events
GPU Detective Stories: PyTorch Recompiles & Hidden Bottlenecks
Due interventi sulle prestazioni dell’IA: contenere i picchi di ricompilazione di PyTorch nell’inferenza e profilare un’esecuzione di addestramento per individuare i colli di bottiglia della GPU.

In inglese
- Starts
- 2026-11-16 17:00 UTC
- Organizer
- AI Performance Engineering Meetup (San Francisco, Global)
- Location
- Online event
The AI Performance Engineering Meetup runs two technical talks on where AI workloads lose time. Chaim Rand, a machine-learning algorithm developer, compares six approaches to handling PyTorch recompilation triggered by new input shapes during inference, from cache preloading and warmup seeding to ahead-of-time compilation with torch.export. Israel Ogbole of Zymtrace presents a profiling case study on training and serving the model Jev, describing how compute, memory and communication bottlenecks were located, measured and changed. Organizers Chris Fregly and Antje Barth open with meetup updates.
Event website