HyperPod model caching cuts LLM cold starts from 30 minutes to seconds
AWS ML Blog · rss · 2026-09-11
Amazon SageMaker HyperPod launched model caching: pre-loading model weights and container images onto node-local NVMe (7 GB/s) so pods start serving in seconds instead of waiting 30+ minutes of downloads. Cold starts were dominated by two sequential downloads: multi-GB inference images (5–7 min) plus weights (20+ min for a 145GB model, 30+ min for DeepSeek-R1 at 600+GB), repeated on every scale-out.
Two independently toggleable caches:
- Weights cache: the operator pre-downloads weights from S3/FSx/HF Hub to node NVMe, labels nodes cache-ready before creating the deployment, and the cache persists across pod restarts.
- Image cache: a DaemonSet pre-pulls the image, skipping ECR downloads; multiple deployments share one cache.
Scheduling is preferred, not required: pods favor cached nodes but fall back to normal downloads without failure. Enable via modelCacheConfig (weightsCache/imageCache); model updates roll out a new cache and clean the old one with zero downtime.
More from Infra
- M.2-Oculink eGPU Link Silently Downgrades to PCIe Gen1 — Here's How to Check — El_90 · 2026-09-11
- DeepSeek launches V4.1-Flash with 1M-token context and 4x smaller KV-cache — matlabulous · 2026-09-11
- What Can You Still Run on 8GB VRAM? User Asks for Small Models With Tool Use — riceinmybelly · 2026-09-11
- Spain's hourly 80% renewable matching rules clash as France fast-tracks 700MW sites, UK cuts grid queues — eherrerosj · 2026-09-11
- AI could add 0.3-0.4 points to Europe's productivity growth, but the EU holds under 5% of global compute — rohanpaul_ai · 2026-09-11
- Qualcomm's next-gen Hexagon NPU runs 30B MoE models with 32K context on-device — lee_stott · 2026-09-11