fix: Add Linux NPU discovery through sysfs accel devices (#28703)

the0cp · web-flow · commit 5c7ba1315d75 · 2026-06-11T09:04:38.000-07:00
## Description This PR adds Linux NPU discovery through sysfs accel devices Currently, `DeviceDiscovery::DiscoverDevicesForPlatform()` on Linux discovers CPU and GPU devices, but NPU discovery is still missing. As a result, plugin execution providers that filter devices by `OrtHardwareDeviceType_NPU` do not receive any NPU hardware devices on Linux, even when the NPU is present and exposed by the kernel. This change scans `/sys/class/accel` for `accelN` devices and creates `OrtHardwareDevice` entries with: - `type = OrtHardwareDeviceType_NPU` - PCI `vendor_id` - PCI `device_id` - `accel_idx` metadata - `pci_bus_id` metadata when available This enables Linux systems with NPUs exposed through the accel subsystem, such as AMD Ryzen AI / XDNA devices, to be reported through ORT device discovery and made available to plugin EP factories. ## Changes - Add Linux sysfs discovery for NPU devices under `/sys/class/accel`. - Read NPU PCI vendor and device IDs from the underlying sysfs device path. - Add NPU metadata including `accel_idx` and `pci_bus_id`. - Include discovered NPU devices in `DeviceDiscovery::DiscoverDevicesForPlatform()`. - Add a `kSysfsAccelPath` constant for the Linux accel sysfs path. ## Motivation Linux plugin EPs that target NPUs rely on ORT passing `OrtHardwareDeviceType_NPU` devices into `GetSupportedDevices()`. Without Linux NPU discovery, those EPs cannot claim NPU devices and provider selection policies such as `PREFER_NPU` silently fall back to CPU. Fixes #28660.
diff --git a/onnxruntime/core/platform/linux/device_discovery.cc b/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/npu_device_discovery.h"
 #include "core/platform/linux/pci_device_discovery.h"
 
 #include <filesystem>
@@ -123,7 +124,7 @@ Status GetPciBusId(const std::filesystem::path& sysfs_path, std::optional<std::s
 
   std::error_code error_code;
   auto pci_bus_id_path = std::filesystem::canonical(sysfs_path / "device", error_code);  // resolves symlink to PCI bus id, e.g. 0000:65:00.0
-  ORT_RETURN_IF_ERROR(ErrorCodeToStatus(error_code, sysfs_path / "device", "Getting PCI bus id from DRM device by resolving symlink"));
+  ORT_RETURN_IF_ERROR(ErrorCodeToStatus(error_code, sysfs_path / "device", "Getting PCI bus id from sysfs device by resolving symlink"));
 
   auto pci_bus_id_filename = pci_bus_id_path.filename();
   if (std::regex_match(pci_bus_id_filename.string(), pci_bus_id_regex)) {
@@ -269,6 +270,7 @@ Status GetGpuDeviceFromPci(const GpuPciPathInfo& path_info, OrtHardwareDevice& g
 namespace {
 
 constexpr const char* kSysfsPciDevicesPath = "/sys/bus/pci/devices";
+constexpr const char* kSysfsAccelPath = "/sys/class/accel";
 
 Status GetGpuDevices(std::vector<OrtHardwareDevice>& gpu_devices_out) {
   std::vector<GpuSysfsPathInfo> gpu_sysfs_path_infos{};
@@ -315,6 +317,119 @@ Status GetGpuDevices(std::vector<OrtHardwareDevice>& gpu_devices_out) {
 
