Add PCI bus fallback for Linux GPU device discovery in containerized environments

cmake/onnxruntime_common.cmake

@@ -84,7 +84,8 @@ if (WIN32)
          "${ONNXRUNTIME_ROOT}/core/platform/windows/device_discovery.cc")
 elseif (LINUX)
     list(APPEND onnxruntime_common_src_patterns
-         "${ONNXRUNTIME_ROOT}/core/platform/linux/device_discovery.cc")
+         "${ONNXRUNTIME_ROOT}/core/platform/linux/device_discovery.cc"
+         "${ONNXRUNTIME_ROOT}/core/platform/linux/pci_device_discovery.h")
 elseif (APPLE)
     list(APPEND onnxruntime_common_src_patterns
          "${ONNXRUNTIME_ROOT}/core/platform/apple/device_discovery.cc")

cmake/onnxruntime_unittests.cmake

@@ -461,6 +461,11 @@ if(WIN32)
     "${TEST_SRC_DIR}/platform/windows/logging/*.cc" )
 endif()
 
+if(LINUX)
+  list(APPEND onnxruntime_test_framework_src_patterns
+    "${TEST_SRC_DIR}/platform/linux/*.cc" )
+endif()
+
 if(NOT onnxruntime_MINIMAL_BUILD AND NOT onnxruntime_REDUCED_OPS_BUILD)
 
   if(onnxruntime_USE_CUDA)

onnxruntime/core/platform/linux/device_discovery.cc

@@ -2,6 +2,7 @@
 // Licensed under the MIT License.
 
 #include "core/platform/device_discovery.h"
+#include "core/platform/linux/pci_device_discovery.h"
 
 #include <filesystem>
 #include <fstream>
@@ -175,6 +176,99 @@
   return Status::OK();
 }
 
+}  // namespace
+
+// PCI bus-based GPU detection as a fallback for environments where DRM sysfs entries
+// are not available (e.g., AKS/Kubernetes containers where the nvidia-drm kernel module
+// is not loaded but GPU PCI devices are still exposed via sysfs).
+
+namespace pci_device_discovery {
+
+Status DetectGpuPciPaths(const fs::path& sysfs_pci_devices_path,
+                         std::vector<GpuPciPathInfo>& gpu_pci_paths_out) {
+  std::error_code error_code{};
+  const bool path_exists = fs::exists(sysfs_pci_devices_path, error_code);
+  ORT_RETURN_IF_ERROR(ErrorCodeToStatus(error_code));
+
+  if (!path_exists) {
+    gpu_pci_paths_out = {};
+    return Status::OK();
+  }
+
+  std::vector<GpuPciPathInfo> gpu_pci_paths{};
+
+  auto dir_iterator = fs::directory_iterator{sysfs_pci_devices_path, error_code};
+  ORT_RETURN_IF_ERROR(ErrorCodeToStatus(error_code));
+
+  for (const auto& dir_item : dir_iterator) {
+    const auto& device_path = dir_item.path();
+
+    // Read PCI class code to identify GPU devices.
+    // The class file contains a 24-bit value: 0xCCSSpp (class/subclass/prog-if).
+    uint32_t pci_class{};
+    if (auto status = ReadValueFromFile(device_path / "class", pci_class); !status.IsOK()) {
+      continue;
+    }
+
+    // Check for GPU/display controller PCI class codes:
+    //   Base class 0x03 = Display controller
+    //   Sub-class 0x00 = VGA compatible controller
+    //   Sub-class 0x02 = 3D controller (common for NVIDIA data center/compute GPUs)
+    // Reference: PCI Code and ID Assignment Specification
+    //   https://pcisig.com/pci-code-and-id-assignment-specification-agreement
+    //   See section on base class 03h.
+    const uint8_t base_class = static_cast<uint8_t>((pci_class >> 16) & 0xFF);
+    const uint8_t sub_class = static_cast<uint8_t>((pci_class >> 8) & 0xFF);
+    if (base_class != 0x03 || (sub_class != 0x00 && sub_class != 0x02)) {
+      continue;
+    }
+
+    GpuPciPathInfo path_info{};
+    path_info.path = device_path;
+    path_info.pci_bus_id = device_path.filename().string();
+    gpu_pci_paths.emplace_back(std::move(path_info));
+  }
+
+  gpu_pci_paths_out = std::move(gpu_pci_paths);
+  return Status::OK();
+}
+
+Status GetGpuDeviceFromPci(const GpuPciPathInfo& path_info, OrtHardwareDevice& gpu_device_out) {
+  OrtHardwareDevice gpu_device{};
+  const auto& pci_path = path_info.path;
+
+  // vendor id - directly under PCI device path
+  uint16_t vendor_id{};
+  ORT_RETURN_IF_ERROR(ReadValueFromFile(pci_path / "vendor", vendor_id));
+  gpu_device.vendor_id = vendor_id;
+
+  // device id - directly under PCI device path
+  uint16_t device_id{};
+  ORT_RETURN_IF_ERROR(ReadValueFromFile(pci_path / "device", device_id));
+  gpu_device.device_id = device_id;
+
+  // metadata
+  if (const auto is_gpu_discrete = IsGpuDiscrete(vendor_id, device_id);
+      is_gpu_discrete.has_value()) {
+    gpu_device.metadata.Add("Discrete", (*is_gpu_discrete ? "1" : "0"));
+  }
+
+  if (!path_info.pci_bus_id.empty()) {
+    gpu_device.metadata.Add("pci_bus_id", path_info.pci_bus_id);
+  }
+
+  gpu_device.type = OrtHardwareDeviceType_GPU;
+
+  gpu_device_out = std::move(gpu_device);
+  return Status::OK();
+}
+
+}  // namespace pci_device_discovery
+
+namespace {
+
+constexpr const char* kSysfsPciDevicesPath = "/sys/bus/pci/devices";
+
 Status GetGpuDevices(std::vector<OrtHardwareDevice>& gpu_devices_out) {
   std::vector<GpuSysfsPathInfo> gpu_sysfs_path_infos{};
   ORT_RETURN_IF_ERROR(DetectGpuSysfsPaths(gpu_sysfs_path_infos));
@@ -188,6 +282,26 @@
     gpu_devices.emplace_back(std::move(gpu_device));
   }
 
