add support for injecting /dev/dri* devices nodes for gfx MIGs

[tariq1890]

Currently, the container toolkit skips adding the /dev/dri device nodes to a device if it's found to be a MIG device. Today, with the availability of GPUs like RTX Pro 6000D Blackwell, MIG slices with graphics capabilities can be created.

In this PR, we update the NVSandboxutils and NVML discoverers to check if the detected MIG device supports graphics and inject the /dev/dri device nodes if true.

[coveralls]

Coverage Report for CI Build 25753963182

Coverage decreased (-0.05%) to 43.296%

Details

• Coverage decreased (-0.05%) from the base build.
• Patch coverage: 34 uncovered changes across 3 files (17 of 51 lines covered, 33.33%).
• 21 coverage regressions across 1 file.

Uncovered Changes

File	Changed	Covered	%
internal/platform-support/dgpu/nvml.go	41	11	26.83%
internal/platform-support/dgpu/dgpu.go	5	3	60.0%
internal/platform-support/dgpu/nvsandboxutils.go	5	3	60.0%

Coverage Regressions

21 previously-covered lines in 1 file lost coverage.

File	Lines Losing Coverage	Coverage
cmd/nvidia-cdi-hook/cudacompat/cuda-elf-header.go	21	55.22%

---

Coverage Stats

Relevant Lines:	14902
Covered Lines:	6452
Line Coverage:	43.3%
Coverage Strength:	0.48 hits per line

---

💛 - Coveralls

[elezar]

This looks reasonable @tariq1890. Could we link the related go-nvlib PRs that will be required for this in the description?

One other thing to note. This handles the injection of these devices when nvidiasandbox-utils is used, but not when pure nvml is used. Is this a concern?

[tariq1890]

Thanks for the review @elezar . There are no go-nvlib or go-nvml changes required for this PR. These toolkit changes just work as they are.

Regarding the NVML discoverer, I thought the changes would just work looking at the code. I can try testing that with the nvsandboxutils discoverer disabled

[tariq1890]

I stand corrected :). Changes are indeed needed in the NVML MIG discoverer. I'll make those changes

[tariq1890]

This handles the injection of these devices when nvidiasandbox-utils is used, but not when pure nvml is used. Is this a concern?

This is now addressed. I was able to confirm through local testing that the pure NVML discoverer will inject the /dev/dri* device nodes if the MIG slices support graphics.

[cdesiniotis]

/cherry-pick release-1.19

[github-actions]

🤖 Backport PR created for release-1.19: #1831 ✅