Evgeny

Evgeny Potapov

Tech Leader

Yerevan, Armenia

Talk

GPU Training Failures that Hide From Your Dashboard
Track: Software Engineering Duration: 50 minutes View on Schedule
Large Language Models (LLMs) Distributed Systems GPU Performance Machine Learning Infrastructure

Large-scale model training rarely fails the way engineers expect it to. The intuition borrowed from other domains treats it as a batch job - provision the hardware, launch the run, wait for convergence - but published accounts from Meta's Llama 3 pre-training and ByteDance's training infrastructure describe interruptions every few hours and thousands of jobs per quarter that continue holding GPUs and incurring cost long after progress has stopped. The costly failures are usually not crashes, but runs that look healthy while computing nothing.


Part of the reason is that the most familiar metric misleads. GPU utilization indicates only that a kernel was resident during the sample window, so a stalled job can report near-total engine activity alongside almost no work in the streaming multiprocessors, and every dashboard stays green.


This talk goes through five failure profiles observed on real training clusters - checkpoint stalls, distributed hangs, stragglers, communication-bound runs, and silently throttled GPUs - showing for each what the telemetry actually looks like and which standard tooling exposes it, from DCGM counters and NCCL debug output to the PyTorch flight recorder, eBPF probes, and host stack sampling.


About the Speaker

Experienced tech leader with a demonstrated history of working on large-scale projects from the ground up and helping organizations define their strategy. Skilled in software development, cloud architecture, performance optimization and scalability, Agile, DevOps, CI/CD, Cloud Services, DevSecOps practices, and Infrastructure as Code (IaC) platforms. Strong engineering professional.

Recording

Video will be available after the conference.