antirez explains tiered prefill loading: SSD streaming hides expert cache thrashing in seconds
antirez · x · 2026-09-11
antirez discusses scheduling details of his local inference setup: a fully resident decoder evicts the experts cache, but the new SSD streaming implementation hides loading well and recovers within seconds.
The tiered strategy:
- Small prefills use the single-token path with the experts cache, like decoding
- Larger prefills use layer-major loading, where Layer N+1's loading is hidden by Layer N's processing
- Very large prefills switch to a full resident encoder
He considers this near-optimal across prefill sizes.
More from Infra
- RTX 5090 hits $7K as PC gaming community goes into 'denial' over AI boom — PAstynome · 2026-10-27
- Wan2GP lands on Pinokio: 1-click AI video on 6GB GPUs — cocktailpeanut · 2026-10-27
- AWS on agent prompt caching: explicit + implicit patterns to slash token costs — AWS ML Blog · 2026-10-27
- Thieves spoofing GPS steal $100M+ in data center cargo since June — Ars Technica AI · 2026-10-27
- Lam Research: weekly token usage up ~250x in 18 months — Beth_Kindig · 2026-10-27
- AWS shows how to build enterprise agentic retrieval on Bedrock with CDK — AWS ML Blog · 2026-10-27