以下部分引用 AI 总结，现阶段 AI 仍然有幻觉。请以内容的原文为准。

数据新鲜度群聊 07-14 ✓卖方 07-14 ✓Wrap 断档15天 ✕News 07-15 ✓X 断档12天 ✕页面生成 07-15

Fluxion

6 atoms · 跨 1 天 · 首见 2026-05-11 · 最近 2026-05-11

三色: 🟦 fact 6 · 🟥 take 0 · stance ▲3/▼0/◆3

来源: X 6

时态: fresh:6

标签: 好信源:3 · 好数字:3

2026-05-11 · 系统类型
In this paper is presented Fluxion, an efficient hybrid sparse-attention system for long-context inference with CPU-GPU parallelism.
tags: 好信源 · → daily
2026-05-11 · 首创优化维度
Fluxion is the first system to identify top-k selection granularity as a key optimization dimension for CPU-side sparse attention, and to expose its predictable coupling with KV budget.
tags: 好信源 · → daily
2026-05-11 · 系统组件
Fluxion combines a lightweight head-property predictor, a granularity-budget selector, and a priority-based scheduler to jointly optimize budget allocation, sparse configuration, and CPU-GPU execution overlap.
tags: 好信源 · → daily
2026-05-11 · 每输出Token加速比
The results show that Fluxion delivers 1.9×–3.4x time per output token speedup on Qwen and 2.5×–3.7x on Llama.
tags: 好数字 · → daily
value: qty=1.9×–3.4x
2026-05-11 · 每输出Token加速比 (Llama)
Fluxion delivers 2.5×–3.7x on Llama.
tags: 好数字 · → daily
value: qty=2.5×–3.7x
2026-05-11 · GPU空转比例改善
under a representative 32K-context setting on Qwen, the idle ratio drops from 70.65% to 45.78%.
tags: 好数字 · → daily
value: qty=70.65%降至45.78% · date=32K上下文