A Ray-Focused Guide to PyTorch Conference North America

PyTorch Conference North America 2026 will take place in San Jose in a few weeks, and the program includes a cluster of sessions centered on Ray, the distributed compute framework. Ray co-creator Ion Stoica, of Anyscale, Databricks and UC Berkeley, will give a keynote titled Evolving Ray and Kubernetes Together for the AI Era on Tuesday at 9:35 AM in the Grand Ballroom.

A related session, Crossing the Divide: Co-Evolving Kubernetes and Ray for the AI Era, features Jago Macleod of Google with Stoica on Tuesday at 2:50 PM in Room 210BF. It addresses the split between the CNCF, which owns Kubernetes, Envoy, OpenTelemetry, llm-d and containerd, and the PyTorch Foundation, which contains PyTorch, vLLM and Ray, and discusses collaboration between Google, Anyscale and the broader community on an open, vertically integrated stack.

Production talks cover several companies. LinkedIn's Ray All the Way Down, Tuesday at 3:25 PM, describes a three-layer Ray-based training stack: a Ray-actor-based data loading library streaming Avro, Parquet and Iceberg; Ray Train with FSDP, HSDP, elastic scaling and async checkpointing; and on-demand cluster provisioning. Routing data loading through dedicated CPU nodes over a zero-copy bridge cut dataloader memory 50-70%, and the team had to work around Ray to reclaim 24% of step time. Uber's session on migrating the Eats recommendation stack from TensorFlow/Horovod to PyTorch uses Ray Data to replace Spark-based preprocessing, reporting a 5x data transformation speedup, 90% lower memory usage and 20x training throughput improvement. Google's Keeping GPUs Busy presents Rapid Storage, a gRPC-based protocol for PyTorch via fsspec that the team says extends to Ray, Dask, HF Datasets and vLLM.

Post-training and RL sessions include Pinterest's PTEnv lifecycle harness, which supports subprocess execution with Ray for RL and in-process with MS-Swift using PyTorch FSDP/DDP for SFT, making framework choice a config line. Anyscale's SkyRL talk covers fully async RL training on 350 billion-plus parameter MoE models with Megatron and vLLM, a multi-tenant Tinker Engine and HTTP-based API redesign. A Demo Theater session on Ray Core covers scheduling across 10,000-node clusters and Ray Direct Transport, which moves PyTorch tensors directly between GPUs.

Why it matters: teams building AI infrastructure on PyTorch can see concrete, reusable scaling patterns and benchmarks from LinkedIn, Uber, Pinterest and Google, and follow ongoing work on Ray and Kubernetes integration.