Vai al contenuto
AI.info

events

GPU Detective Stories: PyTorch Recompiles & Hidden Bottlenecks

Due interventi su come gestire la ricompilazione durante l’inferenza in PyTorch e sulla profilazione dei colli di bottiglia della GPU nell’addestramento e nell’erogazione del modello Jev.

GPU Detective Stories: PyTorch Recompiles & Hidden Bottlenecks | Eventi sull’IA

In inglese

Starts
2026-11-16 17:00 UTC
Organizer
AI Performance Engineering Meetup (Paris)
Location
Online event

An online session from the AI Performance Engineering Meetup in Paris, hosted by Chris Fregly and Antje Barth, with two technical talks on hidden costs in AI workloads. Chaim Rand compares six ways to handle PyTorch recompilation spikes during inference, from cache preloading and bounded recompilation with eager fallback to ahead-of-time compilation with torch.export and AOTInductor. Israel Ogbole of Zymtrace then walks through a profiling case study of training and serving the Jev model, covering compute, memory, communication and synchronization bottlenecks.

Event website