•Three pre-defined StorageClasses are provided: gcsfusecsi-training (high-throughput reads), gcsfusecsi-serving (model loading/inference with Rapid Cache), and gcsfusecsi-checkpointing (fast writes for large checkpoint files)
•The CSI driver dynamically scans bucket size/object count and analyzes node resources (RAM, Local SSD, GPU/TPU) to calculate optimal cache settings automatically
•Manual tuning previously required navigating dozens of pages of configuration guides with settings varying by workload type and infrastructure
•Using the inference profile reduced model loading time for a Qwen3-235B-A22B workload on TPUs (480GB) from 39 hours to 14 minutes
•
Available in GKE version 1.35.1-gke.1616000 or later with the Cloud Storage FUSE CSI driver enabled
This summary was automatically generated by AI based on the original article and may not be fully accurate.
Google Cloud
AWS
Airbnb