OpenBMB Open-Sources ForgeKernel: DeepSeek V4.1 Prefill Kernel 9.8% Faster
OpenBMB announced the open-sourcing of ForgeKernel; its agent completed 42 iterations in about 22 hours and produced 32 kernel versions, ultimately achieving 1.098x performance on the DeepSeek V4.1 FlashMLA Prefill on an NVIDIA B200, at 472 µs versus the 518 µs baseline. The official FlashMLA implementation remained closed throughout; the agent could only run and time its black-box CLI and could not read source code, documentation, or profiling data. The open-sourced content includes the optimized DeepSeek V4.1 fused prefill kernel, the kernels from the paper covering attention, Gated DeltaNet, Kimi Delta Attention, GEMM, and convolution, as well as the ForgeKernel harness that can be used to optimize one's own kernels; across 27 operators, ForgeKernel matched or outperformed 18 of 22 expert-library kernels and outperformed all 7 kernels published by three kernel-generation systems.
Why read this
AI score 60
OpenBMB has open-sourced an agent toolchain that automatically optimizes kernels, along with measured speedup data on DeepSeek V4.1 relative to the FlashMLA baseline.
Topics
View original postForgeKernel Live: Optimising a kernel of DeepSeek https://t.co/vYaS4apfT3
View original
ForgeKernel Live: Optimising a kernel of DeepSeek https://t.co/vYaS4apfT3
Body
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