cuda.core: Fix graphics tests by juenglin · Pull Request #2701 · NVIDIA/cuda-python

juenglin · 2026-08-25T23:47:06Z

Description

The GL context helpers in cuda_core/tests/test_graphics.py and cuda_bindings/tests/test_graphics_apis.py wrapped yield in except Exception and called pytest.skip(), so CUDA errors, assertion failures, and TypeErrors raised in the test body were recorded as "could not create GL context" skips rather than failures. The cuda_core graphics tests also never requested init_cuda, so whether from_gl_buffer / from_gl_image succeeded depended on leftover thread state from the previous test (pytest-randomly seed).

This PR:

Splits each helper into a plain setup function (_configure_pyglet_headless, _open_gl_window, _setup_gl_buffer, _setup_gl_texture) and a thin context manager that wraps only the setup call in try/except so test-body exceptions propagate as failures.
Adds init_cuda to every cuda_core graphics test that touches CUDA, and uses init_cuda.create_stream() instead of a local _create_stream() helper.
Converts the cuda_core test classes to plain test functions (they carried no shared state or fixtures).
Fixes two StridedMemoryView.from_buffer call sites that passed dtype=np.float32 (a scalar type class) instead of np.dtype(np.float32).
Marks test_close_while_mapped_passes_stream_override as xfail(strict=True) because Buffer is an immutable Cython type and neither patch.object nor __class__ assignment can intercept Buffer.close.

Checklist

New or existing tests cover these changes.
The documentation is up to date with these changes.

copy-pr-bot · 2026-08-25T23:47:09Z

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

Stop treating graphics test-body failures as GL-unavailable skips, and bind a CUDA context in those tests so missing thread state cannot hide real errors.

juenglin · 2026-08-26T02:46:55Z

/ok to test baf3e23

juenglin · 2026-08-26T14:51:03Z


-            with patch.object(Buffer, "close", new=tracking_close):
-                resource.close(stream=close_stream)
+@pytest.mark.xfail(


I couldn't get this test to work. Should I remove it?

============================================================= short test summary info ============================================================= FAILED cuda_core/tests/test_graphics.py::test_close_while_mapped_passes_stream_override - TypeError: cannot set 'close' attribute of immutable type 'cuda.core._memory._buffer.Buffer' ========================================================== 1 failed, 29 passed in 0.63s ===========================================================

Yes, I agree with removing it. The stream behavior of Buffer.close() is already covered elsewhere.

juenglin · 2026-08-26T23:55:26Z

/ok to test 8fbb9cf

juenglin · 2026-08-27T14:36:14Z

pre-commit.ci run

juenglin · 2026-08-27T15:59:14Z

/ok to test 5dc9814

Andy-Jost

Looks good. The structure is improved and the main problem is addressed. Found a few minor issues to optionally clean up.

Andy-Jost · 2026-08-27T18:31:57Z

Were the changes to this file intentional?

Yes, I updated the lock file after adding the missing pyglet test dependency.

Andy-Jost · 2026-08-27T18:37:33Z

+def _setup_gl_buffer(pyglet, nbytes):
+    """Open a GL context and allocate a buffer. Returns (win, buf_id)."""
+    win = _open_gl_window(pyglet)
+    from pyglet.gl import gl as _gl
+
+    buf_id = _gl.GLuint(0)
+    _gl.glGenBuffers(1, ctypes.byref(buf_id))
+    _gl.glBindBuffer(_gl.GL_ARRAY_BUFFER, buf_id.value)
+    _gl.glBufferData(_gl.GL_ARRAY_BUFFER, nbytes, None, _gl.GL_DYNAMIC_DRAW)
+    return win, buf_id


Optional: if setup fails after creating win or buf_id, those resources may leak. Low severity.

Andy-Jost · 2026-08-27T18:37:46Z

+def _setup_gl_texture(pyglet, width, height):
+    """Open a GL context and allocate a 2-D RGBA8 texture. Returns (win, tex_id, target)."""
+    win = _open_gl_window(pyglet)
+    from pyglet.gl import gl as _gl
+
+    tex_id = _gl.GLuint(0)
+    _gl.glGenTextures(1, ctypes.byref(tex_id))
+    target = _gl.GL_TEXTURE_2D
+    _gl.glBindTexture(target, tex_id.value)
+    _gl.glTexParameteri(target, _gl.GL_TEXTURE_MIN_FILTER, _gl.GL_NEAREST)
+    _gl.glTexParameteri(target, _gl.GL_TEXTURE_MAG_FILTER, _gl.GL_NEAREST)
+    _gl.glTexImage2D(target, 0, _gl.GL_RGBA8, width, height, 0, _gl.GL_RGBA, _gl.GL_UNSIGNED_BYTE, None)
+    return win, tex_id, target


Optional: similarly, a failure after creating win or tex_id may leave those resources uncleaned. Low severity.

Andy-Jost · 2026-08-27T18:42:51Z


-            with patch.object(Buffer, "close", new=tracking_close):
-                resource.close(stream=close_stream)
+@pytest.mark.xfail(


Yes, I agree with removing it. The stream behavior of Buffer.close() is already covered elsewhere.

