Skip to content

Validate FastGelu fusion scale node #32016

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 5 commits into from
Aug 14, 2026
Merged

Validate FastGelu fusion scale node #32016

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
f50effb
Validate FastGelu fusion scale node
apsonawane Aug 12, 2026
c8436d4
Harden FastGelu fusion input matching
apsonawane Aug 12, 2026
9bb7e7b
Scope FastGelu scale input index
apsonawane Aug 13, 2026
4d098b6
Narrow FastGelu input indices explicitly
apsonawane Aug 13, 2026
7c2a928
Fix lint
apsonawane Aug 13, 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
83 changes: 58 additions & 25 deletions onnxruntime/core/optimizer/fast_gelu_fusion.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -7,6 +7,7 @@
#include "core/graph/graph_utils.h"
#include "float.h"
#include <deque>
#include <optional>

Check warning on line 10 in onnxruntime/core/optimizer/fast_gelu_fusion.cc

View workflow job for this annotation

GitHub Actions / Optional Lint C++ 

[cpplint] reported by reviewdog 🐶
 Found C++ system header after other header. Should be: fast_gelu_fusion.h, c system, c++ system, other.  [build/include_order] [4]

Raw Output:
onnxruntime/core/optimizer/fast_gelu_fusion.cc:10:  Found C++ system header after other header. Should be: fast_gelu_fusion.h, c system, c++ system, other.  [build/include_order] [4]

using namespace ONNX_NAMESPACE;
using namespace onnxruntime::common;
Expand All @@ -32,10 +33,24 @@
!graph.NodeProducesGraphOutput(node);
}

static std::optional<size_t> OtherInputIndex(const Node& node, const NodeArg& input) {
if (node.InputDefs().size() != 2) {
return std::nullopt;
}

const int input_index = optimizer_utils::IndexOfNodeInput(node, input);
if (input_index < 0 || input_index >= 2) {
return std::nullopt;
}

return 1u - onnxruntime::narrow<size_t>(input_index);
}

MatchResult FastGeluFusion::CheckFirstFormula(Graph& graph, Node& mul1_node,
InlinedVector<std::reference_wrapper<Node>>& nodes_to_fuse) const {
MatchResult matchResult{false, nullptr, nullptr};
if (!graph_utils::IsSupportedOptypeVersionAndDomain(mul1_node, "Mul", {7, 13, 14}) ||
mul1_node.InputDefs().size() != 2 ||
!graph_utils::IsSupportedProvider(mul1_node, GetCompatibleExecutionProviders()) ||
mul1_node.GetOutputEdgesCount() != 1 ||
!IsSupportedDataType(mul1_node)) {
Comment thread
apsonawane marked this conversation as resolved.
Resolved
Expand All @@ -53,24 +68,26 @@

if (input_index == -1) return matchResult;

NodeArg* gelu_without_bias_input_arg = mul1_node.MutableInputDefs()[(input_index + 1) % 2];
NodeArg* gelu_without_bias_input_arg = mul1_node.MutableInputDefs()[1u - onnxruntime::narrow<size_t>(input_index)];
nodes_to_fuse.push_back(mul1_node);

Node& mul2_node = *graph.GetNode(mul1_node.OutputNodesBegin()->Index());
input_index = optimizer_utils::IndexOfNodeInput(mul2_node, *mul1_node.MutableOutputDefs()[0]);
auto other_input_index = OtherInputIndex(mul2_node, *mul1_node.MutableOutputDefs()[0]);
if (!(graph_utils::IsSupportedOptypeVersionAndDomain(mul2_node, "Mul", {7, 13, 14}) &&
CheckNode(graph, mul2_node, mul1_node.GetExecutionProviderType(), true)) ||
mul2_node.MutableInputDefs()[(input_index + 1) % 2]->Name() != gelu_without_bias_input_arg->Name()) {
!other_input_index ||
mul2_node.InputDefs()[*other_input_index]->Name() != gelu_without_bias_input_arg->Name()) {
return matchResult;
}
nodes_to_fuse.push_back(mul2_node);

