TorchNeuron Brings Native PyTorch to AWS Trainium
AWS released TorchNeuron in public beta alongside Neuron 2.33. It is an open source PyTorch backend that runs models on Trainium by switching the device to Neuron, without rewriting model code. Eager execution and torch.compile both keep working, and it integrates with the Neuron Kernel Interface for custom operators.
Supported hardware is Trainium2 and Trainium3. Previously, moving a PyTorch workload to Trainium meant adapting it to the Neuron-specific tracing workflow; a native backend makes Trainium much closer to a drop-in target for teams already training on GPUs, which matters for anyone evaluating Trainium's pricing for fine-tuning or post-training jobs.
Read more — AWS News Blog
SageMaker AI Training Sessions and Structured Outputs for Claude on Bedrock
Amazon SageMaker AI training sessions are now generally available. They let you own the post-training loop (custom loss functions and reward signals) while SageMaker AI handles GPU allocation, distributed execution, fault recovery and checkpoint storage. Billing is per token processed, and charges pause while your reward function scores a batch, which makes reinforcement-style post-training with slow external graders cheaper. Muse Glimmer, Qwen 3.8 and Ministral 3 were added as supported base models.
Amazon Bedrock now supports structured outputs for all Anthropic Claude models, returning JSON that conforms to a supplied schema on both the bedrock-runtime and bedrock-mantle endpoints. That removes the custom parsing and retry logic teams typically wrap around extraction or tool-style calls.
Read more — AWS News Blog
ElastiCache Serverless for Valkey Adds Zero-Downtime Online Migration; Polly Generative Voices Expand
Amazon ElastiCache Serverless for Valkey can now migrate self-managed Valkey or Redis OSS running on EC2 with no client-facing errors. The workflow is prepare, start data sync, then cut over manually or automatically, without third-party replication tools. It is available in all commercial Regions, the China Regions and GovCloud (US), which removes one of the main blockers for teams wanting to move off self-managed Redis clusters.
Amazon Polly's Generative engine gained 14 voices across 9 locales, bringing the catalog to 57 generative voices in 11 languages and 25 locales. Hindi and Finnish are generally available for the first time, and the upgrade is drop-in: voice names, the SynthesizeSpeech call and pricing are unchanged.
Read more — AWS News Blog
Gemini Enterprise Agent Platform: Computer Use Tool and Inference Metadata Logging Go GA
On October 30, Google made the computer use tool generally available on Gemini Enterprise Agent Platform. It ships with human-in-the-loop safety policies: developers can configure or override specific confirmation policies, for example requiring approval before financial transactions or data modification, and administrators can enforce organization-wide guardrails through the configurableComputerUseToolSafetyPolicies organization policy constraint.
Inference request metadata logging also reached GA. It writes one Cloud Logging entry per request with model, region, consumption option, token counts, Provisioned Throughput usage, latency and error code, but never prompts or responses. It covers Gemini models via generateContent and streamGenerateContent and Claude models via rawPredict and streamRawPredict. Requests can be tagged with headers prefixed x-goog-agent-platform-observability-label- and queried with SQL in Observability Analytics, which gives per-team or per-feature cost attribution without logging sensitive content.
Read more — Google Cloud
Managed OpenTelemetry for GKE Is Generally Available
Managed OpenTelemetry for GKE reached GA on October 29 for clusters on 1.37.1-gke.1697000 or later. It replaces a self-managed OpenTelemetry Collector with a Google-managed one that accepts OTLP traces, metrics and logs over gRPC on port 4317 as well as HTTP on 4318, configured through the otlp_grpc setting of the Instrumentation custom resource. Logs now flow through the Cloud Telemetry API alongside traces and metrics, with improved per-signal queueing and autoscaling for high-throughput clusters.
Also GA: multimodal prompt and response collection for LangGraph ReAct and Agent Development Kit agents, the --managed-otel-scope gcloud flag (gcloud 576.0.0+), and managed_opentelemetry_config in the Terraform Google provider (v8.8.0+). Clusters on Preview versions need upgrading to pick up the GA collector.
Read more — Google Cloud
Cloud Run Jobs Can Run GPUs for Up to 12 Hours; AlloyDB Scales Vector Indexes to 10 Billion Rows
Cloud Run jobs can now run with GPUs attached for up to 12 hours, in Preview as of October 30. That makes serverless jobs viable for longer batch inference, embedding generation and fine-tuning runs that previously needed GKE or Compute Engine.
On the data side, AlloyDB for PostgreSQL made four-level ScaNN vector indexes generally available on October 28, supporting up to 10 billion rows, and added ai.extract and ai.classify functions for in-query text processing in Preview. BigQuery made Git repositories for SQL scripts and notebooks and the AI.COUNT_TOKENS function GA, and Spanner now supports equality and ordering comparisons on JSON values with search-index acceleration.
Read more — Google Cloud
Links & Sources
You May Also Enjoy
