Skip to content

add: detect Q/DQ with int16/uint16 initializers for GPU Scale Transform Pass #768

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
wants to merge 1 commit into
base: ovep-develop
Choose a base branch
from
Open

add: detect Q/DQ with int16/uint16 initializers for GPU Scale Transform Pass #768

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
44 changes: 43 additions & 1 deletion onnxruntime/core/providers/openvino/backend_manager.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -387,6 +387,44 @@ static bool IsModelBF16(const onnxruntime::GraphViewer& graph_viewer) {
return false;
}

static bool Is16BitTensor(const onnxruntime::NodeArg* node_arg) {
const auto* type_proto = node_arg ? node_arg->TypeAsProto() : nullptr;
return type_proto && type_proto->has_tensor_type() &&
(type_proto->tensor_type().elem_type() == ONNX_NAMESPACE::TensorProto_DataType_UINT16 ||
type_proto->tensor_type().elem_type() == ONNX_NAMESPACE::TensorProto_DataType_INT16);
}

// Check to see if the graph has Q/DQ nodes with int16 or uint16 quantization
static bool IsQDQGraphWithUint16OrInt16(const onnxruntime::GraphViewer& graph_viewer) {
std::unordered_set<std::string> qdq_ops = {"QuantizeLinear", "DequantizeLinear"};
const auto& node_indices = graph_viewer.GetNodesInTopologicalOrder();

for (size_t i = 0; i < node_indices.size(); i++) {
gsl::not_null<const onnxruntime::Node*> node(graph_viewer.GetNode(node_indices[i]));

if (qdq_ops.find(node->OpType()) != qdq_ops.end()) {
const auto& input_defs = node->InputDefs();

if (node->OpType() == "DequantizeLinear") {
// DequantizeLinear: [quantized_input, scale, zero_point] -> [float_output]
// Check quantized input tensor and optional zero point
if (Is16BitTensor(input_defs.empty() ? nullptr : input_defs[0]) ||
(input_defs.size() >= 3 && Is16BitTensor(input_defs[2]))) {
return true;
}
} else if (node->OpType() == "QuantizeLinear") {
// QuantizeLinear: [float_input, scale, zero_point] -> [quantized_output]
const auto& output_defs = node->OutputDefs();
if (Is16BitTensor(output_defs.empty() ? nullptr : output_defs[0]) ||
(input_defs.size() >= 3 && Is16BitTensor(input_defs[2]))) {
return true;
}
}
}
}
return false;
}

static void DumpOpenVINOEPModel([[maybe_unused]] const std::filesystem::path& onnx_model_path_name,
[[maybe_unused]] ONNX_NAMESPACE::ModelProto* model_proto,
[[maybe_unused]] const onnxruntime::Node& fused_node) {
Expand Down Expand Up @@ -445,6 +483,10 @@ BackendManager::GetModelProtoFromFusedNode(const onnxruntime::Node& fused_node,
}
#endif

// Check if the graph is QDQ and has int16 or uint16 quantization
// If so, we will apply the QDQ scales fix transformation (for GPU device only)
bool is_qdq_graph_uint16_or_int16 = IsQDQGraphWithUint16OrInt16(subgraph);

const auto& onnx_model_path_name = subgraph.ModelPath();
// QDQ stripping enabled only for the NPU and experimentally on the GPU
if ((session_context_.device_type.find("NPU") != std::string::npos) &&
Expand All @@ -458,7 +500,7 @@ BackendManager::GetModelProtoFromFusedNode(const onnxruntime::Node& fused_node,
ORT_ENFORCE(status.IsOK(), status.ErrorMessage());
return model_proto;
} else if ((session_context_.device_type.find("GPU") != std::string::npos) &&
enable_ovep_qdq_optimizer) {
is_qdq_graph_uint16_or_int16) {
// Create a copy of the model
std::unique_ptr<onnxruntime::Model> model;
Status status = qdq_scales_fix::Transform(subgraph, logger, model);
Expand Down
Loading