KernelAgent Artifacts

Curated, minimal artifacts showcasing KernelAgent/Fuser outputs across L1/L2/L3.

Source repo: https://github.com/meta-pytorch/KernelAgent Blog post: https://pytorch.org/blog/kernelfalcon-autonomous-gpu-kernel-generation-via-deep-agents/

Layout

• L1/ — selected Level 1 problems (original + generated Triton kernel).

• L2/ — selected Level 2 problems (original + generated Triton kernel).

• L3/ — selected Level 3 problems:

  • Fuser route: original + fused.py + subgraphs.json + per-subgraph kernels + composed_kernel.py + verify logs.

  • KernelAgent route: original + final_kernel.py + test.py + result.json.

• kernelfalcon-artifacts/ — tiny examples for quick browsing.

Minimal Examples

• L1-examples/matmul/ — original + generated Triton kernel + short PASS log.

• L2-examples/conv-bn-relu/ — original + generated Triton kernel (KernelAgent route).

• L3-examples/resnet-block/ — input_model + subgraphs.json + composed_kernel + verify log (Fuser route).

Reproduction

Artifacts were produced using the KernelAgent repo: https://github.com/meta-pytorch/KernelAgent

• Auto-router: python -m Fuser.auto_agent --problem <abs/path/to/problem.py> --verify

• Full pipeline: python -m Fuser.pipeline --problem <abs/path> --extract-model gpt-5 --dispatch-model o4-mini --compose-model o4-mini --dispatch-jobs auto --verify

• KernelAgent direct (Python): see triton_kernel_agent in the main repo.

Provenance

Each example folder contains a manifest.json with route, files, and verification info.

We trimmed logs for readability; full run artifacts live under .fuse/run_* and agent sessions in the main workspace and can be reproduced with the commands above.

Notes

• Kernels here are Triton implementations (no PyTorch compute helpers in wrappers).

• Verification gates success on execution-based checks with tolerances (default rtol=1e-3, atol=1e-3; caps for fp16/bf16 at 1e-2).