AI Infra Summit at PyTorch Conference 2025
2025-10-21
San Francisco, United States
who it's for
platform engineers, infrastructure engineers, research scientists, reinforcement learning researchers, security researchers, compiler engineers
topics
fine-tuning pipelines at scale; data and compute orchestration; serving stack sharding and caching; custom silicon kernel optimizations; FlashInfer inference benchmarking; TVMI in-memory tensor ABI; large-scale decentralized pretraining; gradient and activation compression; multicluster asynchronous RL training; environments hub and verifiers SDK; cybersecurity agent evaluation suites; automatic kernel generation with Triton/Mojo
| title | author | length |
|---|---|---|
| Keynote: Scaling Intelligence: Infrastructure as the Foundation | Joe Spisak | 02m |
Walks through concrete infrastructure pain points and the engineering responses: why fine-tuning pipelines break at scale and how data/compute orchestration… | ||
| Keynote: Building ML Systems Foundations at the Age of AI | Tianqi Chen | 16m |
Launches a curated benchmark and workload spec called flashing for trace that records real LM inference shapes and operator patterns, plus a hosted leaderboard… | ||
| The Four Pillars of Large-Scale Decentralized Pre-Training: Methods, Bottlenecks... | Bowen Peng | 27m |
Survey of four required capabilities: optimization, coordination, verification, and data/privacy, with concrete bottlenecks and partial fixes. Optimization… | ||
| Can Reinforcement Learning Lead to AGI? | Daniel Han | 26m |
Claims that RL can scale to AGI rest on concrete pivots: scaling inference (currently ~90% of compute), automating reward functions, and handling long context… | ||
| Panel Discussion: Training and Inference at Planet Scale | Jana van Greunen, Lianmin Zheng, Michael Suo, Robert Nishihara, Simon Mo, DMYTRO DZHULGAKOV | 34m |
Panel covers concrete engineering moves: using five- or six-dimensional parallelism and mixture-of-experts sharding, splitting prefill vs decode into separate… | ||
| Scaling an Open Environments Ecosystem for Reinforcement Learning | Will Brown | 27m |
Describes an open stack centering on verifiers SDK and the Environments Hub plus prime-rl/Primaril trainer, claims multicluster async RL works with up to… | ||
| Environments and Benchmarks for Cybersecurity Agents | Andy Zhang | 28m |
Cybench evaluates offensive ability on 40 professional CTFs with verifiable tasks, subtasks for partial credit, and a first-solve time metric tied to human… | ||
| Race to Superintelligence: Building Low-Friction AI Infrastructure for a Heterogen... | Ajit Mathews | 25m |
Meta scales PyTorch into a hardware OS using Triton and Torch Inductor to target a family of MTIA accelerators with an MLIR Triton backend, production… | ||
| Panel Discussion: The AI Kernel... | Robert Lange, Simran Arora, Nathan Lambert, Mark Saroufim | 32m |
Kernel-level optimization uses LLMs plus tooling like Triton, Thunderkittens and Mojo to generate CUDA/HIP/PTX kernels but many auto‑generated kernels silently… | ||
| nothing matches | ||