跳到正文
RCreddit.com·
暂不在当前实时榜单

7900 XTX — two "low-thinking" Qwen 3.8 27B quants (Swift + ThinkingCap) vs the regular quant

AI 摘要

Performance tests on a 7900 XTX compared two "low-thinking" Qwen 3.8 27B quants, Swift and ThinkingCap, against a regular quant. The regular quant (unsloth) achieved ~530 t/s prefill and ~48 t/s decode, with a runtime of ~1407s. ThinkingCap showed ~100 t/s prefill and ~43 t/s decode over ~1087s, while Swift managed ~614 t/s prefill and ~32 t/s decode over ~1400s. The prefill speed of ThinkingCap was noted as unusually low, suggesting further testing is needed.

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 16:02 UTC

收录
2026年9月25日 16:02
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