events
GPU Detective Stories: PyTorch Recompiles & Hidden Bottlenecks
Two technical talks on hidden costs in AI performance, held online by the AI Performance Engineering Meetup (Toronto). Chaim Rand covers six approaches to handling PyTorch recompilation when a new…

En inglés
- Starts
- 2026-11-16 17:00 UTC
- Organizer
- AI Performance Engineering Meetup (Toronto)
- Location
- Online event
Two technical talks on hidden costs in AI performance, held online by the AI Performance Engineering Meetup (Toronto). Chaim Rand covers six approaches to handling PyTorch recompilation when a new input shape or layout triggers it during inference — cache preloading, bounded recompilation with eager fallback, warmup seeding and ahead-of-time compilation among them — and the tradeoffs between them. Israel Ogbole of Zymtrace presents a case study on training and serving Jev, where continuous CPU/GPU profiling guided changes the study reports as roughly 46% higher training throughput and about 10% better inference.
Event website