[saqibkh]

The CI here has no job that executes the CUDA code: build.yml only builds wheels and an sdist, because the hosted runners have no GPU. This adds a separate, optional workflow .github/workflows/gpu-sim.yml that builds pytigre with CUDA 13.0 and runs the Python tests that need a GPU on a simulated NVIDIA H100, using PantheonSim (via the pantheongpu/setup-pantheonsim action, pinned by SHA). 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 CUDA 13.0 and the simulated H100, then pip install . and runs pytest under vgpu run.
• The simulator runs PTX only, but setup.py builds -gencode=arch=compute_X,code=sm_X (SASS) for every architecture nvcc lists. The workflow therefore applies a two-line sed to setup.py on the runner (not in the repo) so it builds a single compute_90 PTX target. The extensions already link the shared libcudart, so nothing else needs to change.
• Collects only the test_*.py modules in Python/tests and skips test_config_gen.py. algorithm_test.py, stdout_test.py and test_config_gen.py do not import (tigre.demos is not packaged in the install, and stdout_test.py imports Python 2's StringIO), so they are left out; I did not change them.
• Triggers: workflow_dispatch and pull_request limited to Common/, Python/, setup.py, pyproject.toml and the workflow itself. A concurrency group cancels superseded runs, with permissions: contents: read.

What is left out:

• Only one architecture (H100, PTX for compute_90) and one CUDA version (13.0) are covered.
• The MATLAB interface and the three legacy test scripts above are not exercised.

Result on a fork: the run passed. 70 tests passed and none were skipped, including the 9 Krylov smoke tests (test_krylov_smoke.py, which run Ax and Atb on a 64^3 phantom with 50 views) and the GPU-id tests. The whole job took about 24 minutes on a cold cache, of which about 15.7 minutes was the action setting up the simulator on its first run. The build step took about 48 seconds and the tests about 7 minutes.

Run: https://github.com/saqibkh/TIGRE/actions/runs/42660482796

[AnderBiguri]

Thanks @saqibkh ! I will leave this open for a bit, as I want to first to a push of making a good CI/tests for TIGRE (hopefully less computationally costly) and then see if this CI idea of simulating a GPU is worth doing for TIGRE. This is nevertheless quite interesting, so thanks.

[saqibkh]

Thanks, no rush. If it helps, I can adjust the workflow to whatever test set you settle on.