Skip to content

Only add support for CUDA devices in GetSupportedDevices #4

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
chilo-ms merged 12 commits into from
Apr 15, 2026
Merged

Only add support for CUDA devices in GetSupportedDevices #4

Show file tree
Hide file tree
Changes from 3 commits
Commits
Show all changes
12 commits
Select commit Hold shift + click to select a range
21653db
update
chilo-ms Apr 1, 2026
1de07bd
check cuda api return value
chilo-ms Apr 1, 2026
ab98037
address reviewer's comments
chilo-ms Apr 1, 2026
0fa5f54
address reviewer's comments
chilo-ms Apr 2, 2026
d8a8d84
address reviewer's comments
chilo-ms Apr 2, 2026
bb325a2
update
chilo-ms Apr 2, 2026
86b17b0
refactor
chilo-ms Apr 10, 2026
41ea68c
add TRT builder place holder for test
chilo-ms Apr 10, 2026
3699c7f
address reviewer's comments
chilo-ms Apr 14, 2026
0e36718
address reviewer's comments
chilo-ms Apr 14, 2026
100f440
address reviewer's comment
chilo-ms Apr 15, 2026
e8e0de6
add try/catch as C API boundary guard
chilo-ms Apr 15, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
7 changes: 2 additions & 5 deletions src/tensorrt_execution_provider_data_transfer.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -23,12 +23,9 @@ bool ORT_API_CALL TRTEpDataTransfer::CanCopyImpl(const OrtDataTransferImpl* this
auto src_vendor_id = impl.ep_api.MemoryDevice_GetVendorId(src_memory_device);
auto dst_vendor_id = impl.ep_api.MemoryDevice_GetVendorId(dst_memory_device);

// 0x10DE is the PCI vendor ID for NVIDIA
constexpr uint32_t nvidia_vendor_id = 0x10DE;

// Reject if GPU device is not NVIDIA
if ((src_type == OrtMemoryInfoDeviceType_GPU && src_vendor_id != nvidia_vendor_id) ||
(dst_type == OrtMemoryInfoDeviceType_GPU && dst_vendor_id != nvidia_vendor_id)) {
if ((src_type == OrtMemoryInfoDeviceType_GPU && src_vendor_id != kNvidiaVendorId) ||
(dst_type == OrtMemoryInfoDeviceType_GPU && dst_vendor_id != kNvidiaVendorId)) {
return false;
}

Expand Down
20 changes: 16 additions & 4 deletions src/tensorrt_provider_factory.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -55,7 +55,7 @@ OrtStatus* TensorrtExecutionProviderFactory::CreateMemoryInfoForDevices(int num_
for (int device_id = 0; device_id < num_devices; ++device_id) {
OrtMemoryInfo* mem_info = nullptr;
RETURN_IF_ERROR(ort_api.CreateMemoryInfo_V2("Cuda", OrtMemoryInfoDeviceType_GPU,
/*vendor OrtDevice::VendorIds::NVIDIA*/ 0x10DE,
/* vendor_id */ kNvidiaVendorId,
/* device_id */ device_id, OrtDeviceMemoryType_DEFAULT,
/*alignment*/ 0, OrtAllocatorType::OrtDeviceAllocator, &mem_info));

Expand All @@ -64,7 +64,7 @@ OrtStatus* TensorrtExecutionProviderFactory::CreateMemoryInfoForDevices(int num_
// HOST_ACCESSIBLE memory should use the non-CPU device type
mem_info = nullptr;
RETURN_IF_ERROR(ort_api.CreateMemoryInfo_V2("CudaPinned", OrtMemoryInfoDeviceType_GPU,
/*vendor OrtDevice::VendorIds::NVIDIA*/ 0x10DE,
/* vendor_id */ kNvidiaVendorId,
/* device_id */ device_id, OrtDeviceMemoryType_HOST_ACCESSIBLE,
/*alignment*/ 0, OrtAllocatorType::OrtDeviceAllocator, &mem_info));

Expand All @@ -87,7 +87,15 @@ OrtStatus* ORT_API_CALL TensorrtExecutionProviderFactory::GetSupportedDevicesImp
// Create two memory infos per device.
// The memory info is required to create allocator and gpu data transfer.
int num_cuda_devices = 0;
cudaGetDeviceCount(&num_cuda_devices);
cudaError_t cuda_err = cudaGetDeviceCount(&num_cuda_devices);
if (cuda_err != cudaSuccess) {

@yuslepukhin yuslepukhin Apr 1, 2026

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

the body of the first if block uses 6-space indentation while the rest of the file uses 4-space:

@chilo-ms chilo-ms Apr 2, 2026

Copy link
Copy Markdown
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Made it to use 4-space.

