose, organizations can access near real-time insights with reduced pipeline complexity. Admins configure S3 table bucket permissions once, and authors can immediately create datasets and use Dataset Q&A to query the lakehouse in natural language.

🤖 Amazon SageMaker AI introduces an agentic experience that dramatically shortens model customization from months to days or hours. Using SageMaker AI model customization agent skills, developers interact via natural language coding agents to prepare data, fine-tune models, evaluate quality with LLM-as-a-judge metrics, and generate reusable code artifacts. Skills can be installed into IDEs via the sagemaker-ai agent plugin or used pre-installed in SageMaker Studio Notebooks, and support deployment to Amazon Bedrock or SageMaker AI endpoints.

🚀 Amazon SageMaker AI endpoints now support prioritized instance pools for flexible provisioning. When your preferred instance type has insufficient capacity, SageMaker AI automatically provisions from the next option in your prioritized list for endpoint creation, updates, and autoscaling — keeping endpoints reliable without manual intervention. You can specify hardware-optimized model artifacts per instance type and monitor per-instance-type CloudWatch metrics for latency, throughput, GPU utilization, and instance counts.

🤖 AWS has added Google DeepMind's instruction‑tuned Gemma 4 E4B, Gemma 4 26B‑A4B, and Gemma 4 31B to SageMaker JumpStart, making multimodal foundation models directly accessible to AWS customers. The models offer configurable step‑by‑step reasoning, interleaved text and image inputs, video and image understanding, native function calling, and multilingual support across 140+ languages. Gemma 4 E4B also supports audio input for ASR and speech‑to‑translated‑text workflows. Customers can deploy these models via SageMaker Studio or the SageMaker Python SDK for rapid experimentation and production.

🆕 Amazon SageMaker JumpStart now includes paraphrase-multilingual-MiniLM-L12-v2, Microsoft Table Transformer Detection, and Bielik-11B-v3.0-Instruct. The MiniLM model maps sentences to 384-dimensional dense vectors across 50+ languages for cross-lingual semantic search, multilingual clustering, and sentence similarity scoring. The Microsoft Table Transformer is a DETR-based detector trained on PubTables-1M to locate tables in PDFs and scanned images for document digitization. Bielik-11B offers an 11B-parameter multilingual generative model focused on Polish and 32 European languages for dialogue, STEM reasoning, and enterprise NLP.

🚀 Amazon SageMaker HyperPod now supports G7e and r5d.16xlarge instances to improve large-model development, training, and deployment at scale. G7e uses NVIDIA RTX PRO 6000 Blackwell GPUs, offering up to 2.3x better inference performance than G6e and up to 768 GB GPU memory for larger LLMs, multimodal, and agentic AI workloads. The r5d.16xlarge provides 64 vCPUs, 512 GB RAM, and NVMe storage for distributed preprocessing, feature engineering, and memory-heavy orchestration; G7e is available in select US and Asia Pacific regions while r5d.16xlarge is available across all HyperPod regions.

📊 Amazon SageMaker Training Plans now publishes Amazon CloudWatch metrics to track utilization of capacity reservations tied to purchased Flexible Training Plans. Administrators gain both historical and real‑time views of instance usage at the individual plan level and across an account, enabling informed decisions about capacity allocation and cost optimization. This observability helps teams align compute consumption with AI budgets and timelines while reducing wasted reserved capacity.

🔒 Amazon SageMaker Unified Studio now runs notebook kernels inside the domain-configured Amazon VPC, providing network isolation for interactive ML and data workloads. Kernels inherit VPC settings, subnets, and security groups defined at the domain level, enabling centralized network policy and secure access to private databases, internal APIs, and non-public data sources. This VPC configuration applies to the interactive compute where Python code and dataframes execute; other compute engines have separate VPC considerations. VPC-enabled kernels are available in all Regions where SageMaker Unified Studio is supported.

🧑‍💻 Amazon SageMaker Unified Studio now lets data workers create and manage multiple code spaces within a single project for IAM domains. Each space maintains its own persistent Amazon EBS volume and independent compute and storage settings, and can be paused, resumed, or connected to a local IDE while preserving files and session state. This enables parallel workstreams and isolated experiments with tailored runtimes and is available in all Regions where SageMaker Unified Studio is offered.

🧩 Amazon SageMaker AI now supports serverless model customization for Qwen3.5, enabling supervised fine-tuning (SFT) and reinforcement fine-tuning (RFT) of 4B, 9B, and 27B parameter models. With serverless customization, SageMaker handles infrastructure provisioning and training orchestration so teams can focus on data, evaluation, and domain adaptation while paying only for consumed resources. This capability is available in US East (N. Virginia), US West (Oregon), Asia Pacific (Tokyo), and EU (Ireland) and can be launched from SageMaker Studio or the SageMaker Python SDK.

