Pytest CUDA Marker

[jeff-hykin]

GPU tests are different than cuda tests (most GPU tests run on MacOS, its only the cuda ones that dont)

[jeff-hykin]

NOTE: this one doesnt require cuda, even though its still metric3d (only some of the metric3d tests need cuda)

[greptile-apps]

Greptile Overview

Greptile Summary

This PR introduces a new @pytest.mark.cuda marker to distinguish between general GPU tests (which can run on macOS with Metal) and CUDA-specific tests (which require NVIDIA CUDA). The implementation adds a pytest hook in conftest.py that automatically skips CUDA tests when torch.cuda.is_available() returns False.

Key changes:

• Added _has_cuda() helper and pytest_collection_modifyitems hook to auto-skip CUDA tests
• Registered cuda marker in pyproject.toml pytest configuration
• Applied @pytest.mark.cuda to 3 Metric3D tests and 1 ContactGraspNet test that require CUDA

Issues found:

• Two Metric3D tests (test_metric3d_update_intrinsic and test_metric3d_update_intrinsic_invalid) are missing the @pytest.mark.cuda marker even though they instantiate Metric3D, which defaults to device="cuda". These tests will fail on non-CUDA systems.

Confidence Score: 3/5

• This PR is mostly safe but has incomplete marker coverage that could cause test failures
• The implementation is correct and the pytest hook works properly, but two tests in test_metric3d.py are missing the cuda marker despite requiring CUDA. This inconsistency means these tests will fail on non-CUDA systems instead of being properly skipped.
• Pay close attention to dimos/models/depth/test_metric3d.py - ensure all tests that instantiate Metric3D have the cuda marker

Important Files Changed

Filename	Overview
dimos/conftest.py	Added pytest hook to automatically skip tests marked with @pytest.mark.cuda when CUDA is not available
pyproject.toml	Registered new cuda pytest marker for tests requiring CUDA-specific GPU acceleration
dimos/models/depth/test_metric3d.py	Added @pytest.mark.cuda to 3 tests that require CUDA, but 2 other GPU tests still lack the marker despite instantiating Metric3D
dimos/models/manipulation/contact_graspnet_pytorch/test_contact_graspnet.py	Added @pytest.mark.cuda to integration test that requires CUDA for contact graspnet inference

Sequence Diagram

sequenceDiagram
    participant Pytest
    participant conftest
    participant torch
    participant Test

    Note over Pytest: Test Collection Phase
    Pytest->>conftest: pytest_collection_modifyitems(config, items)
    conftest->>conftest: _has_cuda()
    conftest->>torch: import torch
    conftest->>torch: torch.cuda.is_available()
    torch-->>conftest: True/False
    
    alt CUDA not available
        conftest->>conftest: Create skip marker
        loop For each test item
            conftest->>Test: item.get_closest_marker("cuda")
            alt Has cuda marker
                conftest->>Test: item.add_marker(skip_marker)
                Note over Test: Test will be skipped
            end
        end
    else CUDA available
        Note over conftest: No action, tests run normally
    end
    
    Note over Pytest: Test Execution Phase
    Pytest->>Test: Run test_metric3d_init()
    Note over Test: @pytest.mark.cuda<br/>@pytest.mark.gpu
    alt CUDA available
        Test->>Test: Execute test
    else CUDA not available
        Note over Test: Skipped: CUDA is not available
    end

Loading

[greptile-apps]

_{4 files reviewed, 2 comments}

_{Edit Code Review Agent Settings | Greptile}

[greptile-apps]

Missing @pytest.mark.cuda - this test instantiates Metric3D which defaults to device="cuda" (metric3d.py:35), so it requires CUDA to run.

Suggested change

	@pytest.mark.gpu
	@pytest.mark.cuda
	@pytest.mark.gpu

[greptile-apps]

Additional Comments (1)

dimos/models/depth/test_metric3d.py
Missing @pytest.mark.cuda - this test instantiates Metric3D which defaults to device="cuda" (metric3d.py:35), so it requires CUDA to run.

@pytest.mark.cuda
@pytest.mark.gpu