Kernel Forge uses MCTS to optimize CUDA kernels and beats PyTorch baselines on 14 cases
omarsar0 · x · 2026-07-29
A paper introduces Kernel Forge, an open-source agent harness that lets an unmodified PyTorch model generate and optimize CUDA kernels for vision, diffusion, and LLM workloads.
Key ideas:
- It replaces a simple generate-and-fix loop with Monte Carlo Tree Search over multiple optimization paths.
- It ships with a GUI for monitoring progress, inspecting candidate kernels, and debugging failures.
- On an NVIDIA DGX Spark with a GB10 GPU, 50 optimization iterations per kernel produced 14 kernels that beat their PyTorch baselines across four models.
- Reported gains include 1.52× on adaptiveavgpool2d in ResNet-50, 1.70× on groupnorm in Stable Diffusion 3.5 Medium, 2.83× on softmax in Gemma 4 E2B, and 1.54× on softmax in Qwen 3.5 35B-A3B.
More from coding & agent
- Sightmap open-sources a YAML runtime memory map for AI coding agents — hwchase17 · 2026-07-29
- Developers are pointing Claude straight at the database and skipping the usual stack — EricBuess · 2026-07-29
- Claude Opus 5 still needed iterations to build a custom racing game — mattshumer_ · 2026-07-29
- Verdent says Kimi K3 is tuned for real-world agentic coding workflows — socialwithaayan · 2026-07-29
- Hermes agent now connects to Buzz for workflow automations — intellectronica · 2026-07-29
- How to wire custom MCP servers into ChatGPT and Claude chat interfaces — jeremyjordan · 2026-07-29