Repository navigation
test(ci): enable CUDA GPU execution validation in GPU CI #1487
Description
Description
Enable the CUDA GPU execution validation test in GPU CI once a published, immutable CUDA validation image reference is available.
Issue #1472 adds the GPU validation harness and CUDA execution test. That test can run locally when OPENSHELL_E2E_GPU_CUDA_WORKLOAD_IMAGE points at a suitable image, but it should visibly skip when the variable is unset. This follow-up wires the test into CI so CUDA execution is actually enforced by the GPU e2e workflow.
Context
Parent roadmap: #1444
Depends on: #1472
Related image artifacts: #1476 and PR #1484
PR #1484 defines local GPU validation image artifacts, but does not publish stable image references for CI. This issue should be completed after there is a published CUDA validation image that CI can pull by immutable tag or digest.
Proposed Scope
- Choose or publish a stable CUDA validation image reference for CI.
- Configure GPU CI to set
OPENSHELL_E2E_GPU_CUDA_WORKLOAD_IMAGEto that image. - Ensure the Docker GPU e2e workflow runs the CUDA execution validation path, not only the skip path.
- Keep local/unconfigured runs allowed to skip when the env var is unset.
Definition of Done
- A CUDA validation image is available to CI by immutable tag or digest.
- GPU CI sets
OPENSHELL_E2E_GPU_CUDA_WORKLOAD_IMAGE. - GPU CI runs the CUDA execution validation added by Define GPU validation tests for GPU-enabled drivers #1472.
- GPU CI fails if the workload image exits unsuccessfully.
- GPU CI fails if the workload image output does not contain
OPENSHELL_GPU_WORKLOAD_SUCCESS. - Documentation names the CI image reference and explains how local developers can reproduce the same validation.
Activity
github-actions commentedon Jun 14, 2026
This issue has had no activity for 14 days and is now marked stale. It may be closed in 7 days if there is no further activity. Comment or remove the state:stale label to keep it open.
elezar commentedon Aug 18, 2026
Closing as completed: GPU CI now validates real CUDA workloads.
The reusable GPU E2E workflow runs on ARM64 L4 and AMD64 RTX Pro 6000 runners (with an experimental WSL lane), verifies Docker CDI and nvidia-smi, builds the GPU workload manifest, and runs the Docker GPU E2E suite. The manifest includes cuda-basic, which runs CUDA deviceQuery and vectorAdd and requires OPENSHELL_GPU_WORKLOAD_SUCCESS. The current workflow therefore meets this issue’s core intent: CI exercises GPU workload execution and fails on workload or success-marker failure.
One possible future follow-up is artifact provenance: this workflow builds its validation images during CI instead of consuming a separately published immutable CUDA workload image via OPENSHELL_E2E_GPU_CUDA_WORKLOAD_IMAGE. That is not needed to establish GPU workload coverage, but can be reconsidered if reproducibility or independent image lifecycle becomes important.