Performant Training for All Workstreams | Latitude AI | Ray Summit 2026 @anyscale
Performant Training for All Workstreams | Latitude AI | Ray Summit 2026  @anyscale
Uploaded September 2026 | Updated November 2026, 20 hours ago
GPU utilization is a useful first signal, but past a point it decorrelates from real progress: a busy GPU is not always a fast one. Throughput maps directly to faster training, lower cost, and shorter iteration loops, so it is the metric Latitude optimizes against.

At Ray Summit 2026, Nicolas Cebron and Krishna Toshniwal from Latitude AI show how they tune the three stages that determine throughput together, the Ray Data pipeline, data transfer onto the GPU, and the training step, measuring each with Ray actors that report per-stage speed and rolling the signals into a fleet-wide live dashboard with one-click profiles for any run.

You'll leave with the throughput gains this made possible and a blueprint for making training throughput a metric your whole org watches and steadily drives up.

Liked this video? Check out other Ray Summit breakout session recordings: youtube.com/playlist?list=PLZNaJyYZRPdo

Subscribe to our YouTube channel to stay up-to-date on the future of AI: youtube.com/c/anyscale

đź”— Connect with us:
LinkedIn: linkedin.com/company/joinanyscale
X: https://x.com/anyscalecompute
Website: anyscale.com
Performant Training for All Workstreams | Latitude AI | Ray Summit 2026How Coinbase Uses Ray, vLLM & LiteLLM to Power Secure LLM Services | Ray Summit 2025MLOps with Ray on Anyscale | Ray Summit 2025Distributed Embeddings at Scale: Processing 10M+ Rows/ Day with Ray, GPUs & Qdrant | Ray Summit 2025Scaling Physical AI & Robotics Systems with Ray | Anyscale | Ray Summit 2026Serving vLLM on Intel GPUs, CPUs, and Gaudi | Intel | Ray Summit 2026Elastic Expert Parallelism for vLLM | Ray Summit 2025Post-training Nemotron with Ray | Bryan Catanzaro (Nvidia) | Ray Summit 2026Apple’s Approach to Scalable Machine Learning Infrastructure on Ray | Ray Summit 2025Serving Fast and Smart: Getting the Most Out of Ray Serve LLM on GKE | Google | Ray Summit 2026Anyscale on Azure: Build, Train, and Serve AI in Your Own Tenant | Microsoft | Ray Summit 2026Improved Scheduling Flexibility with Label Selectors in Ray | Ray Summit 2025
Anyscale |

Performant Training for All Workstreams | Latitude AI | Ray Summit 2026

SHARE TO X SHARE TO REDDIT SHARE TO FACEBOOK WALLPAPER