Infra behind Krea 2: How to train and serve at scale — Gabriel Jorge Menezes, Krea.ai
- Source
- youtube.com
- Author
- AI Engineer
- Date
Context
Krea's account of training Krea 2 emphasizes recovering from failures quickly. Its infrastructure talk describes communication timeouts that could leave dashboards looking healthy, making failures difficult to diagnose. The team's response included frequent checkpoints, saved training states from which a run could resume. It reports a filesystem able to write a terabyte in under 30 seconds. Training infrastructure account.
The connection is operational: frequent snapshots preserve recent progress when a large training job crashes, while fast storage makes those snapshots practical to take repeatedly. The reported storage capability explains the recovery strategy; it is not a measurement of overall training uptime. Checkpointing discussion.
videoVertical Mobility: Inference from MVP to Trillion-Parameter Workloads — Sitanshu Gupta, CoreWeaveAI Engineer
videoGuardians of the State: An Air-Gapped AI Fortress for Consumer Data — Rachna Srivastava, DFPIAI Engineer
videoThe Messy Reality of Scale: Synthetic Data and Pre-Training — Marah Abdin & Robert McHardy, poolsideAI Engineer
Checking sign-in…
Loading comments…



