initial commit of OptArena #1

	name: agentbench-gpu

	# Runs the agent_bench suite on a real NVIDIA GPU so the device-residency CUDA
	# path executes for real (GitHub-hosted runners have no GPU -- this needs a
	# self-hosted runner labelled [self-hosted, gpu] with the CUDA toolkit/nvcc and
	# a recent driver). Adjust the `runs-on` labels to match your runner.

	on:
	push:
	branches: [ extended, main ]
	pull_request:
	branches: [ extended, main ]

	jobs:
	gpu-agentbench:
	runs-on: [self-hosted, gpu]
	steps:
	- uses: actions/checkout@v4

	- name: GPU + toolchain sanity
	run: \|
	nvidia-smi
	nvcc --version
	gcc --version \| head -1
	gfortran --version \| head -1

	- name: Install
	run: \|
	python -m pip install --upgrade pip
	python -m pip install -r requirements/nvidia.txt
	python -m pip install jinja2 tree-sitter tree-sitter-languages pytest
	python -m pip install -e .

	- name: cupy attaches to the device
	run: \|
	python -c "import cupy; n=cupy.cuda.runtime.getDeviceCount(); assert n>0, n; print('devices', n, 'cc', cupy.cuda.Device(0).compute_capability)"

	- name: agent_bench tests (device-residency CUDA runs for real)
	# On this runner the cupy/nvcc gates are satisfied, so the GPU
	# device-residency test executes instead of skipping.
	run: python -m pytest tests/test_agent_bench.py -v

	- name: Codegen frameworks on the GPU (Triton + TVM-GPU execute for real)
	# requirements/nvidia.txt installs triton + apache-tvm; with a real GPU
	# present the triton / tvm (cuda) framework tests run instead of skipping.
	env:
	OPTARENA_TVM_NOTUNE: "1"
	run: python -m pytest tests/test_frameworks.py -v -rs

Navigation Menu