[saqibkh]

The hosted runners have no GPU, so test_gpu_matches_cpu is always skipped (cuda_available() is False) and the Mandelbrot kernel is only compiled, never run. This adds a separate, optional workflow .github/workflows/gpu-sim.yml that builds the extension and runs the tests on a simulated NVIDIA H100 using PantheonSim (no GPU needed). I am the author of PantheonSim, so please weigh this accordingly.

It does not touch the existing workflows and is not meant to be a required check. A simulator cannot reproduce hardware-specific behavior or timing, so it complements real-GPU testing and does not replace it.

What it does:

• installs the CUDA 13.0 toolkit and the simulated H100
• builds with pip install . --group test, passing CMAKE_CUDA_ARCHITECTURES=90-virtual (the simulator runs the embedded PTX)
• the simulator needs a shared CUDA runtime, and CMakeLists.txt sets CUDA_RUNTIME_LIBRARY Static on the target, so the job changes that one word with sed in its own checkout; the repository files are not modified
• runs pytest -rs tests under vgpu run
• runs on manual dispatch and on pull requests to master that touch src/, tests/, CMakeLists.txt or pyproject.toml, with a concurrency group that cancels superseded runs

Notes:

• Run on my fork: https://github.com/saqibkh/cuda_example/actions/runs/42658608952 . All 3 tests passed with none skipped, including test_gpu_matches_cpu, which compares the GPU image with the CPU image.
• That run took 15m42s: about 15 minutes for the simulator and toolkit setup, 15 seconds for the build and install, and about 1 second for the tests.