Skip to content

fix(optimizer): guard Relu quant fusion for non-positive scales #33988

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

Open
自分 (2Zest) wants to merge 1 commit into
base: main
Choose a base branch
from
Open

fix(optimizer): guard Relu quant fusion for non-positive scales #33988

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
14 changes: 14 additions & 0 deletions onnxruntime/core/optimizer/qdq_transformer/relu_quantizelinear.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -45,6 +45,20 @@ Status ReluQuantFusion::Apply(Graph& graph, Node& node, RewriteRuleEffect& rule_
return Status::OK();
}

constexpr size_t scale_idx = 1;
const ONNX_NAMESPACE::TensorProto* scale_tensor_proto = nullptr;
if (!graph_utils::NodeArgIsConstant(graph, *q_input_defs[scale_idx]) ||
!graph.GetInitializedTensor(q_input_defs[scale_idx]->Name(), scale_tensor_proto)) {
return Status::OK();
}

Initializer scale(graph, *scale_tensor_proto, graph.ModelPath());
if (scale.size() != 1 ||
scale.data_type() != ONNX_NAMESPACE::TensorProto::DataType::TensorProto_DataType_FLOAT ||
scale.data<float>()[0] <= 0.0f) {
return Status::OK();
}

using ONNX_TENSOR_ELEM_TYPE = ONNX_NAMESPACE::TensorProto::DataType;
Initializer zero_point(graph, *zp_tensor_proto, graph.ModelPath());
if (zero_point.size() != 1 ||
Expand Down
30 changes: 30 additions & 0 deletions onnxruntime/test/optimizer/qdq_transformer_test.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3996,6 +3996,36 @@ TEST(QDQTransformerTests, ReluQuantFusion_Level2Only) {
test_case(TransformerLevel::Level3, 0); // Will not fuse Relu into QuantizeLinear due to zero-point != -128
}

TEST(QDQTransformerTests, ReluQuantFusionSkipsNonPositiveScale) {
auto test_case = [](float scale, bool scale_is_initializer, bool expect_fusion) {
auto build_test_case = [scale, scale_is_initializer](ModelTestBuilder& builder) {
auto* input_arg = builder.MakeInput<float>({1, 2, 2, 2}, -4.0f, 4.0f);
auto* output_arg = builder.MakeOutput();
auto* relu_output = builder.MakeIntermediate();
builder.AddNode("Relu", {input_arg}, {relu_output});

NodeArg* scale_arg = scale_is_initializer
? builder.MakeScalarInitializer<float>(scale)
: builder.MakeInput<float>({1}, scale, scale);
auto* zero_point = builder.MakeScalarInitializer<uint8_t>(0);
builder.AddNode("QuantizeLinear", {relu_output, scale_arg, zero_point}, {output_arg});
};

auto check_graph = [expect_fusion](InferenceSessionWrapper& session) {
auto op_to_count = CountOpsInGraph(session.GetGraph());
EXPECT_EQ(op_to_count["Relu"], expect_fusion ? 0 : 1);
};

TransformerTester(build_test_case, check_graph, TransformerLevel::Level1,
TransformerLevel::Level2);
};

test_case(0.5f, true, true);
test_case(0.0f, true, false);
test_case(-0.5f, true, false);
test_case(-0.5f, false, false);
}

// Test skip removing node when min/max come from DequantizeLinear nodes instead of initializers.
TEST(QDQTransformerTests, ClipQuantFusion_MultipleInputEdges) {
auto build_test_case = [&](ModelTestBuilder& builder) {
Expand Down