Node& add1_node = *graph.GetNode(mul2_node.OutputNodesBegin()->Index());
input_index = optimizer_utils::IndexOfNodeInput(add1_node, *mul2_node.MutableOutputDefs()[0]);
other_input_index = OtherInputIndex(add1_node, *mul2_node.MutableOutputDefs()[0]);

if (!(graph_utils::IsSupportedOptypeVersionAndDomain(add1_node, "Add", {7, 13, 14}) &&
CheckNode(graph, add1_node, mul1_node.GetExecutionProviderType(), true)) ||
!optimizer_utils::IsInitializerWithExpectedValue(graph, *(add1_node.InputDefs()[(input_index + 1) % 2]), 1.0f, true)) {
!other_input_index ||
!optimizer_utils::IsInitializerWithExpectedValue(graph, *add1_node.InputDefs()[*other_input_index], 1.0f, true)) {
return matchResult;
}
nodes_to_fuse.push_back(add1_node);
Expand All @@ -82,11 +99,14 @@
}
nodes_to_fuse.push_back(mul3_node);

input_index = optimizer_utils::IndexOfNodeInput(mul3_node, *add1_node.MutableOutputDefs()[0]);
const Node* p_mul3_input_node = graph_utils::GetInputNode(mul3_node, (input_index + 1) % 2);
other_input_index = OtherInputIndex(mul3_node, *add1_node.MutableOutputDefs()[0]);
if (!other_input_index) return matchResult;
const Node* p_mul3_input_node =
graph_utils::GetInputNode(mul3_node, onnxruntime::narrow<int>(*other_input_index));
if (p_mul3_input_node == nullptr) return matchResult;
Node& mul4_node = const_cast<Node&>(*p_mul3_input_node);
if (!(graph_utils::IsSupportedOptypeVersionAndDomain(mul3_node, "Mul", {7, 13, 14}) &&
if (!(graph_utils::IsSupportedOptypeVersionAndDomain(mul4_node, "Mul", {7, 13, 14}) &&
mul4_node.InputDefs().size() == 2 &&
CheckNode(graph, mul4_node, mul1_node.GetExecutionProviderType(), true))) {
return matchResult;
}
Expand All @@ -100,7 +120,8 @@
}
}

if (input_index == -1 || mul4_node.InputDefs()[(input_index + 1) % 2]->Name() != gelu_without_bias_input_arg->Name())
if (input_index == -1 ||
mul4_node.InputDefs()[1u - onnxruntime::narrow<size_t>(input_index)]->Name() != gelu_without_bias_input_arg->Name())
return matchResult;
nodes_to_fuse.push_back(mul4_node);

Expand All @@ -114,6 +135,7 @@
InlinedVector<std::reference_wrapper<Node>>& nodes_to_fuse) const {
MatchResult matchResult{false, nullptr, nullptr};
if (!graph_utils::IsSupportedOptypeVersionAndDomain(pow1_node, "Pow", {7, 12, 13, 15}) ||
pow1_node.InputDefs().size() != 2 ||
!graph_utils::IsSupportedProvider(pow1_node, GetCompatibleExecutionProviders()) ||
pow1_node.GetOutputEdgesCount() != 1 ||
!IsSupportedDataType(pow1_node)) {
Expand All @@ -128,20 +150,22 @@
nodes_to_fuse.push_back(pow1_node);

