benchmarks/gpu_mode/vecadd/README.md

GPU Mode: Float16 Vector Addition

Evolve a Triton kernel for float16 vector addition using SkyDiscover.

Operation: C = A + B (element-wise, float16)

Quick Start

From the repo root:

uv run skydiscover-run \
  benchmarks/gpu_mode/vecadd/initial_program.py \
  benchmarks/gpu_mode/vecadd/evaluator.py \
  -c benchmarks/gpu_mode/vecadd/config.yaml \
  -s [your_algorithm] -i 50

Scoring

• Correctness weight: 0.3 (must return float16, rtol/atol=1e-3)
• Speedup weight: 1.0 (geometric mean vs PyTorch reference, capped at 10x)
• Combined: 0.3 * correctness + speedup

Modal Cloud GPU Support

GPUMODE_USE_MODAL=true GPUMODE_MODAL_GPU=H100 \
  uv run skydiscover-run \
  benchmarks/gpu_mode/vecadd/initial_program.py \
  benchmarks/gpu_mode/vecadd/evaluator.py \
  -c benchmarks/gpu_mode/vecadd/config.yaml \
  -s [your_algorithm] -i 50