I think i will include lintrunner to have a nice format in other PR.

return factory->ort_api.CreateStatus(ORT_EP_FAIL, cudaGetErrorString(cuda_err));
}

if (num_cuda_devices == 0) {

@yuslepukhin yuslepukhin Apr 1, 2026

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Returning error when num_cuda_devices == 0 is semantically questionable.

GetSupportedDevices is an enumeration function — its job is to report which devices this EP can service. If there are no CUDA devices, the correct behavior is to report 0 supported devices and return success (nullptr), not to return a hard failure. An ORT_EP_FAIL status here may cause ORT to log spurious errors or abort initialization on machines that legitimately have no NVIDIA GPUs (e.g., an AMD-only system where the TRT EP plugin is still present).

@chilo-ms chilo-ms Apr 2, 2026

Copy link
Copy Markdown
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

That makes sense and i made the return success (nullptr) when there is no CUDA device available.

Also, add a check for CUDA devices in CreateEpFactories()

return factory->ort_api.CreateStatus(ORT_EP_FAIL, "No CUDA devices found.");
}

RETURN_IF_ERROR(factory->CreateMemoryInfoForDevices(num_cuda_devices));

int32_t device_id = 0;
Comment thread
chilo-ms marked this conversation as resolved.
Outdated
Expand All @@ -96,7 +104,8 @@ OrtStatus* ORT_API_CALL TensorrtExecutionProviderFactory::GetSupportedDevicesImp
// C API
const OrtHardwareDevice& device = *devices[i];

if (factory->ort_api.HardwareDevice_Type(&device) == OrtHardwareDeviceType::OrtHardwareDeviceType_GPU) {
if (factory->ort_api.HardwareDevice_Type(&device) == OrtHardwareDeviceType::OrtHardwareDeviceType_GPU &&

@yuslepukhin yuslepukhin Apr 1, 2026

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

In tensorrt_provider_factory.cc device_id is a zero-based counter that increments for each NVIDIA GPU found in the ORT hardware device list. This value is then used to index into cuda_gpu_memory_infos (keyed 0..num_cuda_devices-1). The implicit assumption is that the ORT hardware device enumeration order matches the CUDA device ordinals.

If ORT enumerates devices in a different order than CUDA, the device_id mapping could silently use the wrong memory info. The newly added bounds check (device_id < num_cuda_devices) catches overflow, which is good, but doesn't guard against order mismatches.

Suggestion: Consider using the actual CUDA device ID from the OrtHardwareDevice (e.g., via a device-index getter API if available) rather than relying on positional enumeration order.

@chilo-ms chilo-ms Apr 2, 2026

Copy link
Copy Markdown
Collaborator Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I updated the code to use actual CUDA device ID from the OrtHardwareDevice.

factory->ort_api.HardwareDevice_VendorId(&device) == kNvidiaVendorId) {
// These can be returned as nullptr if you have nothing to add.
OrtKeyValuePairs* ep_metadata = nullptr;
OrtKeyValuePairs* ep_options = nullptr;
Expand All @@ -120,6 +129,9 @@ OrtStatus* ORT_API_CALL TensorrtExecutionProviderFactory::GetSupportedDevicesImp
return status;
}

RETURN_IF_NOT(device_id < num_cuda_devices,
"The device_id for supported device exceeds the number of CUDA devices.");

const OrtMemoryInfo* cuda_gpu_mem_info = factory->cuda_gpu_memory_infos[device_id].get();
const OrtMemoryInfo* cuda_pinned_mem_info = factory->cuda_pinned_memory_infos[device_id].get();

Expand Down
2 changes: 2 additions & 0 deletions src/utils/ep_utils.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -23,6 +23,8 @@ struct ApiPtrs {

namespace trt_ep {

constexpr uint32_t kNvidiaVendorId = 0x10DE;

#define ENFORCE(condition, ...) \
do { \
if (!(condition)) { \
Expand Down