Node& mul1_node = *graph.GetNode(pow1_node.OutputNodesBegin()->Index());
auto input_index = optimizer_utils::IndexOfNodeInput(mul1_node, *pow1_node.MutableOutputDefs()[0]);
auto other_input_index = OtherInputIndex(mul1_node, *pow1_node.MutableOutputDefs()[0]);
if (!(graph_utils::IsSupportedOptypeVersionAndDomain(mul1_node, "Mul", {7, 13, 14}) &&
CheckNode(graph, mul1_node, pow1_node.GetExecutionProviderType(), true)) ||
!optimizer_utils::IsInitializerWithExpectedValue(graph, *(mul1_node.InputDefs()[(input_index + 1) % 2]),
!other_input_index ||
!optimizer_utils::IsInitializerWithExpectedValue(graph, *mul1_node.InputDefs()[*other_input_index],
0.044714998453855515f, true)) {
return matchResult;
}
nodes_to_fuse.push_back(mul1_node);

Node& add1_node = *graph.GetNode(mul1_node.OutputNodesBegin()->Index());
input_index = optimizer_utils::IndexOfNodeInput(add1_node, *mul1_node.MutableOutputDefs()[0]);
other_input_index = OtherInputIndex(add1_node, *mul1_node.MutableOutputDefs()[0]);
if (!(graph_utils::IsSupportedOptypeVersionAndDomain(add1_node, "Add", {7, 13, 14}) &&
CheckNode(graph, add1_node, pow1_node.GetExecutionProviderType(), true)) ||
add1_node.MutableInputDefs()[(input_index + 1) % 2]->Name() != pow_input_arg->Name()) {
!other_input_index ||
add1_node.InputDefs()[*other_input_index]->Name() != pow_input_arg->Name()) {
return matchResult;
}
nodes_to_fuse.push_back(add1_node);
Expand All @@ -152,6 +176,7 @@
Node& cast1_node = *graph.GetNode(p_cast1_node->Index());
// this is fused Cast node, so expect 2 output edges
if (!(graph_utils::IsSupportedOptypeVersionAndDomain(cast1_node, "Cast", {9, 13, 19, 21, 23, 24, 25}) &&
cast1_node.InputDefs().size() == 1 &&
CheckNode(graph, cast1_node, pow1_node.GetExecutionProviderType(), false)) ||
cast1_node.GetOutputEdgesCount() != 2) {
return matchResult;
Expand All @@ -167,10 +192,11 @@
}

Node& mul2_node = *graph.GetNode(add1_node.OutputNodesBegin()->Index());
input_index = optimizer_utils::IndexOfNodeInput(mul2_node, *add1_node.MutableOutputDefs()[0]);
other_input_index = OtherInputIndex(mul2_node, *add1_node.MutableOutputDefs()[0]);
if (!(graph_utils::IsSupportedOptypeVersionAndDomain(mul2_node, "Mul", {7, 13, 14}) &&
CheckNode(graph, mul2_node, pow1_node.GetExecutionProviderType(), true)) ||
!optimizer_utils::IsInitializerWithExpectedValue(graph, *(mul2_node.InputDefs()[(input_index + 1) % 2]),
!other_input_index ||
!optimizer_utils::IsInitializerWithExpectedValue(graph, *mul2_node.InputDefs()[*other_input_index],
0.7978845834732056f, true)) {
return matchResult;
}
Expand Down Expand Up @@ -233,8 +259,9 @@
continue;
}

auto input_index = optimizer_utils::IndexOfNodeInput(add2_node, *tanh_node.MutableOutputDefs()[0]);
if (!optimizer_utils::IsInitializerWithExpectedValue(graph, *(add2_node.InputDefs()[(input_index + 1) % 2]), 1.0f, true)) {
auto other_input_index = OtherInputIndex(add2_node, *tanh_node.MutableOutputDefs()[0]);
if (!other_input_index ||
!optimizer_utils::IsInitializerWithExpectedValue(graph, *add2_node.InputDefs()[*other_input_index], 1.0f, true)) {
continue;
}

