
Time correction: GPU Detective Stories (see official listing)
About this event
This is a schedule correction, not a second meetup. Please use the official Tokyo listing: https://www.meetup.com/ai-performance-engineering-meetup-tokyo/events/317171624/ GPU Detective Stories: PyTorch Recompiles & Hidden Bottlenecks Tuesday, November 17, 2026 · 2:00–3:00 a.m. JST Monday, November 16, 2026 · 9:00–10:00 a.m. PST Online · Zoom: https://us02web.zoom.us/j/82308186562 Talk #0: Introductions and Meetup Updates Chris Fregly and Antje Barth Talk #1: The Recompile Ambush: Taming PyTorch Inference Spikes Chaim Rand Chaim will explain why torch.compile can trigger costly recompilation when input shapes or layouts change, and compare practical ways to control inference latency: cache preloading, bounded recompilation with eager fallback, Dynamo and input-layout tuning, warmup seeding, background compilation, and ahead-of-time compilation. Talk #2: Follow the Profile: Hunting Jev’s GPU Bottlenecks Israel Ogbole, Co-founder and CEO of Zymtrace Israel will explore the Zymtrace/TypeSafe Jev case study and how continuous CPU/GPU profiling helped identify and improve compute, memory, communication, and inference-synchronization bottlenecks. The published case study reports approximately 46% higher training throughput on the same hardware and approximately 10% better inference performance. Please use the official Tokyo listing above for the full agenda and RSVP. This page is a correction to the previously advertised time, not an additional event.
Questions & comments
Ask the host anything — replies are visible to everyone.
—