Ray Summit 2021
2021-06-22..2021-06-24
-- online --
[Ray Summit 2020, Ray Summit 2021, Ray Summit 2024, Ray Summit 2026]
| title | author | length |
|---|---|---|
| [Opening Keynote] State of Ray | Robert Nishihara | 09m |
Ray reports concrete growth and engineering wins: 480 contributors and production wins like XGBoost running 5x faster on Ray and Dask workloads running 13x… | ||
| [Opening Keynote] Making Distributed Computing Easy | Ion Stoica | 16m |
AI workloads outpace single machines — compute demand grows 35x every 18 months while per‑core performance inches up, so training and serving go distributed… | ||
| [Opening Keynote] Anyscale Demo: Machine Learning Application from Dev to Prod | Edward Oakes | 27m |
Shows a runnable pipeline that trains an intent classifier with Ray Tune and XGBoost on Ray, runs 100 hyperparameter trials across ~155 CPUs on 10 worker… | ||
| Responsible AI at Scale | Sarah Bird | 16m |
Azure AI describes a concrete responsible-AI pipeline: a 360° impact assessment that yields design requirements, continuous fairness testing using labeled test… | ||
| Human-in-the-Loop Reinforcement Learning | Pieter Abbeel | 14m |
A pipeline converts human pairwise preferences over rollouts into a learned reward model (Bradley Terry style, trained with cross entropy) and then runs… | ||
| End-to-End AutoML with Ludwig on Ray | Travis Addair | 21m |
Ludwig on Ray runs the same Ludwig config locally or at cluster scale with one parameter on the command line, enabling out‑of‑core preprocessing via Dask on… | ||
| Offline RL with RLlib | Edi Palencia | 30m |
RLlib is extended to run Conservative Q‑Learning (CQL) by modifying DQN losses and reusing existing RLlib components and CLI. A dataset generator produces… | ||
| Ray Serve: Patterns of ML Models in Production | Simon Mo | 25m |
Ray Serve is a Ray-based serving framework developed from 100+ user interviews that claims to deploy many models as composable microservices with per-component… | ||
| Petabyte Scale Datalake Table Management with Ray, Arrow, Parquet, and S3 | Patrick Ames | 26m |
Amazon builds a production flash compactor using Ray, Arrow, Parquet, and S3 to apply upserts/deletes at petabyte scale, using partitioned primary key indexes… | ||
| A 24x Speedup for Reinforcement Learning with RLlib + Ray | Raoul Khouri | 19m |
Migration from Stable Baselines to RLlib + Ray turns a CPU‑bound financial simulator into a 24x faster workflow: tests go from ~7–10 hours to ~2 hours on a… | ||
| Distributed XGBoost on Ray | Patrick Dolan | 28m |
XGBoost-Ray implements a Ray-based backend that uses Ray actors and Rabit all-reduce to run multi-node, multi-GPU XGBoost with locality-aware loading from… | ||
| Ray and Anyscale: An Optimization Journey | Nachi Mehta | 13m |
Delivers a concrete pipeline that turns a trained model into a scalable REST endpoint using Ray Serve on Anyscale, Elasticsearch and Locust for load testing… | ||
| Detection of Sensitive Personal Data in Large Data Sets Using Ray with Deep Learning | Srinivasa Rao Aravilli | 23m |
(Srinivasa Rao Aravilli, Visa)
In this talk, Srinivasa discusses:
* Traditional approaches to detect the sensitive data for data protection and data loss… | ||
| The Quick Journey to Using Ray: How We Implement Ray and Anyscale to Speed up our ML Processes | Juan Roberto Honorato, Domingo Ortuzar | 20m |
Anastasia migrates a demand-forecasting pipeline from AWS Batch to Ray with Anyscale, reworking code to queues and actors so every node handles every model. A… | ||
| Using Reinforcement Learning to Optimize IAP Offer Recommendations in Mobile Games | Emiliano Castro, Vinicius Alves, Felipe Antunes | 27m |
Uses offline reinforcement learning with behavior cloning then policy improvement, running on a Kubernetes cluster with Ray (RLlib and Tune) to recommend IAP… | ||
| Data Science at Scale: From Notebook to Production with Airflow and Ray | Daniel Imberman, Rob Deeb | 25m |
Airflow now runs Ray jobs natively: add the ray_task decorator to a Python function and the task executes on a Ray cluster, returning a Ray object ID stored in… | ||
| Scaling Interactive Data Science with Modin and Ray | Devin Petersohn | 21m |
Modin runs as a drop-in replacement for pandas with native Ray and Dask execution and a plugin to hand work to external engines like OmniSci via Apache… | ||
| Per-epoch Shuffling Data Loader: Mix It Up As You Train! | Clark Zinzow | 28m |
Ray-based per-epoch shuffling data loader yields globally shuffled GPU batches to dozens of trainers via an iterable dataset API, using a mapper→reducer… | ||
| W&B Tables and Ray | Chris Van Pelt | 26m |