 }  // namespace
 
+namespace npu_device_discovery {
+
+Status DetectNpuSysfsPaths(const fs::path& sysfs_accel_path,
+                           std::vector<NpuSysfsPathInfo>& npu_sysfs_paths_out) {
+  std::error_code error_code{};
+
+  const bool sysfs_accel_path_exists = fs::exists(sysfs_accel_path, error_code);
+  ORT_RETURN_IF_ERROR(ErrorCodeToStatus(error_code, sysfs_accel_path, "Checking existence of accel sysfs path"));
+
+  if (!sysfs_accel_path_exists) {
+    npu_sysfs_paths_out = {};
+    return Status::OK();
+  }
+
+  const auto detect_accel_path = [](const fs::path& sysfs_path, size_t& accel_idx) -> bool {
+    const auto filename = sysfs_path.filename();
+    const auto filename_str = std::string_view{filename.native()};
+
+    // Look for a filename matching "accelN". N is a number.
+    constexpr std::string_view prefix = "accel";
+    if (filename_str.find(prefix) != 0) {
+      return false;
+    }
+
+    size_t parsed_accel_idx{};
+    if (!TryParseStringWithClassicLocale<size_t>(filename_str.substr(prefix.size()), parsed_accel_idx)) {
+      return false;
+    }
+
+    accel_idx = parsed_accel_idx;
+    return true;
+  };
+
+  std::vector<NpuSysfsPathInfo> npu_sysfs_paths{};
+
+  auto dir_iterator = fs::directory_iterator{sysfs_accel_path, error_code};
+  ORT_RETURN_IF_ERROR(ErrorCodeToStatus(error_code, sysfs_accel_path, "Iterating over accel sysfs devices"));
+
+  for (const auto& dir_item : dir_iterator) {
+    const auto& dir_item_path = dir_item.path();
+
+    if (size_t accel_idx{}; detect_accel_path(dir_item_path, accel_idx)) {
+      NpuSysfsPathInfo path_info{};
+      path_info.accel_idx = accel_idx;
+      path_info.path = dir_item_path;
+      npu_sysfs_paths.emplace_back(std::move(path_info));
+    }
+  }
+
+  npu_sysfs_paths_out = std::move(npu_sysfs_paths);
+  return Status::OK();
+}
+
+Status GetNpuDeviceFromSysfs(const NpuSysfsPathInfo& path_info,
+                             OrtHardwareDevice& npu_device_out) {
+  OrtHardwareDevice npu_device{};
+
+  const auto& sysfs_path = path_info.path;
+
+  uint16_t vendor_id{};
+  const auto vendor_id_path = sysfs_path / "device" / "vendor";
+  ORT_RETURN_IF_ERROR(ReadValueFromFile(vendor_id_path, vendor_id));
+  npu_device.vendor_id = vendor_id;
+
+  uint16_t device_id{};
+  const auto device_id_path = sysfs_path / "device" / "device";
+  ORT_RETURN_IF_ERROR(ReadValueFromFile(device_id_path, device_id));
+  npu_device.device_id = device_id;
+
+  npu_device.metadata.Add("accel_idx", MakeString(path_info.accel_idx));
+
+  std::optional<std::string> pci_bus_id;
+  ORT_RETURN_IF_ERROR(GetPciBusId(sysfs_path, pci_bus_id));
+  if (pci_bus_id) {
+    npu_device.metadata.Add("pci_bus_id", std::move(*pci_bus_id));
+  }
+
+  npu_device.type = OrtHardwareDeviceType_NPU;
+
+  npu_device_out = std::move(npu_device);
+
+  return Status::OK();
+}
+
+}  // namespace npu_device_discovery
+
+namespace {
+
+Status GetNpuDevices(std::vector<OrtHardwareDevice>& npu_devices_out) {
+  std::vector<npu_device_discovery::NpuSysfsPathInfo> npu_sysfs_path_infos{};
+  ORT_RETURN_IF_ERROR(npu_device_discovery::DetectNpuSysfsPaths(kSysfsAccelPath, npu_sysfs_path_infos));
+
+  std::vector<OrtHardwareDevice> npu_devices{};
+  npu_devices.reserve(npu_sysfs_path_infos.size());
+
+  for (const auto& npu_sysfs_path_info : npu_sysfs_path_infos) {
+    OrtHardwareDevice npu_device{};
+    if (auto status = npu_device_discovery::GetNpuDeviceFromSysfs(npu_sysfs_path_info, npu_device); !status.IsOK()) {
+      LOGS_DEFAULT(WARNING) << MakeString("Failed to detect devices under ",
+                                          npu_sysfs_path_info.path,
+                                          ": ",
+                                          status.ErrorMessage());
+      continue;
+    }
+    npu_devices.emplace_back(std::move(npu_device));
+  }
+
+  npu_devices_out = std::move(npu_devices);
+
+  return Status::OK();
+}
+}  // namespace
+
 std::unordered_set<OrtHardwareDevice> DeviceDiscovery::DiscoverDevicesForPlatform() {
   std::unordered_set<OrtHardwareDevice> devices;
 
@@ -334,7 +449,16 @@ std::unordered_set<OrtHardwareDevice> DeviceDiscovery::DiscoverDevicesForPlatfor
   }
 
   // get NPU devices
-  // TODO figure out how to discover these
+  {
+    std::vector<OrtHardwareDevice> npu_devices{};
+    Status npu_device_discovery_status = GetNpuDevices(npu_devices);
+    if (npu_device_discovery_status.IsOK()) {
+      devices.insert(std::make_move_iterator(npu_devices.begin()),
+                     std::make_move_iterator(npu_devices.end()));
+    } else {
+      LOGS_DEFAULT(WARNING) << "NPU device discovery failed: " << npu_device_discovery_status.ErrorMessage();
+    }
+  }
 
