Event ·Multi-source

Kernel Forge: 用MCTS优化CUDA内核的开源Agent Harness

Kernel Forge是一个开源agent harness，能直接改写PyTorch模型的CUDA内核。它覆盖视觉、扩散和LLM三类工作负载。采用蒙特卡洛树搜索(MCTS)而非线性生成-修复链，并提供GUI监控进度。在NVIDIA DGX Spark（GB10 GPU）上，经过每核50次优化迭代，它在四个模型中优化了14个内核，使其超越PyTorch基线。性能提升主要来自harness结构和原地重整合，而非更强的LLM。

Current read

想又快又稳地优化CUDA内核？Kernel Forge用MCTS自动探索，比手动写靠谱多了，实测14个内核超基线。

11 sources62° AI heatLast updated 2026/7/29 14:55:02

Evidence

11 sources

Primary sourceelvis

2026/7/29 14:55:02View source

Related source 1IT之家

2026/7/30 06:51:42View source

Related source 2Junyang Lin

2026/7/27 14:56:08View source

Related source 3Cognition

2026/7/27 16:26:50View source

Related source 4arXiv cs.AI

2026/7/27 17:08:27View source

Related source 5NVIDIA AI

2026/7/27 22:01:26View source

Related source 6The Rundown AI

2026/7/27 22:44:26View source

Related source 7Aravind Srinivas

2026/7/28 16:40:33View source

Related source 8Amjad Masad

2026/7/28 23:26:11View source

Related source 9Genspark

2026/7/30 20:02:22View source

Related source 10Cohere

2026/7/30 20:53:23View source

Conflict check

Dedup data only shows that these sources discuss the same event, not that they agree. There is no structured supporting or refuting evidence yet, so no conflict is inferred automatically.

Open the story