面壁智能开源 ForgeKernel：DeepSeek V4.1 预填充内核提速 9.8%
ForgeKernel Live has wrapped up with a 9.8% speedup over the FlashMLA baseline on DeepSeek V4.1 FlashMLA Prefill, measured on NVIDIA B200. Now, we’re open-sourc
面壁智能宣布开源 ForgeKernel，其智能体在约 22 小时内完成 42 轮迭代、产出 32 个内核版本，最终在 NVIDIA B200 上对 DeepSeek V4.1 FlashMLA Prefill 实现 1.098 倍性能，472 µs 对比基线 518 µs。官方 FlashMLA 实现全程封闭，智能体只能运行并计时其黑盒 CLI，无法读取源码、文档或性能剖析数据。此次开源内容包括优化后的 DeepSeek V4.1 融合预填充内核、论文中覆盖 attention、Gated DeltaNet、Kimi Delta Attention、GEMM 和卷积的内核，以及可用来优化自有内核的 ForgeKernel harness；在 27 个算子上，ForgeKernel 与 22 个专家库内核中的 18 个持平或更优，并优于三个内核生成系统已发布的全部 7 个内核。
查看原帖ForgeKernel 直播：优化 DeepSeek 的一个内核 https://t.co/vYaS4apfT3
查看原文
ForgeKernel Live: Optimising a kernel of DeepSeek https://t.co/vYaS4apfT3
正文
ForgeKernel Live has wrapped up with a 9.8% speedup over the FlashMLA baseline on DeepSeek V4.1 FlashMLA Prefill, measured on NVIDIA B200. Now, we’re open-sourcing ForgeKernel! 🚀
A quick recap: ⌛️ Around 22 hours of autonomous optimization 🛠️ 32 kernel versions across 42 iteration rounds ✅ Version 13 first beat the baseline within 10 hours 📈 Version 31 reached 1.098× performance: 472 µs vs. 518 µs
The official FlashMLA implementation stayed sealed. The agent could run and time its black-box CLI, but could not read its source, documentation, or profiles.
We’re releasing: 1️⃣ The optimized DeepSeek V4.1 fused prefill kernel. 2️⃣ Kernels from our ForgeKernel paper, covering attention, Gated DeltaNet, Kimi Delta Attention, GEMM, and convolution. Across 27 operators, ForgeKernel matches or beats 18 of 22 expert-library kernels and outperforms all 7 published kernels from three kernel-generation systems in our evaluation. 3️⃣ The ForgeKernel harness, so you can optimize your own kernels and explore successive generations of improvement.
Thank you to everyone who followed the live run. We’re excited to see what the community builds next!
🔗 Live: https://t.co/JfiI1jnzLe 💻 Code: https://t.co/C0yX3vxlf7