Skip to content

[WebGPU] Enable GeluFusion and BiasGeluFusion for the WebGPU EP #32053

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
Hariharan Seshadri (hariharans29) merged 4 commits into from
Aug 19, 2026
Merged

[WebGPU] Enable GeluFusion and BiasGeluFusion for the WebGPU EP #32053

Show file tree
Hide file tree
Changes from 3 commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
9c1a64a
Enable GeluFusion and BiasGeluFusion for the WebGPU EP
fanchenkong1 Jun 22, 2026
634499a
Add ActivationOpTest.Gelu_half
fanchenkong1 Aug 14, 2026
b26a601
Add WebGPU regression tests for Gelu/BiasGelu fusion
fanchenkong1 Aug 18, 2026
bb2370b
Potential fix for pull request finding
fanchenkong1 Aug 18, 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
9 changes: 7 additions & 2 deletions onnxruntime/core/optimizer/graph_transformer_utils.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -349,6 +349,11 @@ InlinedVector<std::unique_ptr<GraphTransformer>> GenerateTransformers(
onnxruntime::kAclExecutionProvider,
onnxruntime::kCudaExecutionProvider,
onnxruntime::kDmlExecutionProvider};
const InlinedHashSet<std::string_view> cpu_acl_cuda_dml_webgpu_eps = {onnxruntime::kCpuExecutionProvider,
onnxruntime::kAclExecutionProvider,
onnxruntime::kCudaExecutionProvider,
onnxruntime::kDmlExecutionProvider,
onnxruntime::kWebGpuExecutionProvider};
const InlinedHashSet<std::string_view> cpu_acl_cuda_dml_js_webgpu_eps = {onnxruntime::kCpuExecutionProvider,
onnxruntime::kAclExecutionProvider,
onnxruntime::kCudaExecutionProvider,
Expand Down Expand Up @@ -401,15 +406,15 @@ InlinedVector<std::unique_ptr<GraphTransformer>> GenerateTransformers(

transformers.emplace_back(std::make_unique<ConvActivationFusion>(cpu_acl_js_webgpu_eps));

transformers.emplace_back(std::make_unique<GeluFusion>(cpu_acl_cuda_dml_eps, level));
transformers.emplace_back(std::make_unique<GeluFusion>(cpu_acl_cuda_dml_webgpu_eps, level));
transformers.emplace_back(std::make_unique<LayerNormFusion>(cpu_acl_cuda_dml_eps, level));
transformers.emplace_back(std::make_unique<SimplifiedLayerNormFusion>(cpu_cuda_eps));
transformers.emplace_back(std::make_unique<AttentionFusion>(cpu_acl_cuda_dml_eps));
transformers.emplace_back(std::make_unique<EmbedLayerNormFusion>(cpu_acl_cuda_dml_eps));
transformers.emplace_back(std::make_unique<GatherSliceToSplitFusion>(cpu_cuda_eps));
transformers.emplace_back(std::make_unique<GatherToSliceFusion>(cpu_cuda_eps));
transformers.emplace_back(std::make_unique<MatmulTransposeFusion>(cpu_cuda_dml_eps));
transformers.emplace_back(std::make_unique<BiasGeluFusion>(cpu_acl_cuda_dml_eps));
transformers.emplace_back(std::make_unique<BiasGeluFusion>(cpu_acl_cuda_dml_webgpu_eps));
transformers.emplace_back(std::make_unique<GroupQueryAttentionFusion>(cuda_eps));
// Run MatMulAddFusion again after *AttentionFusion transforms with `preserve_attention_pattern = false`,
// to cleanup the remaining MatMul-Add that were part of the attention pattern but not detected or fused.
Expand Down
16 changes: 16 additions & 0 deletions onnxruntime/test/contrib_ops/activation_op_test.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -60,6 +60,22 @@ TEST_F(ActivationOpTest, Gelu) {
}
#endif

TEST_F(ActivationOpTest, Gelu_half) {
const std::vector<MLFloat16>& X = input_values_fp16[0];
std::vector<MLFloat16> Y;
Y.reserve(X.size());
for (const MLFloat16& x_half : X) {
const float x = x_half.ToFloat();
Y.push_back(MLFloat16(x * 0.5f * (1.0f + std::erf(x * static_cast<float>(M_SQRT1_2)))));
}

OpTester tester("Gelu", 1, onnxruntime::kMSDomain);
const std::vector<int64_t> dims{1, 1, static_cast<int64_t>(X.size())};
tester.AddInput<MLFloat16>("X", dims, X);
tester.AddOutput<MLFloat16>("Y", dims, Y);
tester.Run(OpTester::ExpectResult::kExpectSuccess, "", {kTensorrtExecutionProvider});
}

#if defined(USE_DNNL)
std::vector<BFloat16> expected_output_bfloat16(const std::vector<float>& input_data) {
std::vector<float> output;
Expand Down
99 changes: 99 additions & 0 deletions onnxruntime/test/optimizer/graph_transform_test.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -7330,6 +7330,105 @@ TEST_F(GraphTransformationTests, BiasGeluFusionCurrentOpsetTest) {
ModelOptions{kAllowReleasedOpsetsOnly, /*strict_shape_type_inference*/ false}));
}