   return devices;
 }
diff --git a/onnxruntime/core/platform/linux/npu_device_discovery.h b/onnxruntime/core/platform/linux/npu_device_discovery.h
@@ -0,0 +1,32 @@
+// Copyright (c) Microsoft Corporation. All rights reserved.
+// Licensed under the MIT License.
+
+// This header exposes Linux NPU device discovery internals for testing.
+
+#pragma once
+
+#include <cstddef>
+#include <filesystem>
+#include <vector>
+
+#include "core/common/status.h"
+#include "core/session/abi_devices.h"
+
+namespace onnxruntime {
+namespace npu_device_discovery {
+
+struct NpuSysfsPathInfo {
+  size_t accel_idx;
+  std::filesystem::path path;
+};
+
+// Scans the given sysfs accel directory for NPU accel devices.
+Status DetectNpuSysfsPaths(const std::filesystem::path& sysfs_accel_path,
+                           std::vector<NpuSysfsPathInfo>& npu_sysfs_paths_out);
+
+// Reads vendor/device IDs and populates an OrtHardwareDevice from an accel sysfs path.
+Status GetNpuDeviceFromSysfs(const NpuSysfsPathInfo& path_info,
+                             OrtHardwareDevice& npu_device_out);
+
+}  // namespace npu_device_discovery
+}  // namespace onnxruntime
diff --git a/onnxruntime/test/platform/linux/npu_device_discovery_test.cc b/onnxruntime/test/platform/linux/npu_device_discovery_test.cc
@@ -0,0 +1,103 @@
+// Copyright (c) Microsoft Corporation. All rights reserved.
+// Licensed under the MIT License.
+
+#include "core/platform/linux/npu_device_discovery.h"
+
+#include <algorithm>
+#include <filesystem>
+#include <fstream>
+#include <string>
+#include <vector>
+
+#include "gtest/gtest.h"
+#include "test/util/include/asserts.h"
+
+namespace fs = std::filesystem;
+
+namespace onnxruntime::test {
+namespace {
+
+void WriteFile(const fs::path& path, const std::string& value) {
+  std::ofstream f(path);
+  f << value;
+}
+
+class NpuDeviceDiscoveryTest : public ::testing::Test {
+ protected:
+  void SetUp() override {
+    temp_dir_ = fs::temp_directory_path() / "ort_npu_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(NpuDeviceDiscoveryTest, ReturnsEmptyForNonexistentPath) {
+  std::vector<npu_device_discovery::NpuSysfsPathInfo> npu_paths;
+  ASSERT_STATUS_OK(npu_device_discovery::DetectNpuSysfsPaths(temp_dir_ / "nonexistent", npu_paths));
+  EXPECT_TRUE(npu_paths.empty());
+}
+
+TEST_F(NpuDeviceDiscoveryTest, DetectsAccelDevices) {
+  fs::create_directories(temp_dir_ / "accel0");
+  fs::create_directories(temp_dir_ / "accel12");
+  fs::create_directories(temp_dir_ / "renderD128");
+  fs::create_directories(temp_dir_ / "accelabc");
+
+  std::vector<npu_device_discovery::NpuSysfsPathInfo> npu_paths;
+  ASSERT_STATUS_OK(npu_device_discovery::DetectNpuSysfsPaths(temp_dir_, npu_paths));
+
+  ASSERT_EQ(npu_paths.size(), 2u);
+
+  std::vector<size_t> accel_indices;
+  accel_indices.reserve(npu_paths.size());
+  for (const auto& npu_path : npu_paths) {
+    accel_indices.push_back(npu_path.accel_idx);
+  }
+
+  std::sort(accel_indices.begin(), accel_indices.end());
+
+  EXPECT_EQ(accel_indices[0], 0u);
+  EXPECT_EQ(accel_indices[1], 12u);
+}
+
+TEST_F(NpuDeviceDiscoveryTest, GetNpuDeviceFromSysfsReadsVendorDeviceAndMetadata) {
+  const auto pci_device_dir = temp_dir_ / "pci_devices" / "0000:65:00.0";
+  const auto accel_dir = temp_dir_ / "class_accel" / "accel0";
+
+  fs::create_directories(pci_device_dir);
+  fs::create_directories(accel_dir);
+
+  WriteFile(pci_device_dir / "vendor", "0x1022");
+  WriteFile(pci_device_dir / "device", "0x1502");
+
+  std::error_code error_code{};
+  fs::create_directory_symlink(pci_device_dir, accel_dir / "device", error_code);
+  ASSERT_FALSE(error_code) << error_code.message();
+
+  npu_device_discovery::NpuSysfsPathInfo path_info{};
+  path_info.accel_idx = 0;
+  path_info.path = accel_dir;
+
+  OrtHardwareDevice npu_device{};
+  ASSERT_STATUS_OK(npu_device_discovery::GetNpuDeviceFromSysfs(path_info, npu_device));
+
+  EXPECT_EQ(npu_device.type, OrtHardwareDeviceType_NPU);
+  EXPECT_EQ(npu_device.vendor_id, 0x1022u);
+  EXPECT_EQ(npu_device.device_id, 0x1502u);
+
+  const auto& entries = npu_device.metadata.Entries();
+  EXPECT_NE(entries.find("accel_idx"), entries.end());
+  EXPECT_EQ(entries.at("accel_idx"), "0");
+  EXPECT_NE(entries.find("pci_bus_id"), entries.end());
+  EXPECT_EQ(entries.at("pci_bus_id"), "0000:65:00.0");
+}
+
+}  // namespace onnxruntime::test
