events
GPU Detective Stories: PyTorch Recompiles & Hidden Bottlenecks
Online meetup from the AI Performance Engineering group, hosted by Chris Fregly and Antje Barth, with two talks on overlooked costs in AI workloads. Chaim Rand explains how torch.compile can…

Auf Englisch
- Starts
- 2026-11-16 17:00 UTC
- Organizer
- AI Performance Engineering Meetup (London)
- Location
- Online event
Online meetup from the AI Performance Engineering group, hosted by Chris Fregly and Antje Barth, with two talks on overlooked costs in AI workloads. Chaim Rand explains how torch.compile can recompile mid-request when an input shape or layout changes, comparing six ways of handling it, including warmup seeding, background compilation and ahead-of-time export, and their tradeoffs. Israel Ogbole of Zymtrace presents a case study on training and serving the Jev model, where continuous CPU/GPU profiling guided three changes; the published study reports about 46% higher training throughput on the same hardware.
Event website