X · 百度文心·· 2 天前AI 评分18
ERNIE Evolve 通过主模型、MoE 调度、量化融合与草稿模型迭代提升 14.9% 吞吐
Extreme decode speed was already heavily optimized. ERNIE Evolve still found 14.9% more throughput by iterating across the main model, MoE scheduling, quantization/fusion, and the draft model. https://t.co/DDkEh8BGcP
AI 导读
ERNIE Evolve 在解码速度已高度优化的基础上，通过在主模型、MoE 调度、量化/融合和草稿模型上迭代，额外获得 14.9% 的吞吐提升。
博主
Baidu ERNIE@Baidu_ERNIE
粉丝 3786
Foundation model family built by @Baidu_Inc. 🤗 https://t.co/xdqfVxgtTh 💬 https://t.co/kv5s6nWDsb
互动数据
- 浏览
- 78
- 点赞
- 0
- 评论
- 1
- 分享
- 0
- 收藏
- 0
数字来自平台，为最近一次抓取时的数据。
来源：X · 百度文心 · x.com