+  // If DRM-based detection found no GPUs, fall back to PCI bus scanning.
+  // This handles containerized environments (e.g., AKS/Kubernetes) where the DRM
+  // subsystem (nvidia-drm) may not be available but GPU PCI devices are still
+  // exposed via /sys/bus/pci/devices/.
+  if (gpu_devices.empty()) {
+    LOGS_DEFAULT(VERBOSE) << "No GPUs found via /sys/class/drm. "
+                          << "Falling back to PCI bus scanning via " << kSysfsPciDevicesPath << ".";
+
+    std::vector<pci_device_discovery::GpuPciPathInfo> gpu_pci_path_infos{};
+    ORT_RETURN_IF_ERROR(pci_device_discovery::DetectGpuPciPaths(kSysfsPciDevicesPath, gpu_pci_path_infos));
+
+    gpu_devices.reserve(gpu_pci_path_infos.size());
+
+    for (const auto& gpu_pci_path_info : gpu_pci_path_infos) {
+      OrtHardwareDevice gpu_device{};
+      ORT_RETURN_IF_ERROR(pci_device_discovery::GetGpuDeviceFromPci(gpu_pci_path_info, gpu_device));
+      gpu_devices.emplace_back(std::move(gpu_device));
+    }
+  }
+
   gpu_devices_out = std::move(gpu_devices);
   return Status::OK();
 }

onnxruntime/core/platform/linux/pci_device_discovery.h

@@ -0,0 +1,33 @@
+// Copyright (c) Microsoft Corporation. All rights reserved.
+// Licensed under the MIT License.
+
+// This header exposes Linux PCI device discovery internals for testing.
+
+#pragma once
+
+#include <filesystem>
+#include <string>
+#include <vector>
+
+#include "core/common/status.h"
+#include "core/session/abi_devices.h"
+
+namespace onnxruntime {
+namespace pci_device_discovery {
+
+struct GpuPciPathInfo {
+  std::filesystem::path path;
+  std::string pci_bus_id;
+};
+
+// Scans the given sysfs PCI devices directory for GPU devices.
+// Filters by PCI class codes: 0x0300 (VGA) and 0x0302 (3D controller).
+Status DetectGpuPciPaths(const std::filesystem::path& sysfs_pci_devices_path,
+                         std::vector<GpuPciPathInfo>& gpu_pci_paths_out);
+
+// Reads vendor/device IDs and populates an OrtHardwareDevice from a PCI device sysfs path.
+Status GetGpuDeviceFromPci(const GpuPciPathInfo& path_info,
+                           OrtHardwareDevice& gpu_device_out);
+
+}  // namespace pci_device_discovery
+}  // namespace onnxruntime

onnxruntime/test/platform/device_discovery_test.cc

@@ -30,5 +30,15 @@ TEST(DeviceDiscoveryTest, HasCpuDevice) {
 #endif  // defined(CPUINFO_SUPPORTED)
 }
 
