Skip to content

Handle empty tensors in CUDA InstanceNormalization #31647

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
Akshay Sonawane (apsonawane) merged 2 commits into from
Aug 5, 2026
Merged

Handle empty tensors in CUDA InstanceNormalization #31647

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
419d10e
Handle empty tensors in CUDA InstanceNormalization
apsonawane Aug 4, 2026
d617362
Fix missing CUDA include in InstanceNormalization test
apsonawane Aug 5, 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
12 changes: 12 additions & 0 deletions onnxruntime/core/providers/cuda/nn/instance_norm.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -46,6 +46,12 @@ Status InstanceNorm<T>::ComputeInternal(OpKernelContext* p_op_kernel_context) co
const TensorShape& x_shape = X->Shape();
Tensor* Y = p_op_kernel_context->Output(0, x_shape);

// Empty inputs are valid in ONNX and should produce empty outputs.
// Early return avoids stats_count == 0 in the N != 1 path.
if (x_shape.Size() == 0) {
return Status::OK();
}

auto* y_data = reinterpret_cast<CudaT*>(Y->MutableData<T>());
const auto* x_data = reinterpret_cast<const CudaT*>(X->Data<T>());
const auto* scale_data = reinterpret_cast<const CudaT*>(scale->Data<T>());
Expand Down Expand Up @@ -174,6 +180,12 @@ Status InstanceNorm<MLFloat16>::ComputeInternal(OpKernelContext* p_op_kernel_con
const TensorShape& x_shape = X->Shape();
Tensor* Y = p_op_kernel_context->Output(0, x_shape);

// Empty inputs are valid in ONNX and should produce empty outputs.
// Early return avoids stats_count == 0 in the N != 1 path.
if (x_shape.Size() == 0) {
return Status::OK();
}

auto* y_data = reinterpret_cast<CudaT*>(Y->MutableData<MLFloat16>());
const auto* x_data = reinterpret_cast<const CudaT*>(X->Data<MLFloat16>());
const auto* scale_data = reinterpret_cast<const CudaT*>(scale->Data<MLFloat16>());
Expand Down
43 changes: 43 additions & 0 deletions onnxruntime/test/providers/cpu/nn/instance_norm_op_test.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -6,6 +6,10 @@
#include "test/common/tensor_op_test_utils.h"
#include "test/util/include/default_providers.h"

#if defined(USE_CUDA)
#include "test/common/cuda_op_test_utils.h"
#endif

using namespace std;
namespace onnxruntime {
namespace test {
Expand Down Expand Up @@ -224,6 +228,45 @@ TEST(InstanceNormalizationOpTest, InstanceNormBatch2_fp16) {
}

#endif

#ifdef USE_CUDA
TEST(InstanceNormalizationOpTest, InstanceNormEmptyChannel_Cuda) {
if (!HasCudaEnvironment(0)) {
GTEST_SKIP() << "CUDA not available";
}

OpTester test("InstanceNormalization");
test.AddAttribute("epsilon", 0.3F);

test.AddInput<float>("input", {2, 0, 4, 4}, {});
test.AddInput<float>("scale", {0}, {});
test.AddInput<float>("B", {0}, {});
test.AddOutput<float>("Y", {2, 0, 4, 4}, {});

std::vector<std::unique_ptr<IExecutionProvider>> cuda_only_ep;
cuda_only_ep.push_back(DefaultCudaExecutionProvider());
test.Run(OpTester::ExpectResult::kExpectSuccess, "", {}, nullptr, &cuda_only_ep);
}

TEST(InstanceNormalizationOpTest, InstanceNormEmptyBatch_Cuda) {
if (!HasCudaEnvironment(0)) {
GTEST_SKIP() << "CUDA not available";
}

OpTester test("InstanceNormalization");
test.AddAttribute("epsilon", 0.3F);

test.AddInput<float>("input", {0, 3, 4, 4}, {});
test.AddInput<float>("scale", {3}, {1.0f, 1.0f, 1.0f});
test.AddInput<float>("B", {3}, {0.0f, 0.0f, 0.0f});
test.AddOutput<float>("Y", {0, 3, 4, 4}, {});

std::vector<std::unique_ptr<IExecutionProvider>> cuda_only_ep;
cuda_only_ep.push_back(DefaultCudaExecutionProvider());
test.Run(OpTester::ExpectResult::kExpectSuccess, "", {}, nullptr, &cuda_only_ep);
}
#endif

TEST(InstanceNormalizationOpTest, InstanceNorm_2) {
OpTester test("InstanceNormalization");
test.AddAttribute("epsilon", 0.3F);
Expand Down
Loading