events
GPU Detective Stories: PyTorch Recompiles & Hidden Bottlenecks
Due interventi online su come tenere sotto controllo le ricompilazioni di PyTorch nell’inferenza e sulla profilazione dei colli di bottiglia relativi a calcolo, memoria e comunicazione…

In inglese
- Starts
- 2026-11-16 17:00 UTC
- Organizer
- AI Performance Engineering Meetup (Seattle)
- Location
- Online event
An online session of the AI Performance Engineering Meetup in Seattle, opening with introductions and group updates before two technical talks. Chaim Rand compares six ways to handle `torch.compile` recompilations that appear in the request path when input shapes or layouts change, using a dynamic graph-neural-network example and covering tradeoffs against model, input distribution and latency needs. Israel Ogbole of Zymtrace then walks through a profiling case study on training and serving Jev, describing how bottlenecks were found and measured across compute, memory, communication and inference synchronization.
Event website