+TEST(DeviceDiscoveryTest, GpuDevicesHaveValidProperties) {
+  const auto gpu_devices = GetDevicesByType(OrtHardwareDeviceType_GPU);
+
+  // GPU detection should not crash. If GPUs are present, validate their properties.
+  for (const auto& gpu_device : gpu_devices) {
+    EXPECT_NE(gpu_device.vendor_id, 0u);
+    // Note: device_id may be 0 on some platforms (e.g., Apple Silicon) where it is not populated.
+  }
+}
+
 }  // namespace onnxruntime::test
 #endif  // !defined(ORT_MINIMAL_BUILD) && !defined(_GAMING_XBOX)

onnxruntime/test/platform/linux/pci_device_discovery_test.cc

@@ -0,0 +1,165 @@
+// Copyright (c) Microsoft Corporation. All rights reserved.
+// Licensed under the MIT License.
+
+#include "core/platform/linux/pci_device_discovery.h"
+
+#include <filesystem>
+#include <fstream>
+
+#include "test/util/include/asserts.h"
+#include "gtest/gtest.h"
+
+namespace fs = std::filesystem;
+
+namespace onnxruntime::test {
+
+namespace {
+
+// Helper to create a fake PCI device directory with the given class, vendor, and device files.
+void CreateFakePciDevice(const fs::path& device_dir, const std::string& pci_class,
+                         const std::string& vendor, const std::string& device) {
+  fs::create_directories(device_dir);
+  {
+    std::ofstream f(device_dir / "class");
+    f << pci_class;
+  }
+  {
+    std::ofstream f(device_dir / "vendor");
+    f << vendor;
+  }
+  {
+    std::ofstream f(device_dir / "device");
+    f << device;
+  }
+}
+
+class PciDeviceDiscoveryTest : public ::testing::Test {
+ protected:
+  void SetUp() override {
+    temp_dir_ = fs::temp_directory_path() / "ort_pci_discovery_test";
+    fs::remove_all(temp_dir_);
+    fs::create_directories(temp_dir_);
+  }
+
+  void TearDown() override {
+    fs::remove_all(temp_dir_);
+  }
+
+  fs::path temp_dir_;
+};
+
+}  // namespace
+
+TEST_F(PciDeviceDiscoveryTest, DetectsNvidiaVgaController) {
+  // PCI class 0x030000 = VGA compatible controller
+  CreateFakePciDevice(temp_dir_ / "0000:01:00.0", "0x030000", "0x10de", "0x2204");
+
+  std::vector<pci_device_discovery::GpuPciPathInfo> gpu_paths;
+  ASSERT_STATUS_OK(pci_device_discovery::DetectGpuPciPaths(temp_dir_, gpu_paths));
+  ASSERT_EQ(gpu_paths.size(), 1u);
+  EXPECT_EQ(gpu_paths[0].pci_bus_id, "0000:01:00.0");
+}
+
+TEST_F(PciDeviceDiscoveryTest, DetectsNvidia3DController) {
+  // PCI class 0x030200 = 3D controller (common for NVIDIA datacenter GPUs like A100/H100)
+  CreateFakePciDevice(temp_dir_ / "0000:65:00.0", "0x030200", "0x10de", "0x20b5");
+
+  std::vector<pci_device_discovery::GpuPciPathInfo> gpu_paths;
+  ASSERT_STATUS_OK(pci_device_discovery::DetectGpuPciPaths(temp_dir_, gpu_paths));
+  ASSERT_EQ(gpu_paths.size(), 1u);
+  EXPECT_EQ(gpu_paths[0].pci_bus_id, "0000:65:00.0");
+}
+
+TEST_F(PciDeviceDiscoveryTest, FiltersOutNonGpuDevices) {
+  // PCI class 0x020000 = Network controller (should be skipped)
+  CreateFakePciDevice(temp_dir_ / "0000:02:00.0", "0x020000", "0x8086", "0x1533");
+  // PCI class 0x010600 = SATA controller (should be skipped)
+  CreateFakePciDevice(temp_dir_ / "0000:00:1f.2", "0x010600", "0x8086", "0xa102");
+  // PCI class 0x030000 = VGA controller (should be detected)
+  CreateFakePciDevice(temp_dir_ / "0000:01:00.0", "0x030000", "0x10de", "0x2204");
+
+  std::vector<pci_device_discovery::GpuPciPathInfo> gpu_paths;