#if !defined(DISABLE_CONTRIB_OPS)
// Regression test for the WebGPU entry added to the Level-2 GeluFusion allowlist
// (cpu_acl_cuda_dml_webgpu_eps in graph_transformer_utils.cc).
TEST_F(GraphTransformationTests, GeluFusionWebGpu) {
constexpr const ORTCHAR_T* model_uri = MODEL_FOLDER "fusion/gelu.onnx";
std::shared_ptr<Model> p_model;
ASSERT_STATUS_OK(Model::Load(model_uri, p_model, nullptr, *logger_));
Graph& graph = p_model->MainGraph();
#if defined(USE_WEBGPU)
const std::string expected_ep = kWebGpuExecutionProvider;
#else
const std::string expected_ep = kCpuExecutionProvider;
#endif
for (auto& node : graph.Nodes()) {
node.SetExecutionProviderType(expected_ep);
}

SessionOptions session_options;
auto cpu_ep = std::make_unique<CPUExecutionProvider>(CPUExecutionProviderInfo());
const InlinedHashSet<std::string> gelu_transformer_names = {"GeluFusionL1", "GeluFusionL2"};
onnxruntime::GraphTransformerManager graph_transformation_mgr{5};
for (auto level : {TransformerLevel::Level1, TransformerLevel::Level2}) {
for (auto& transformer : optimizer_utils::GenerateTransformers(level, session_options, *cpu_ep, *logger_, {})) {
if (gelu_transformer_names.count(transformer->Name()) != 0) {
ASSERT_STATUS_OK(graph_transformation_mgr.Register(std::move(transformer), level));
}
}
}
ASSERT_STATUS_OK(graph_transformation_mgr.ApplyTransformers(graph, TransformerLevel::Level1, *logger_));
ASSERT_STATUS_OK(graph_transformation_mgr.ApplyTransformers(graph, TransformerLevel::Level2, *logger_));

std::map<std::string, int> op_to_count = CountOpsInGraph(graph);
ASSERT_EQ(op_to_count["com.microsoft.Gelu"], 1);
ASSERT_EQ(op_to_count["Div"], 0);
ASSERT_EQ(op_to_count["Erf"], 0);
ASSERT_EQ(op_to_count["Add"], 0);
ASSERT_EQ(op_to_count["Mul"], 0);

const Node* gelu_node = nullptr;
for (auto& node : graph.Nodes()) {
if (node.OpType() == "Gelu" && node.Domain() == kMSDomain) {
gelu_node = &node;
break;
}
}
ASSERT_NE(gelu_node, nullptr);
EXPECT_EQ(gelu_node->GetExecutionProviderType(), expected_ep);
}

// Regression test for the WebGPU entry added to the BiasGeluFusion allowlist.
TEST_F(GraphTransformationTests, BiasGeluFusionWebGpu) {
constexpr const ORTCHAR_T* model_uri = MODEL_FOLDER "fusion/bias_gelu_fusion.onnx";
std::shared_ptr<Model> p_model;
ASSERT_STATUS_OK(Model::Load(model_uri, p_model, nullptr, *logger_));
Graph& graph = p_model->MainGraph();
#if defined(USE_WEBGPU)
const std::string expected_ep = kWebGpuExecutionProvider;
#else
const std::string expected_ep = kCpuExecutionProvider;
#endif
for (auto& node : graph.Nodes()) {
node.SetExecutionProviderType(expected_ep);
}

SessionOptions session_options;
auto cpu_ep = std::make_unique<CPUExecutionProvider>(CPUExecutionProviderInfo());
const InlinedHashSet<std::string> gelu_transformer_names = {"GeluFusionL1", "BiasGeluFusion"};
Comment thread
fanchenkong1 marked this conversation as resolved.
Resolved
onnxruntime::GraphTransformerManager graph_transformation_mgr{5};
for (auto level : {TransformerLevel::Level1, TransformerLevel::Level2}) {
for (auto& transformer : optimizer_utils::GenerateTransformers(level, session_options, *cpu_ep, *logger_, {})) {
if (gelu_transformer_names.count(transformer->Name()) != 0) {
ASSERT_STATUS_OK(graph_transformation_mgr.Register(std::move(transformer), level));
}
}
}
ASSERT_STATUS_OK(graph_transformation_mgr.ApplyTransformers(graph, TransformerLevel::Level1, *logger_));
ASSERT_STATUS_OK(graph_transformation_mgr.ApplyTransformers(graph, TransformerLevel::Level2, *logger_));

std::map<std::string, int> op_to_count = CountOpsInGraph(graph);
ASSERT_EQ(op_to_count["com.microsoft.BiasGelu"], 1);
ASSERT_EQ(op_to_count["com.microsoft.Gelu"], 0);
ASSERT_EQ(op_to_count["Gelu"], 0);
ASSERT_EQ(op_to_count["Add"], 0);
ASSERT_EQ(op_to_count["Div"], 0);
ASSERT_EQ(op_to_count["Erf"], 0);
ASSERT_EQ(op_to_count["Mul"], 0);

const Node* bias_gelu_node = nullptr;
for (auto& node : graph.Nodes()) {
if (node.OpType() == "BiasGelu" && node.Domain() == kMSDomain) {
bias_gelu_node = &node;
break;
}
}
ASSERT_NE(bias_gelu_node, nullptr);
EXPECT_EQ(bias_gelu_node->GetExecutionProviderType(), expected_ep);
}
#endif // !defined(DISABLE_CONTRIB_OPS)

TEST_F(GraphTransformationTests, MatMulAddFusionCurrentOpsetTest) {
// MatMul + Add -> Gemm fusion
int current_opset = GetCurrentOnnxOpset();
Expand Down