Back to feed
AR
arXiv CS.AI
7/29/2026
Kernel Forge: An Agent Harness for LLM-based Generation and Optimization of CUDA Kernels

Kernel Forge: An Agent Harness for LLM-based Generation and Optimization of CUDA Kernels

Short summary

Kernel Forge is an open-source agentic harness that automatically generates and optimizes CUDA kernels for any unmodified PyTorch model using Monte Carlo Tree Search. It supports vision, diffusion, and LLM workloads, and includes a GUI for monitoring and debugging. With 50 optimization iterations per kernel, it achieves 1.5x–2.8x speedups over PyTorch eager mode across multiple models.

  • •Open-source agentic harness for automatic CUDA kernel generation and optimization
  • •Uses MCTS to explore multiple optimization paths rather than linear refinement
  • •Achieves 1.52x–2.83x speedups over PyTorch eager mode on ResNet-50, Stable Diffusion 3.5, Gemma 4, and Qwen 3.5

Generated with AI, which can make mistakes.

#open-source#ai-tools#ai-agents#research-breakthrough
Read full article at arXiv CS.AI

Is this a good recommendation for you?

Comments

Failed to load comments. Please try again.

Explore more

Fused Kernels in LLMs: Reducing Memory Bandwidth Bottlenecks Through GPU Kernel Fusion
Dev.to
From CUDA to MLX: How K-Search Brings Decades of Kernel Expertise to Apple Silicon
Berkeley BAIR
InferenceBench: A Benchmark for Open-Ended LLM Inference Optimization by AI Agents
arXiv CS.AI
MawForge: Memory-Bounded Expert Materialization for Local Mixture-of-Experts Inference
arXiv cs.LG