Expand All @@ -245,8 +272,10 @@
continue;
}

input_index = optimizer_utils::IndexOfNodeInput(mul5_node, *add2_node.MutableOutputDefs()[0]);
const Node* p_mul5_input_node = graph_utils::GetInputNode(mul5_node, (input_index + 1) % 2);
other_input_index = OtherInputIndex(mul5_node, *add2_node.MutableOutputDefs()[0]);
if (!other_input_index) continue;
const Node* p_mul5_input_node =
graph_utils::GetInputNode(mul5_node, onnxruntime::narrow<int>(*other_input_index));
if (p_mul5_input_node == nullptr) continue;

// if this is second formula and if pow node has Cast parent, expect mul5_node has Cast parent as well
Expand All @@ -263,37 +292,41 @@

Node& cast3_node = *graph.GetNode(p_cast3_node->Index());
if (!(graph_utils::IsSupportedOptypeVersionAndDomain(cast3_node, "Cast", {9, 13, 19, 21, 23, 24, 25}) &&
cast3_node.InputDefs().size() == 1 &&
CheckNode(graph, cast3_node, node.GetExecutionProviderType(), true))) {
continue;
}
// overwrite and continue as usual
p_mul5_input_node = graph_utils::FirstParentByType(cast3_node, "Mul");
if (p_mul5_input_node == nullptr) continue;
nodes_to_fuse.push_back(cast3_node);
// keep cast1_node for reuse, its output edges will be adjusted in FinalizeNodeFusion()
}
}

Node& mul6_node = const_cast<Node&>(*p_mul5_input_node);
if (!(graph_utils::IsSupportedOptypeVersionAndDomain(mul6_node, "Mul", {7, 13, 14}) &&
mul6_node.InputDefs().size() == 2 &&
CheckNode(graph, mul6_node, node.GetExecutionProviderType(), false))) {
continue;
}

input_index = -1;
for (auto i = 0; i < 2; i++) {
std::optional<size_t> scale_input_index;
for (size_t i = 0; i < 2; ++i) {
if (optimizer_utils::IsInitializerWithExpectedValue(graph, *(mul6_node.InputDefs()[i]), 0.5f, true)) {
input_index = i;
scale_input_index = i;
break;
}
}

if (input_index == -1) continue;
if (!scale_input_index) continue;
const size_t mul6_other_input_index = 1u - *scale_input_index;
// check same parent for both mul6 and pow, with or without cast
if (cast_input_arg != nullptr) {
if (mul6_node.InputDefs()[(input_index + 1) % 2]->Name() != cast_input_arg->Name())
if (mul6_node.InputDefs()[mul6_other_input_index]->Name() != cast_input_arg->Name())
continue;
} else {
if (mul6_node.InputDefs()[(input_index + 1) % 2]->Name() != matchRet.gelu_without_bias_input_arg->Name())
if (mul6_node.InputDefs()[mul6_other_input_index]->Name() != matchRet.gelu_without_bias_input_arg->Name())
continue;
}

Expand Down
89 changes: 89 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 @@ -7905,6 +7905,95 @@ TEST_F(GraphTransformationTests, FastGeluFusionTest) {
ASSERT_TRUE(op_to_count["com.microsoft.FastGelu"] == 1);
}

TEST_F(GraphTransformationTests, FastGeluFusionSkipsMalformedScaleMul) {
constexpr const ORTCHAR_T* model_uri = MODEL_FOLDER "fusion/fast_gelu.onnx";
std::shared_ptr<Model> model;
ASSERT_STATUS_OK(Model::Load(model_uri, model, nullptr, *logger_));
Graph& graph = model->MainGraph();

Node* scale_mul = nullptr;
for (auto& node : graph.Nodes()) {
if (node.OpType() != "Mul" || node.InputDefs().size() != 2) {
continue;
}

for (const NodeArg* input : node.InputDefs()) {
if (optimizer_utils::IsInitializerWithExpectedValue(graph, *input, 0.7978845834732056f, true)) {
scale_mul = &node;
break;
}
}

if (scale_mul != nullptr) {
break;
}
}

ASSERT_NE(scale_mul, nullptr);
scale_mul->MutableInputDefs().pop_back();

