CI: give GPU unit-test containers a working NVIDIA Vulkan driver

[swahtz]

Summary

tests/unit/test_viz_shutdown.py::test_viewer_shutdown_with_gpu_buffers, added in #810, needs a Vulkan device. On the GPU runners it fails with Failed to create Vulkan instance. VkResult: -9 (VK_ERROR_INCOMPATIBLE_DRIVER). #810's unit-test jobs were skipped, so #812 is the first PR to hit it. Every PR based on current main will fail the same way.

This PR gives the three GPU unit-test jobs a working NVIDIA Vulkan driver.

Cause

The runner AMI is fine. Its 580.65.06 driver includes the Vulkan driver, and vulkaninfo on the host lists the L4. The unit-test containers could not load that driver, for two reasons:

1. The CUDA images set NVIDIA_DRIVER_CAPABILITIES=compute,utility. The container toolkit therefore does not mount the Vulkan driver (libGLX_nvidia and /etc/vulkan/icd.d/nvidia_icd.json).
2. The driver depends on libraries the slim images lack.
   • libGLX_nvidia links against libXext.so.6.
   • It also opens libEGL.so.1 at run time.
   • Without either one, the driver loads but returns no vkCreateInstance, and instance creation fails with -9.

Change

The fvdb Unit Tests jobs in tests.yml (conda), cu130.yml and cu132.yml now set up Vulkan in their containers:

• All three jobs: set NVIDIA_DRIVER_CAPABILITIES: all.
• pip jobs: install libvulkan1 libxext6 libegl1.
• conda job:
  • install libXext libglvnd-egl;
  • set VK_DRIVER_FILES=/etc/vulkan/icd.d/nvidia_icd.json, in case the conda-provided Vulkan loader doesn't search /etc/vulkan.

The gtest and docs jobs are unchanged.

Test plan

• Diagnosed by hand on a runner instance from the CI AMI (Ubuntu 24.04, driver 580.65.06, L4, container toolkit 1.17.8):
  • The stock nvidia/cuda:13.0.2-cudnn-runtime-ubuntu22.04 container has no NVIDIA Vulkan driver files.
  • With NVIDIA_DRIVER_CAPABILITIES=all but without libxext6, the loader skips the driver (libXext.so.6: cannot open).
  • With libxext6 but without libegl1, the driver loads but returns no vkCreateInstance.
  • With libvulkan1 libxext6 libegl1, vulkaninfo --summary lists NVIDIA L4.
• CI: these workflows run on pull_request_target, which uses main's workflow definitions. This PR's own unit-test jobs will therefore still fail on the viewer test. The fix takes effect once it merges; the next PR, or a rerun of Support empty JaggedTensor structures and validate raw constructors #812, should then pass test_viz_shutdown.
• The conda job's VK_DRIVER_FILES setting is untested. If that job still fails after merge, its loader output will show whether the variable is needed or not enough.

🤖 Generated with Claude Code