https://news.ycombinator.com/user?id=mailonceHacker News RSShttps://hnrss.org/hnrss v2.1.1Fri, 11 Sep 2026 16:56:48 +0000I've been running local image models on an older laptop recently, and memory behavior surprised me more than raw inference time. One experiment briefly pushed private memory past 18 GB before I changed the allocation behavior. After terminating the worker/process, it dropped dramatically. It made me realize how different "the model fits in memory" is from "the whole inference pipeline behaves well in memory."

]]>Thu, 10 Sep 2026 21:32:36 +0000https://news.ycombinator.com/item?id=49650436mailoncehttps://news.ycombinator.com/item?id=49650436https://news.ycombinator.com/item?id=49650436