CUDA pinned memory fixes

Author: gedoensmax

onnxruntime/core/providers/nv_tensorrt_rtx/nv_allocator.h

@@ -6,6 +6,7 @@
 
 #include "core/common/inlined_containers.h"
 #include "core/framework/allocator.h"
+#include "core/providers/shared_library/provider_api.h"
 #include <mutex>
 
 namespace onnxruntime {
@@ -56,9 +57,13 @@ class CUDAPinnedAllocator : public IAllocator {
   CUDAPinnedAllocator(const char* name, OrtDevice::DeviceId device_id)
       : IAllocator(
             OrtMemoryInfo(name, OrtAllocatorType::OrtDeviceAllocator,
-                          OrtDevice(OrtDevice::GPU, OrtDevice::MemType::HOST_ACCESSIBLE, OrtDevice::VendorIds::NVIDIA,
-                                    device_id),
-                          OrtMemTypeCPUOutput)) {}
+                          OrtDevice(OrtDevice::CPU, OrtDevice::MemType::HOST_ACCESSIBLE, OrtDevice::VendorIds::NVIDIA,
+                                    0 /*CPU device always with id 0*/),
+                          OrtMemTypeCPUOutput)) {
+    if (device_id != 0) {
+      LOGS_DEFAULT(WARNING) << "[NvTensorRTRTX EP] Device ID was non 0 for a CudaPinned allocator. The ID will be ignored.";
+    }
+  }
 
   void* Alloc(size_t size) override;
   void Free(void* p) override;

onnxruntime/core/providers/nv_tensorrt_rtx/nv_execution_provider.cc

@@ -1304,7 +1304,7 @@ std::vector<AllocatorPtr> NvExecutionProvider::CreatePreferredAllocators() {
 
   AllocatorCreationInfo pinned_allocator_info(
       [](OrtDevice::DeviceId device_id) {
-        return std::make_unique<CUDAPinnedAllocator>(device_id, CUDA_PINNED);
+        return std::make_unique<CUDAPinnedAllocator>(CUDA_PINNED, device_id);
       },
       narrow<OrtDevice::DeviceId>(device_id_));
 
@@ -3258,8 +3258,8 @@ OrtDevice NvExecutionProvider::GetOrtDeviceByMemType(OrtMemType mem_type) const
   if (mem_type == OrtMemTypeCPUInput)
     return OrtDevice();
   if (mem_type == OrtMemTypeCPUOutput)
-    return OrtDevice(OrtDevice::GPU, OrtDevice::MemType::HOST_ACCESSIBLE, OrtDevice::VendorIds::NVIDIA,
-                     default_device_.Id());
+    return OrtDevice(OrtDevice::CPU, OrtDevice::MemType::HOST_ACCESSIBLE, OrtDevice::VendorIds::NVIDIA,
+                     0 /*CPU device id always be 0*/);
   return default_device_;
 }