GraphTransformerManager graph_transformation_mgr{5};
ASSERT_STATUS_OK(graph_transformation_mgr.Register(std::make_unique<FastGeluFusion>(), TransformerLevel::Level2));
ASSERT_STATUS_OK(graph_transformation_mgr.ApplyTransformers(graph, TransformerLevel::Level2, *logger_));

std::map<std::string, int> op_to_count = CountOpsInGraph(graph);
EXPECT_EQ(op_to_count["com.microsoft.FastGelu"], 0);
}

TEST_F(GraphTransformationTests, FastGeluFusionSkipsMalformedFirstFormulaIntermediateMul) {
constexpr const ORTCHAR_T* model_uri = MODEL_FOLDER "fusion/fast_gelu.onnx";
std::shared_ptr<Model> model;
ASSERT_STATUS_OK(Model::Load(model_uri, model, nullptr, *logger_));
Graph& graph = model->MainGraph();

Node* coefficient_mul = nullptr;
for (auto& node : graph.Nodes()) {
if (node.OpType() == "Mul" && node.InputDefs().size() == 2 &&
(optimizer_utils::IsInitializerWithExpectedValue(graph, *node.InputDefs()[0], 0.044715f, true) ||
optimizer_utils::IsInitializerWithExpectedValue(graph, *node.InputDefs()[1], 0.044715f, true))) {
coefficient_mul = &node;
break;
}
}

ASSERT_NE(coefficient_mul, nullptr);
ASSERT_EQ(coefficient_mul->GetOutputEdgesCount(), 1);
Node& intermediate_mul = *graph.GetNode(coefficient_mul->OutputNodesBegin()->Index());
intermediate_mul.MutableInputDefs().pop_back();

GraphTransformerManager graph_transformation_mgr{5};
ASSERT_STATUS_OK(graph_transformation_mgr.Register(std::make_unique<FastGeluFusion>(), TransformerLevel::Level2));
ASSERT_STATUS_OK(graph_transformation_mgr.ApplyTransformers(graph, TransformerLevel::Level2, *logger_));

EXPECT_EQ(CountOpsInGraph(graph)["com.microsoft.FastGelu"], 0);
}

TEST_F(GraphTransformationTests, FastGeluFusionSkipsMalformedSecondFormulaIntermediateMul) {
constexpr const ORTCHAR_T* model_uri = MODEL_FOLDER "fusion/fast_gelu2.onnx";
std::shared_ptr<Model> model;
ASSERT_STATUS_OK(Model::Load(model_uri, model, nullptr, *logger_));
Graph& graph = model->MainGraph();

Node* pow_node = nullptr;
for (auto& node : graph.Nodes()) {
if (node.OpType() == "Pow") {
pow_node = &node;
break;
}
}

ASSERT_NE(pow_node, nullptr);
ASSERT_EQ(pow_node->GetOutputEdgesCount(), 1);
Node& intermediate_mul = *graph.GetNode(pow_node->OutputNodesBegin()->Index());
intermediate_mul.MutableInputDefs().pop_back();

GraphTransformerManager graph_transformation_mgr{5};
ASSERT_STATUS_OK(graph_transformation_mgr.Register(std::make_unique<FastGeluFusion>(), TransformerLevel::Level2));
ASSERT_STATUS_OK(graph_transformation_mgr.ApplyTransformers(graph, TransformerLevel::Level2, *logger_));

EXPECT_EQ(CountOpsInGraph(graph)["com.microsoft.FastGelu"], 0);
}

TEST_F(GraphTransformationTests, FastGeluUseGraphInputFusionTest) {
constexpr const ORTCHAR_T* model_uri = MODEL_FOLDER "fusion/fast_gelu_use_graph_input.onnx";
std::shared_ptr<Model> p_model;
Expand Down
Loading