🔧 Amazon SageMaker AI now provides inference recommendations that automate optimization and benchmarking to deliver validated, deployment-ready configurations. Customers supply their own generative models, define expected traffic patterns, and set a performance objective — optimize for cost, minimize latency, or maximize throughput. SageMaker analyzes model architecture, benchmarks across multiple instance types using NVIDIA AIPerf, and returns metrics such as time to first token, inter-token latency, request latency percentiles, throughput, and cost projections. The capability is available today in seven AWS Regions.

🆕 Amazon SageMaker HyperPod now supports flexible instance groups, allowing multiple instance types and multiple subnets within a single instance group. Using a new InstanceRequirements parameter, HyperPod provisions the highest-priority instance type first and automatically falls back to lower-priority types when capacity is unavailable. The feature integrates with Karpenter autoscaling and can be created via the CreateCluster/UpdateCluster APIs, AWS CLI, or the Management Console.

🔍 Amazon SageMaker HyperPod now supports on-demand deep health checks for Amazon EKS and Slurm-orchestrated clusters. Administrators can run comprehensive GPU stress and connectivity tests on entire instance groups or specific instances, with progress and results visible at both group and instance levels via the SageMaker console and APIs. Instances under test are isolated from scheduling and are returned to service upon passing or, when paired with automatic node recovery, rebooted or replaced if they fail.

🚀 SageMaker JumpStart now offers optimized deployments for foundation models, providing pre-configured, task-aware settings tailored to specific use cases and performance goals. Customers can choose cost-, throughput-, latency-optimized, or balanced configurations and preview P50 latency, time-to-first-token, and throughput metrics before deployment. Supported models include variants from Meta, Microsoft, Mistral AI, Qwen, Google, and TII, and deployments target SageMaker AI Managed Inference endpoints or HyperPod clusters. The feature leverages VPC deployment capabilities and is available in all regions where JumpStart is supported.

🚀 Amazon SageMaker JumpStart now includes NVIDIA’s Nemotron-3-Super-120B and the Qwen3.5 family (9B and 27B), giving customers turnkey access to foundation models optimized for agentic reasoning, multilingual coding, and advanced instruction following. Nemotron-3-Super-120B employs a hybrid LatentMixture-of-Experts architecture with Mamba-2 and MoE layers to support collaborative agents and high-volume automation such as IT ticket triage and cybersecurity workflows. The Qwen3.5-9B prioritizes efficiency for resource-constrained environments, while Qwen3.5-27B offers deeper contextual and multimodal reasoning for large-scale document processing and complex scenarios. Users can deploy these models directly from the JumpStart catalog or programmatically via the SageMaker Python SDK.

✅ Amazon SageMaker HyperPod task governance now supports gang scheduling for HyperPod clusters using the EKS orchestrator. Administrators can configure readiness timeouts, node-failure behavior, single-workload admission and retry policies so distributed training jobs are only admitted when all required pods are ready. Pulled workloads are automatically requeued to avoid stalls and wasted compute. This reduces deadlocks, resource contention, and unexpected cost overruns for multi-pod training jobs.

⚙️ Amazon SageMaker Unified Studio now supports Serverless Workflows in Identity Center domains, allowing customers to orchestrate data-processing tasks with Apache Airflow (via Managed Workflows for Apache Airflow) without provisioning Airflow infrastructure. Serverless Workflows auto-provision compute during runs and release it afterward, so you pay only for actual run time. Each workflow runs with its own execution role and isolated worker to ensure workflow-level security and prevent cross-workflow interference. The Visual Workflow experience supports around 200 operators and built-in integrations with services such as Amazon S3, Amazon Redshift, Amazon EMR, AWS Glue, and Amazon SageMaker AI.

📝 Amazon has added import/export capabilities to SageMaker Unified Studio notebooks to simplify migration from JupyterLab and other platforms. The feature supports .ipynb, .json, and .py formats while preserving cell types, outputs, execution history, and metadata. Exports are available in four package types (.zip with requirements, .ipynb, .py, and native .json). The release also introduces developer productivity features including cell reordering, keyboard shortcuts, cell renaming, and multi-line SQL with tabbed results.

📊 Amazon SageMaker Data Agent now embeds interactive charting, SQL analytics across Snowflake sources, and materialized view management directly inside SageMaker Unified Studio notebooks. You can ask natural-language prompts like "plot monthly revenue trends by region for 2025" to generate interactive charts that support hover, editing, and refinement without writing code. When analyses span AWS and Snowflake, the agent lets you join Snowflake tables via external connections with AWS Glue Data Catalog data in a single prompt. The agent can also recommend and create materialized views, including refresh schedules, to optimize query performance.