AI @Scale 2022
2022-09-28
-- online --
who it's for
machine learning engineers, ml infrastructure engineers, research scientists, platform engineers, data engineers, site reliability engineers, applied product managers
topics
training 175B parameter models; self-supervised pretraining; PyTorch compiler optimizations; model parallelism with SageMaker; three-tier checkpointing strategies; compiler-accelerated training and inference; continuous learning for autonomous vehicles; large-scale sensor data pipelines; model performance monitoring metrics (PSI/KL/JS/KS); synthetic pretraining for computer vision; ONNX/Triton inference graph execution; quantization and specialized hardware
| title | author | length |
|---|---|---|
| Fireside Chat | Yann LeCun, Ludovic Hauduc | 35m |
Fireside Chat summarizes concrete claims and tools: large models improve with self-supervised learning so one pretrained model serves many tasks; open research… | ||
| OPT-175B: LLM Development Lifecycle & Challenges | Susan Zhang | 17m |
OPT-175B is a 175 billion parameter GPT-style model released for research in May 2022 and requires access request; development faces repeated large-scale… | ||
| ML Infrastructure for Autonomous Vehicles | Alexander Sidorov | 21m |
Cruise runs over a petabyte of high‑resolution temporal sensor data, collects tens of gigabytes per mile, and expects hundreds of thousands of vehicles this… | ||
| AI @Scale | Yuan Yu | 22m |
Microsoft describes a unified Azure AI @Scale stack that actually ships: ACPT containers bundle PyTorch with DeepSpeed and Onnx/Onyx runtime so teams run the… | ||
| The Laws of ML at Scale | Ketan Umare | 19m |
Three concrete laws: scaling triggers a vicious cycle where ad hoc infra requests (Spark, GPUs) balloon cost and never finish; early trade-offs compound into… | ||
| The New Vision of Compiler-Accelerated | Peng Wu | 21m |
Claim: an out-of-box front end funnels most PyTorch models into optimizing backends with no model changes, using TorchDynamo's partial graph capture, runtime… | ||
| At-scale Training | Andrea Olgiati | 18m |
Shows how SageMaker Model Parallel + PyTorch actually partitions and trains a GPT-J–class model across P3 and P4 instances, using pipeline, tensor and data… | ||
| Supporting AI holistically Across the Infra stack | Kaushik Veeraraghavan | 23m |
Meta describes how PyTorch eager mode drives co-design of compilers, chips, racks and data centers, claiming a 50 billion parameter No Language Left Behind… | ||
| Human Friendly, Production Ready Data Science Stack | Savin Goyal | 24m |
An open‑source stack from Netflix called Metaflow provides concrete, production features: Python @step decorator to define DAG nodes, automatic snapshotting… | ||
| Model Performance Monitoring - A Practitioner's Perspective | Krishna Gade | 19m |
Claims a deployable Model Performance Management platform that continuously detects concept, data, label and feature drift, attributes degradations with a… | ||
| Customizable Computer Vision Expands Data Access Without Compromising Privacy | Sujoy Ganguly | 18m |
Open-source Perception package and People Sans People provide a CLI, Mac/Linux binaries, rich labels (boxes, 3D boxes, seg, keypoints, depth, lidar) and… | ||
| The Future of Inference for Complex AI Applications | Jake Mannix | 19m |
Claims a working, stateless synchronous DAG executor — an inference graph execution service — that stitches ONNX/Triton model calls, feature-store fetches… | ||
| Live Q&A Session 1 | Michael Chen | 27m |
Panel defines scale as data, compute, concurrent projects, use-case variety and velocity and claims fragmentation is the main bottleneck; proposes a “one… | ||
| Live Q&A Session 2 | Michael Zhang | 33m |
Panel answers focus on concrete engineering tradeoffs: scale means training huge models, cost-effective serving, and infusing models across apps; current… | ||
| nothing matches | ||