KernelGeneration

KernelGeneration is a repository created to evaluate GPU kernel metrics, including:

• correctness
• performance
• portability

Requirements

• PyTorch nightly (CUDA or ROCm). Install the nightly that matches your system before running the installer.

  • CUDA example:

    pip install --pre torch torchvision --index-url https://download.pytorch.org/whl/nightly/cu124

  • ROCm example:

    pip install --pre torch torchvision --index-url https://download.pytorch.org/whl/nightly/rocm6.4

  Use the official PyTorch install selector to pick the correct nightly wheel for your OS/driver stack.

• (Optional) KernelLLM prompts
  If you want to try KernelLLM-style prompting, see the templates in ./prompt/.
  Note: the installer does not clone KernelLLM or install transformers/accelerate.

Environment

After setup, export:

export TRITONBENCH_RUN_CONFIG="$(pwd)/benchmark_helion_runner.yaml"
export TRITONBENCH_HELION_PATH="$(pwd)/helion"

Project Structure (before running install.py)

kernelGen/
├── README.md
├── install.py
├── benchmark_helion_runner.yaml
├── prompt/
│   └── ...                # optional prompt templates for KernelLLM-style tests
└── operators/
    ├── bf16_layernorm/
    └── bf16_matmul/

• install.py — installs TritonBench, clones Helion (kernel-gen-rh branch) and installs it in editable mode, and copies local operators into TritonBench.
  It does not install PyTorch or KernelLLM.
• operators/ — GPU kernel operators (copied into TritonBench during install).

Evaluated Kernels

• MatMul
• LayerNorm — WIP (optional KernelLLM prompts available in ./prompt/)
• WIP: GELU

Backends Tested

• TorchInductor
• Triton
• Helion
• KernelLLM (optional, prompts only via ./prompt/)
• WIP: Mako (based on KernelLLM)

Tested on

• TritonBench

Usage

1. Install PyTorch nightly (see Requirements above).
2. Run the installer (clones TritonBench, clones & installs Helion, and copies operators):

   python install.py

3. Export environment variables (adjust paths as needed):

   export TRITONBENCH_RUN_CONFIG="$(pwd)/benchmark_helion_runner.yaml"
   export TRITONBENCH_HELION_PATH="$(pwd)/helion"

4. Run the benchmark:

   python tritonbench/run.py