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

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。

Impressive paper! It's on one of the hardest tasks for coding agents today. Of course, I am talking about kernel optimization. Coding agents are usually not so great at this. Reasons: Unfamiliar low-level API, no room for a plausible-looking answer, and every candidate has to actually run and be faster. Kernel Forge is an open-source agent harness that takes an unmodified PyTorch model in place and rewrites its CUDA kernels. It covers vision, diffusion, and LLM workloads. Instead of a linear generate-and-fix chain, it runs Monte Carlo Tree Search over multiple optimization paths, with a GUI for monitoring progress, inspecting candidate kernels, and debugging failures. On an NVIDIA DGX Spark with a GB10 GPU, at 50 optimization iterations per kernel, it optimized 14 kernels past their PyTorch baselines across four models. It looks like the gain came from harness structure and in-place reintegration rather than a stronger model. That lesson repeats for any agent working against an API it was never trained on. Paper: arxiv.org/abs/2607.24762 Learn to build effective AI agents in our academy: academy.dair.ai📊 14 ⚡

This report is categorized under AITOP's "Papers" desk. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows. Key entities and technical topics under continuous tracking include Kernel Forge, CUDA, 代码优化, 论文.

Source attribution: curated from elvis (@omarsar0). Cross-source corroboration: 11 independent sources have covered this development (including IT之家, Junyang Lin, Cognition), confirming industry-wide signal strength.