+  ASSERT_STATUS_OK(pci_device_discovery::DetectGpuPciPaths(temp_dir_, gpu_paths));
+  ASSERT_EQ(gpu_paths.size(), 1u);
+  EXPECT_EQ(gpu_paths[0].pci_bus_id, "0000:01:00.0");
+}
+
+TEST_F(PciDeviceDiscoveryTest, ReturnsEmptyForNonexistentPath) {
+  std::vector<pci_device_discovery::GpuPciPathInfo> gpu_paths;
+  ASSERT_STATUS_OK(pci_device_discovery::DetectGpuPciPaths(temp_dir_ / "nonexistent", gpu_paths));
+  EXPECT_TRUE(gpu_paths.empty());
+}
+
+TEST_F(PciDeviceDiscoveryTest, ReturnsEmptyForEmptyDirectory) {
+  std::vector<pci_device_discovery::GpuPciPathInfo> gpu_paths;
+  ASSERT_STATUS_OK(pci_device_discovery::DetectGpuPciPaths(temp_dir_, gpu_paths));
+  EXPECT_TRUE(gpu_paths.empty());
+}
+
+TEST_F(PciDeviceDiscoveryTest, DetectsMultipleGpus) {
+  // Two NVIDIA GPUs
+  CreateFakePciDevice(temp_dir_ / "0000:01:00.0", "0x030200", "0x10de", "0x20b5");
+  CreateFakePciDevice(temp_dir_ / "0000:41:00.0", "0x030200", "0x10de", "0x20b5");
+
+  std::vector<pci_device_discovery::GpuPciPathInfo> gpu_paths;
+  ASSERT_STATUS_OK(pci_device_discovery::DetectGpuPciPaths(temp_dir_, gpu_paths));
+  EXPECT_EQ(gpu_paths.size(), 2u);
+}
+
+TEST_F(PciDeviceDiscoveryTest, SkipsDevicesWithMissingClassFile) {
+  // Device directory without a class file
+  auto device_dir = temp_dir_ / "0000:03:00.0";
+  fs::create_directories(device_dir);
+  {
+    std::ofstream f(device_dir / "vendor");
+    f << "0x10de";
+  }
+
+  std::vector<pci_device_discovery::GpuPciPathInfo> gpu_paths;
+  ASSERT_STATUS_OK(pci_device_discovery::DetectGpuPciPaths(temp_dir_, gpu_paths));
+  EXPECT_TRUE(gpu_paths.empty());
+}
+
+TEST_F(PciDeviceDiscoveryTest, GetGpuDeviceFromPciReadsVendorAndDevice) {
+  // Create a fake NVIDIA GPU PCI device
+  CreateFakePciDevice(temp_dir_ / "0000:65:00.0", "0x030200", "0x10de", "0x20b5");
+
+  pci_device_discovery::GpuPciPathInfo path_info;
+  path_info.path = temp_dir_ / "0000:65:00.0";
+  path_info.pci_bus_id = "0000:65:00.0";
+
+  OrtHardwareDevice gpu_device{};
+  ASSERT_STATUS_OK(pci_device_discovery::GetGpuDeviceFromPci(path_info, gpu_device));
+
+  EXPECT_EQ(gpu_device.type, OrtHardwareDeviceType_GPU);
+  EXPECT_EQ(gpu_device.vendor_id, 0x10deu);
+  EXPECT_EQ(gpu_device.device_id, 0x20b5u);
+
+  const auto& entries = gpu_device.metadata.Entries();
+  EXPECT_NE(entries.find("pci_bus_id"), entries.end());
+  EXPECT_EQ(entries.at("pci_bus_id"), "0000:65:00.0");
+  EXPECT_NE(entries.find("Discrete"), entries.end());
+  EXPECT_EQ(entries.at("Discrete"), "1");
+}
+
+TEST_F(PciDeviceDiscoveryTest, GetGpuDeviceFromPciNonNvidiaVendor) {
+  // Create a fake AMD GPU PCI device
+  CreateFakePciDevice(temp_dir_ / "0000:03:00.0", "0x030000", "0x1002", "0x731f");
+
+  pci_device_discovery::GpuPciPathInfo path_info;
+  path_info.path = temp_dir_ / "0000:03:00.0";
+  path_info.pci_bus_id = "0000:03:00.0";
+
+  OrtHardwareDevice gpu_device{};
+  ASSERT_STATUS_OK(pci_device_discovery::GetGpuDeviceFromPci(path_info, gpu_device));
+
+  EXPECT_EQ(gpu_device.type, OrtHardwareDeviceType_GPU);
+  EXPECT_EQ(gpu_device.vendor_id, 0x1002u);
+  EXPECT_EQ(gpu_device.device_id, 0x731fu);
+
+  const auto& entries = gpu_device.metadata.Entries();
+  // Non-NVIDIA vendor should not have the Discrete metadata entry
+  EXPECT_EQ(entries.find("Discrete"), entries.end());
+}
+
+}  // namespace onnxruntime::test