events

Tame PyTorch Recompiles & Unlock Jev GPU Throughput with Zymtrace

An online session from the AI Performance Engineering Meetup (Munich2) for developers and researchers working on ML systems, with two technical talks. Chaim Rand compares six ways to handle PyTorch recompilation during real-time inference,

Starts

2026-11-16T17:00:00+00:00

Organizer

AI Performance Engineering Meetup (Munich2)

Location

Online event

An online session from the AI Performance Engineering Meetup (Munich2) for developers and researchers working on ML systems, with two technical talks. Chaim Rand compares six ways to handle PyTorch recompilation during real-time inference, including caching, bounded recompilation with eager fallback, Dynamo and input-layout tuning, warmup, background compilation, and ahead-of-time compilation with torch.export and AOTInductor. Israel Ogbole presents a Zymtrace/TypeSafe case study on profiling Jev training and serving, citing reported gains of about 46% higher training throughput and about 10% better inference after removing a hidden device-to-host synchronization.