Andy-Jost · 2026-08-27T18:46:03Z

+def _setup_gl_texture(pyglet):
+    """Open a GL context and allocate a 2-D RGBA8 texture. Returns (win, tex_id, target)."""
+    if not pyglet.options.get("headless"):
+        # Hidden window path (WGL on Windows, GLX/WLS on Linux)
+        from pyglet import gl
+
+        config = gl.Config(double_buffer=False)
+        win = pyglet.window.Window(visible=False, config=config)
+        win.switch_to()
+    else:
+        # Headless EGL path; pyglet will arrange a pbuffer-like headless context
+        from pyglet.gl import headless  # noqa: F401
+
+        win = None
+
+    # Make a tiny texture so we have a real GL object to register
+    from pyglet.gl import gl as _gl
+
+    tex_id = _gl.GLuint(0)
+    _gl.glGenTextures(1, ctypes.byref(tex_id))
+    target = _gl.GL_TEXTURE_2D
+    _gl.glBindTexture(target, tex_id.value)
+    _gl.glTexParameteri(target, _gl.GL_TEXTURE_MIN_FILTER, _gl.GL_NEAREST)
+    _gl.glTexParameteri(target, _gl.GL_TEXTURE_MAG_FILTER, _gl.GL_NEAREST)
+    width, height = 16, 16
+    _gl.glTexImage2D(target, 0, _gl.GL_RGBA8, width, height, 0, _gl.GL_RGBA, _gl.GL_UNSIGNED_BYTE, None)
+    return win, tex_id, target


Optional: same potential leak issue here

juenglin · 2026-08-27T20:02:14Z

Thanks. I have a few more things for a follow-on PR and will address your comments there.

github-actions · 2026-08-27T21:22:33Z

Doc Preview CI
Preview removed because the pull request was closed or merged.

github-actions Bot added the cuda.core Everything related to the cuda.core module label Aug 25, 2026

juenglin added bug Something isn't working P0 High priority - Must do! test Improvements or additions to tests labels Aug 25, 2026

juenglin added this to the cuda.core 1.2.0 milestone Aug 25, 2026

juenglin self-assigned this Aug 25, 2026

cuda.core: Fix graphics tests

d45b987

Stop treating graphics test-body failures as GL-unavailable skips, and bind a CUDA context in those tests so missing thread state cannot hide real errors.

juenglin force-pushed the graphics-tests branch from f348d94 to d45b987 Compare August 26, 2026 00:02

juenglin requested review from Andy-Jost and rparolin August 26, 2026 00:07

juenglin added 2 commits August 25, 2026 19:44

fix memory view tests

409cb25

defer fix for close_stream test

baf3e23

remove redundant stream helper

3e00825

This comment has been minimized.

Sign in to view

juenglin commented Aug 26, 2026

View reviewed changes

juenglin marked this pull request as ready for review August 26, 2026 14:51

fix broken helpers in corresponding bindings tests

8fbb9cf

github-actions Bot added the cuda.bindings Everything related to the cuda.bindings module label Aug 26, 2026

juenglin added 2 commits August 27, 2026 08:15

accomodate drivers refusing CUDA-GL interop

b004320

add missing test dependency

5dc9814

juenglin force-pushed the graphics-tests branch from c095715 to 5dc9814 Compare August 27, 2026 15:47

Andy-Jost approved these changes Aug 27, 2026

View reviewed changes

juenglin merged commit 190527c into NVIDIA:main Aug 27, 2026
111 checks passed

juenglin mentioned this pull request Aug 27, 2026

cuda.core: fix a few test issues #2714

Open

2 tasks

juenglin added a commit to juenglin/cuda-python that referenced this pull request Aug 27, 2026

address PR NVIDIA#2701 feedback

cccbbfe

Uh oh!

Conversation

juenglin commented Aug 25, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Description

Checklist

Uh oh!

copy-pr-bot Bot commented Aug 25, 2026

Uh oh!

juenglin commented Aug 26, 2026

Uh oh!

This comment has been minimized.

juenglin Aug 26, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Andy-Jost Aug 27, 2026

Choose a reason for hiding this comment

Uh oh!

juenglin commented Aug 26, 2026

Uh oh!

juenglin commented Aug 27, 2026

Uh oh!

juenglin commented Aug 27, 2026

Uh oh!

Andy-Jost left a comment

Choose a reason for hiding this comment

Uh oh!

Andy-Jost Aug 27, 2026

Choose a reason for hiding this comment

Uh oh!

juenglin Aug 27, 2026

Choose a reason for hiding this comment

Uh oh!

Andy-Jost Aug 27, 2026

Choose a reason for hiding this comment

Uh oh!

Andy-Jost Aug 27, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Andy-Jost Aug 27, 2026

Choose a reason for hiding this comment

Uh oh!

Andy-Jost Aug 27, 2026

Choose a reason for hiding this comment

Uh oh!

juenglin commented Aug 27, 2026

Uh oh!

Uh oh!

github-actions Bot commented Aug 27, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

juenglin commented Aug 25, 2026 •

edited

Loading

juenglin Aug 26, 2026 •

edited

Loading

Andy-Jost Aug 27, 2026 •

edited

Loading