开发者在 Hacker News 平台推出了名为 Flide 的人工智能推理基础设施,该工具专门针对 Jev 风格的决策模型进行优化 1。Flide 采用“精确状态分叉”技术,宣称其处理速度相较于 vLLM 和 SGLang 等主流框架提升最高可达 9.5 倍 1。项目遵循 Apache-2.0 开源协议,其中部分内核代码衍生自 flash-linear-attention(MIT 许可证) 1。
在性能基准测试中,当在 H200 GPU 上运行 OneJev1.6-0.8B 模型时,单封电子邮件的处理耗时仅为 10.0 毫秒,而批量处理 10,000 封邮件总共仅需 28 秒 1。该软件支持包括 Gemma 4、Qwen3.5 在内的多种主流及专用决策模型 1。最新发布的 Flide 0.1.15 版本于 2026 年 11 月 12 日上线,此次更新新增了 LiquidAI 的 d1-3B 模型支持,并允许 Gemma 4 以 FP8 精度运行 1。此外,相关学术论文《Flide: Accelerating Decision Inference with Exact State Forking》已于 2026 年 11 月 2 日发布于 arXiv 平台(编号 arXiv:2610.42148) 1。
Flide, a new inference infrastructure designed specifically for Jev-style decision models, has been released by its development team on Hacker News 1. The tool claims to achieve speeds up to 9.5 times faster than existing solutions such as vLLM and SGLang by utilizing a technique called "exact state forking" 1. This performance boost is aimed at optimizing decision-making tasks supported by mainstream and specialized models, including Gemma 4 and Qwen3.5 1.
Recent updates have expanded the platform's capabilities and efficiency. Version 0.1.15, released on November 12, 2026, introduced support for LiquidAI’s d1-3B model and enabled Gemma 4 to run in FP8 precision 1. Benchmark data provided for the H200 GPU demonstrates that processing the OneJev1.6-0.8B model takes only 10.0 milliseconds per email, allowing 10,000 emails to be processed in just 28 seconds 1.
The project operates under the Apache-2.0 license, with some kernel code derived from the MIT-licensed flash-linear-attention library 1. Accompanying technical documentation includes the paper "Flide: Accelerating Decision Inference with Exact State Forking," which was published on arXiv (identifier: 2610.42148) on November 2, 2026 1.
评论
还没有评论,欢迎留下第一条。