Shows W&B Tables and Artifacts used with Ray to log example-level results: a 3-class Inception model evaluates a rock–paper–scissors dataset, producing a table… | ||
| Data Processing on Ray | SangBin Cho | 26m |
Ray adds concrete features to run large-scale ETL and ETL→ML pipelines: it implements memory-aware scheduling and locality-aware placement to avoid needless… | ||
| How Ray and Anyscale Make it Easy to do Massive-scale ML on Aerial Imagery | Richard Decal | 21m |
Dendra runs PyTorch on Ray SGD and Ray Tune to scale from a laptop to dozens of GPUs in one Python script, and reports Ray Tune saves ~$16k through aggressive… | ||
| RayDP: Build Large-scale End-to-end Data Analytics and AI Pipelines Using Spark and Ray | Carson Wang | 26m |
RayDP runs Spark on Ray by launching Spark driver and executors as Ray Java actors, uses the Ray in-memory object store to exchange DataFrames without HDFS… | ||
| Ray Internals: Object Management with the Ownership Model | Stephanie Wang, Yi Cheng | 24m |
Ray restructures object metadata by shards the master across workers so the creator of an object owns its metadata, keeping location and lineage local to… | ||
| Neural MMO: Building a Massively Multiagent Research Platform with Ray and RLlib | Joseph Suarez | 19m |
Neural MMO is an open research environment built with Ray and RLlib that runs up to 1024 agents in procedurally generated persistent worlds and supports long… | ||
| Achieving Scalability and Interactivity with Ray Serve | Jaehyun Sim | 21m |
Ikigai Labs builds Eki, a pipeline platform that lets users inject arbitrary Python steps and needs both massive scale and sub-second peeking. Ray and Ray… | ||
| Building High Availability and Scalability Online Computing Applications on Ray | Tengwei Cai | 14m |
Framework that turns Ray jobs into low‑latency, high‑availability online services by running a long‑running Ray cluster with a long‑running APP master actor… | ||
| Simplify Big Data & AI on Spark and Ray with MLSQL | Dong Li | 30m |
MLSQL fuses Spark and Ray so SQL can orchestrate ETL, feature engineering and model training with embedded Python, using existing Ray clusters and standard… | ||
| Improving Ray for Large-scale Applications | Yi Chang | 23m |
Ant Group runs Ray at >200,000 CPU cores and reports concrete changes: replaces relay actor RPC with direct actor calls via gRPC implemented in C++ for a 6x… | ||
| Mobile Order Click-Through Rate (CTR) Recommendation with Ray on Apache Spark at Burger King | Luyang Wang | 30m |
Burger King deploys a Transformer Cross Transformer that runs a sequence Transformer (current-session token plus averaged purchase-history token) and a context… | ||
| Application of Online Resource Allocation Based on Ray | Fang Fengbin | 17m |
Ant Group deploys an online resource allocation system built on Ray for coupon, recommendation and ad allocation that targets maximizing platform revenue under… | ||
| DeepCoMP: Multi-agent Reinforcement Learning for Multi-cell Selection in 5G and Beyond | Stefan Schneider | 24m |
DeepCoMP is a working multi-agent deep RL system (PPO, OpenAI Gym) that learns dynamic multi‑cell selection from partial observations and a log‑QoE reward… | ||
| Panel: ML Platform on Ray | Zhe Zhang, Xu Ning, Azeem Ahmed, Bhavya Agarwal | 44m |
Companies replace Spark with Ray for unified training and serving, running XGBoost and GPU deep learning with measurable wins: a 150M-parameter BERT variant… | ||
| Lessons from the Kubernetes Adventure | Eric Brewer | 14m |
Kubernetes is a code-first, rapidly changing platform (material changes every quarter) that turns containers into service-level primitives with per-pod IP/port… | ||
| Elements of Learning Systems | Tianqi Chen | 19m |
Covers two concrete systems: XGBoost and TVM. Claims XGBoost removes manual imputation by learning a automatic missing-value default direction inside tree… | ||
| Towards Building a Responsible Data Economy | Dawn Song | 21m |
Argues that secure enclaves, cryptographic MPC, differential privacy and federated learning combine with blockchains to make data usable without copying and to… | ||
| JAX: Accelerated Machine Learning Research via Composable Function Transformations in Python | Matt Johnson | 14m |
JAX gives NumPy-style code hardware acceleration by composing function transformations: grad, jit, vmap implement autodiff, end-to-end XLA compilation, and… | ||
| Cloud Computing Made Better: An AWS View on Serverless Computing | Marvin Theimer | 17m |
Serverless adoption has exploded: over half of new applications built inside Amazon, driving trillions of invocations per month and some customers at tens of… | ||
| Scaling and Unifying SciKit Learn and Spark Pipelines using Ray | Raghu Ganti | 17m |
