Detect Jetson GPUs in Linux DRM device discovery

[Copilot]

Description

• Recognize non-PCI DRM devices using the NVIDIA nvgpu/nvidia driver.
• Surface them as integrated NVIDIA GPUs without PCI identity metadata.
• Continue excluding AMD XCP and other non-NVIDIA platform devices.
• Add focused Jetson-style sysfs discovery coverage.

Motivation and Context

Jetson GPUs use the platform/host1x bus and lack PCI vendor attributes, causing Linux device discovery to skip them. Consequently, the CUDA plugin EP is absent from GetEpDevices() despite loading successfully.

• Fixes CUDA plugin EP GetEpDevices() finds no GPU on Jetson #32541

[azure-pipelines]

Azure Pipelines:
There may be pipelines that require an authorized user to comment /azp run to run.

[Tianlei Wu (tianleiwu)]

Copilot resolve the merge conflicts in this pull request

[Copilot]

Copilot resolve the merge conflicts in this pull request

Resolved the conflict by merging current main and preserving its shared NVIDIA vendor-ID constant in the Jetson discovery path. Merge commit: efe18509. The Release build and all 11 focused DRM/PCI discovery tests pass.

[Copilot]

🔵 Needs a closer look

The accepted nvidia driver path still lacks a dedicated positive test.

Pull request overview

Adds Linux DRM discovery for Jetson platform-bus NVIDIA GPUs, surfacing non-PCI nvgpu/nvidia devices as integrated GPUs.

Changes:

• Detects supported NVIDIA platform DRM devices without PCI metadata.
• Preserves filtering of other non-NVIDIA platform devices.
• Adds Jetson-style sysfs tests and CMake integration.

File summaries

File	Summary
onnxruntime/test/platform/linux/drm_device_discovery_test.cc	Adds Jetson-style discovery and filtering tests.
onnxruntime/core/platform/linux/drm_device_discovery.h	Declares DRM discovery interfaces.
onnxruntime/core/platform/linux/device_discovery.cc	Implements NVIDIA platform GPU detection and metadata handling.
cmake/onnxruntime_common.cmake	Registers the DRM discovery header.

Review details

Suppressed comments (2)

onnxruntime/core/platform/linux/device_discovery.cc:159

• The new contract accepts both nvgpu and nvidia, but the tests only exercise DRIVER=nvgpu; a regression or typo in the second accepted driver would therefore pass. Add a focused case with DRIVER=nvidia and assert that it is recognized as a platform GPU.

                               (HasUeventEntry(uevent, "DRIVER=nvgpu") ||
                                HasUeventEntry(uevent, "DRIVER=nvidia"));

onnxruntime/test/platform/linux/drm_device_discovery_test.cc:45

• The new detector explicitly accepts both DRIVER=nvgpu and DRIVER=nvidia, but this fixture only exercises the former (the other two values are rejection cases). Add a positive nvidia platform-card case and assert that both accepted cards are returned, so the second supported Jetson-style driver cannot regress untested.

  CreatePlatformDrmCard(temp_dir_ / "card0", "nvgpu");
  CreatePlatformDrmCard(temp_dir_ / "card1", "amdgpu");
  CreatePlatformDrmCard(temp_dir_ / "card2", "nvgpu-extra");

• Files reviewed: 4/4 changed files
• Comments generated: 0
• Review effort level: Lite (auto)

Note

Copilot is running an experiment and ran this review at Lite.

---

💡 Configure MCP servers for context-aware, tailored reviews. Learn more in the docs.