AIBRIEF
The open source tool searches a tree of optimizations and returns working CUDA kernels for unmodified PyTorch models, with reported gains on vision, image, and language workloads.
Kernel Forge, an open-source tool released this month, automatically writes the low-level GPU code, called kernels, that machine-learning models run on. It takes an unmodified PyTorch model without code changes and uses an LLM-driven agent to generate and refine candidate kernels, with a graphical interface for inspecting each attempt (Kernel Forge paper).
Prior agentic-kernel tools ran a single refinement chain. Kernel Forge uses Monte Carlo Tree Search to explore many candidate optimizations in parallel and pick the best. With only 50 optimization iterations per kernel on a single NVIDIA DGX Spark, it produced 14 kernels that beat PyTorch's default code.
Reported per-kernel speedups versus PyTorch's default code: 1.52x on adaptive_avgpool2d inside ResNet-50, 1.70x on group_norm inside Stable Diffusion 3.5 Medium, 2.83x on softmax inside Gemma 4 E2B, and 1.54x on softmax inside Qwen 3.5 35B-A3B. The gains are author-reported on a single GPU and have not been independently replicated.
The tool positions itself against earlier agentic-kernel systems that were evaluated on random tensors, emitted standalone CUDA that developers had to reintegrate, and offered limited debugging. By taking stock PyTorch in and returning working kernels out, Kernel Forge targets ML teams that lack dedicated CUDA engineers, not engineers themselves.
The abstract does not specify when the code will be released or whether the per-kernel gains hold against hand-tuned libraries like cuDNN or Triton.