A pipeline abstraction on Ray maps Scikit‑Learn and Spark pipelines to Ray tasks, actors and the distributed object store, using a list of object references as… | ||
| How Smart is the ECG? Lessons from Advanced ECG Analytics | Dr. Peter Noseworthy | 28m |
A convolutional neural network trained on ~100,000 paired 12‑lead ECGs and echocardiograms detects low ejection fraction with AUC 0.93, hypertrophic… | ||
| Leveraging the Possibilities of Ray Serve | Tanja Bayer | 29m |
Built a fully automated digital authentication service called validator that runs on-prem and processes front/back ID video plus face checks, using Ray Serve… | ||
| Collective-on-Ray: High-performance Collective Communication for Distributed Machine Learning on Ray | Hao Zhang | 19m |
Ray-native Python collective primitives run on top of Python bindings to NCCL and Gloo and implement all-reduce, broadcast, point-to-point and declarative… | ||
| NumS: Scalable Array Programming for the Cloud | Melih Elibol | 20m |
NumS compiles NumPy-like code to a Ray-backed runtime that schedules block-wise operations using futures and graph-arrays, trading element-wise overhead for… | ||
| Anomaly Detection on Remote Sensing with Ray + Horovod | Linsong Chu | 13m |
Uses a U-Net frame prediction model that ingests t−k..t−1 frames to predict frame t and flags pixels where prediction versus ground truth error exceeds a… | ||
| The Journey from AI to Bedside: Translating Academic Research to Serving Patients Using Ray | Roger Ng | 27m |
Builds a web app that uses RLlib and reinforcement learning to convert EMR notes, vitals and twice‑daily patient uploads into sequential disease states, trains… | ||
| A Deep Dive into Ray’s Scheduling Policy | Jiajun Yao | 28m |
Ray implements a distributed dynamic scheduler with a centralized GCS and per-node raylet processes and claims scalable execution of hundreds of thousands of… | ||
| Large Scale Data Loading and Data Preprocessing with Ray | Wei Chen | 22m |
NVIDIA describes replacing PyTorch’s multiprocessing DataLoader with Ray actors and the Ray object store to run dataset get_item calls in separate processes… | ||
| Applying Ray and RLlib to Real-life Industrial Use Cases | Edward Junprung, Sahar Esmaeilzadeh | 21m |
Claims that RL policies trained with Pathmind + Ray + AnyLogic beat heuristics/optimizers by measurable margins, with one HVAC demo reporting about 75%… | ||
| Machine Learning and Microbiology in the Cloud Through Ray | Stanley Bishop | 21m |
A cloud pipeline is built that uses JupyterLab to author code, Airflow to DAG and schedule, Ray to execute distributed jobs, Kubernetes/Docker for images… | ||
| Mars-on-Ray: Accelerating Large-scale Tensor and DataFrame Workloads | Chris Qin | 35m |
Mars-on-Ray runs Mars (distributed numpy/pandas/scikit-learn APIs) on Ray with an Oscar-based actor runtime, placement groups, and Ray object store spilling… | ||
| A Growing Ecosystem of Scalable ML Libraries on Ray | Amog Kamsetty | 26m |
Covers concrete integrations: Horovod is checked into its repo to use Ray for process placement and autoscaling, HuggingFace’s RAG moves retrieval into Ray… | ||
| Powering Open Data Hub with Ray | Erik Erlandson | 25m |
Integrates Ray 2.0 client-server into Open Data Hub by adding Kubernetes configmap-based single-user profile and service-template entries that spawn on-demand… | ||
| Using RLLib in an Enterprise Scale Reinforcement Learning Solution | Jeroen Bédorf, Ishaan Sood | 23m |
DeepSim uses RLLib + Tune on a Kubernetes Ray cluster with custom action distributions, loggers and network-config JSONs to run large-scale HPO and export… | ||
| Software 2.0 Needs Data 2.0: A New Way of Storing and Managing Data for Efficient Deep Learning | Davit Buniatyan | 14m |
Converts large unstructured datasets into unified tensor arrays stored on object storage so training reads them like local numpy arrays, supports petabyte… | ||
| Introducing Amazon SageMaker Kubeflow Reinforcement Learning Pipelines for Robotics | Sahika Genc | 25m |
Orchestrates reinforcement learning from perception to controls using SageMaker + Kubeflow pipelines and AnyScale’s Ray RLlib to scale trainers and simulators… | ||
| Fusion Computing at Ant Group | Tengwei Cai | 14m |
Ant Group describes a production fusion computing platform built on Ray and Kubernetes that runs about 200k CPU cores and scales to ~1,000 nodes and 10,000… | ||
| Deeper Dive on Making Boats Fly with AI on Ray | Nicolas Hohn | 28m |
Industrial flight-like yacht control is built: a commercial simulator is virtualized into a Gym environment, trained with deep Reinforcement Learning using a… | ||
| nothing matches | ||