Render Product Device Pinning

RenderProduct Device Selection

When to Use

Use this skill when a scene, test, or example must constrain a RenderProduct to CUDA-visible GPU 0 or another explicit set of CUDA-visible GPUs, especially for multi-GPU CI or viewport picking.

Inputs

Resolve inputs in this order: existing repository files and referenced snippets, explicit user request, then broader agent context.

• Target API surface: Python, C/C++, USD, or a combination.
• RenderProduct prim path or scene file that needs a deviceIds restriction.
• CUDA-visible device index allow-list, usually [0] for picking or deterministic multi-GPU CI.
• Whether the pinning is authored in static USDA, inline USD, or an existing example/test scene.
• Repository source snippets referenced below. Treat these snippets as the API source of truth.

Prerequisites

• Use an ovrtx checkout that contains the referenced examples and docs tests.
• Read the relevant > **Source:** snippet before writing or explaining API usage.
• Confirm device indices are CUDA-visible indices after CUDA_VISIBLE_DEVICES filtering, not physical GPU IDs.
• Pin the RenderProduct only; Camera and RenderVar prims do not own deviceIds.

Instructions

1. Identify the RenderProduct path and the CUDA-visible device indices it should be allowed to use.
2. Author or update the RenderProduct deviceIds attribute in USD; do not pin Camera or RenderVar prims.
3. Interpret device IDs after CUDA_VISIBLE_DEVICES remapping, not as physical PCI bus or global GPU IDs.
4. For picking workflows, pin the relevant RenderProduct to CUDA-visible GPU 0 until the picking limitation is removed.
5. When changing code, run the picking or render-product docs test that owns the GPU-pinning snippet whenever practical.

Output Format

• For explanations, cite the relevant API names, source snippets, and caveats.
• For code changes, summarize the files changed, snippets affected, and validation run.

Scripts

This skill has no scripts.

Limitations

• For picking, pin RenderProducts to CUDA-visible GPU 0 until the current picking limitation is removed.
• The referenced snippets remain the source of truth; update or add tested snippets before documenting new API usage.

Overview

ovrtx can select active CUDA devices at renderer creation and can also restrict an individual RenderProduct to a specific device list. Author uint[] deviceIds on the RenderProduct prim before the stage is loaded.

deviceIds is an allow-list of indices into the CUDA-visible device set, after CUDA_VISIBLE_DEVICES filtering and remapping. ovrtx is free to choose any CUDA-visible GPU from the list. A one-element list such as [0] restricts the RenderProduct to CUDA-visible GPU 0.

Use this when a workflow is device-sensitive, needs to constrain eligible GPUs, or must run on CUDA-visible GPU 0. In the current ovrtx version, viewport picking only works for RenderProducts running on CUDA-visible GPU 0, so picking tests and examples should restrict their picking RenderProduct to deviceIds = [0].

USDA

Restrict a RenderProduct to CUDA-visible GPU 0

Source: tests/docs/data/ovrtx-test-picking-selection.usda snippet doc-pin-render-product-to-gpu-0-usda

Workflow

1. Choose the CUDA-visible device indices that may render the product.
2. Author uint[] deviceIds directly on the RenderProduct prim in USD.
3. If using renderer-level active_cuda_gpus, ensure the RenderProduct allow-list is compatible with that active device set.
4. Load the scene normally from Python or C.
5. Step the constrained RenderProduct path as usual.

For picking, use CUDA-visible GPU 0 until the current picking limitation is removed.

Key Attribute

Troubleshooting

• deviceIds is authored on the RenderProduct prim, not on the renderer config object, camera, or RenderVar.
• deviceIds indexes CUDA-visible devices, not non-CUDA devices, Vulkan device indices, display adapters, or physical GPU ids masked out by CUDA_VISIBLE_DEVICES.
• deviceIds is an allow-list, not a load-balancing or exact-placement policy. Exact device selection requires a singleton list such as [0]; [0, 1] permits ovrtx to choose either listed CUDA-visible GPU.
• Renderer-level CUDA selection and per-RenderProduct deviceIds must agree. For picking, leave active_cuda_gpus unset or ensure it includes CUDA-visible index 0.
• Author the attribute before loading the scene. Do not rely on runtime mutation for tests that need deterministic device assignment during renderer setup.
• Picking currently requires the picked RenderProduct to run on CUDA-visible GPU 0. In multi-GPU CI, omit the CUDA-visible GPU 0 restriction only if the test does not depend on pick results.

Related Skills

• picking-selection for viewport picking, marquee selection, and selection outline drawing.
• renderer-creation for renderer-level CUDA device selection.
• cuda-interop for mapping a selected CUDA-visible ordinal to the exact Vulkan physical device, including sibling MIG instances.

References

• Use the > **Source:** directives in this skill to locate tested snippets before reusing API patterns.
• Keep related skills, docs, and snippets synchronized when changing the workflow.