Keynote: Building ML Systems Foundations at the Age of AI

on YouTube ↗16m

" summary

Launches a curated benchmark and workload spec called flashing for trace that records real LM inference shapes and operator patterns, plus a hosted leaderboard flashing bench showing current AI agents do not yet beat hand-tuned kernels. Claims enablement of day-zero integration so a better AI-generated kernel can swap into FlashInfer with zero lines of user code. Introduces TVMI ABI, a stable C in‑memory tensor API with first‑class PyTorch support to let one library and DSOs interop across Python versions and runtimes.

AI-written summary · about

" chapters (5)

1. 0:00AI-powered ML systems intro
2. 2:25Optimizing inference kernels
3. 7:38ML interoperability challenges
4. 10:47TVMI for cross-framework
5. 15:03Call to action and future

" related talks

• 23mKeynoteLaurence Moroney, Sharbani Roy, Joana Carrasqueira, Dominique Thorne · Women in Machine Learning Symposium 2022
• 12mKeynote: What's New in AIEve Phillips · Cassandra Summit 2023
• 16mKeynote: Accelerating AI/ML Development with Open Source SoftwareBrian E. Granger · AI.dev: Open Source GenAI & ML Summit North America 2023
• 38mFireside chatChris Lattner · Scale By the Bay 2023
• 06mKeynote: PyTorch Lightning: Powering the GenAI Revolution from Research to the...William Falcon · PyTorch Conference 2023
• 16mKeynote: An Open-Source Stack for AI Compute: Kubernetes + Ray + PyTorch + vLLMRobert Nishihara · Open Source Summit North America 2025
• 13mKeynote: Enabling PyTorch’s Thousand Ops for Software First Silicon DesignAndrew Ling · PyTorch Conference 2025
• 24mKeynote: Next-generation frameworks for Large-scale AIAnima Anandkumar · Scale By the Bay 2020