跳到正文
RCreddit.com·
暂不在当前实时榜单

R9V Update: Created and adopted KVA projections based on Deepseek V4.1 Flash + HySparse2/MiMo-V3 for Qwen3.8 Flash Next. This is a game changer for models that don't natively implement it. 1.45-1.85x speedup in prefill to 3k+ at a small deficit to perplexity. [2x R9700, 128GB DDR5]

AI 摘要

R9V has implemented KVA projections, derived from Deepseek V4.1 Flash and HySparse2/MiMo-V3, for Qwen3.8 Flash Next. This innovation significantly boosts prefill speed by 1.45-1.85x up to 3k+ tokens, with a minor impact on perplexity. Performance metrics include a prefill rate of 3,600 t/s at 64k tokens, decode at 74.7 t/s, and a combined concurrent rate of 70.3 t/s across 8 streams. This development is particularly impactful for models lacking native KVA projection support.

时间与来源

时间显示为 UTC

显示时区:UTC

本地时区尚不可用,暂时显示 UTC。

收录当时偏移:UTC+02026年9月25日 10:01 UTC

收录
2026年9月25日 10:01
来源类型
开发者社区

本站未收录正文。

前往源站阅读 →
来源·reddit.com