remove bfloat16 silence conversion

Author: bopeng1234

onnxruntime/core/providers/openvino/backend_manager.cc

@@ -389,18 +389,6 @@ static bool IsQDQGraph(const onnxruntime::GraphViewer& graph_viewer) {
   return false;
 }
 
-static bool IsModelBF16(const onnxruntime::GraphViewer& graph_viewer) {
-  const auto& node_indices = graph_viewer.GetNodesInTopologicalOrder();
-  for (std::size_t i = 0; i < node_indices.size(); i++) {
-    gsl::not_null<const onnxruntime::Node*> node(graph_viewer.GetNode(node_indices[i]));
-    for (auto& output : node->OutputDefs()) {
-      if (output->ToProto().type().tensor_type().elem_type() == ONNX_NAMESPACE::TensorProto_DataType_BFLOAT16)
-        return true;
-    }
-  }
-  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() &&
@@ -598,16 +586,6 @@ BackendManager::GetModelProtoFromFusedNode(const onnxruntime::Node& fused_node,
     DumpOpenVINOEPModel(onnx_model_path_name, model_proto.get(), fused_node);
     ORT_ENFORCE(status.IsOK(), status.ErrorMessage());
     return model_proto;
-  } else if ((session_context_.device_type.find("CPU") == std::string::npos) && IsModelBF16(subgraph)) {
-    LOGS_DEFAULT(INFO) << "[OpenVINO-EP] OVEP bfloat16->float16 optimization pass is enabled";
-    std::unique_ptr<onnxruntime::Model> model;
-    Status status = bfloat16_fix::Transform(subgraph, logger, model);
-    auto model_proto = model->ToProto();
-    model_proto->set_ir_version(ONNX_NAMESPACE::Version::IR_VERSION);
-    print_model_proto_duration();
-    DumpOpenVINOEPModel(onnx_model_path_name, model_proto.get(), fused_node);
-    ORT_ENFORCE(status.IsOK(), status.ErrorMessage());
-    return model_proto;
   } else {
     LOGS_DEFAULT(INFO) << "[OpenVINO-EP] OVEP QDQ optimization pass is disabled";
 

onnxruntime/core/providers/openvino/qdq_transformations/qdq_scales_fix.cpp

@@ -955,60 +955,5 @@ Status Transform(const GraphViewer& src_graph_viewer,
   return status;
 }
 }  // namespace qdq_scales_fix
-
-namespace bfloat16_fix {
-void replace_bf16_with_fp16(qdq_scales_fix::CustomGraph& gen_graph) {
-  for (auto& const_node : gen_graph.original_graph.Nodes()) {
-    auto node = const_cast<ONNX_NAMESPACE::Node*>(const_node);
-    if (node->OpType() == "Cast") {
-      for (auto& [name, const_attribute] : node->GetAttributes()) {
-        auto& attribute = const_cast<ONNX_NAMESPACE::AttributeProto&>(const_attribute);
-        if (name == "to" && attribute.type() == ONNX_NAMESPACE::AttributeProto_AttributeType_INT)
-          if (attribute.i() == ONNX_NAMESPACE::TensorProto_DataType_BFLOAT16)
-            attribute.set_i(ONNX_NAMESPACE::TensorProto_DataType_FLOAT16);
-      }
-    }
-    for (auto& output : node->OutputDefs()) {
-      auto& output_proto = const_cast<ONNX_NAMESPACE::TypeProto&>(output->ToProto().type());
-      if (output_proto.mutable_tensor_type()->elem_type() == ONNX_NAMESPACE::TensorProto_DataType_BFLOAT16)
-        output_proto.mutable_tensor_type()->set_elem_type(ONNX_NAMESPACE::TensorProto_DataType_FLOAT16);
-    }
-  }
-
-  for (auto& node : gen_graph.original_graph.Nodes()) {
-    for (auto& input_def : node->InputDefs()) {
-      ORT_THROW_IF_ERROR(graph_utils::ConvertInMemoryDataToInline(gen_graph.original_graph, input_def->Name()));
-    }
-  }
-
-  const auto& init_set = gen_graph.original_graph.GetAllInitializedTensors();
-  for (auto& [key, const_tensor_proto] : init_set) {
-    auto tensor_proto = const_cast<ONNX_NAMESPACE::TensorProto*>(const_tensor_proto);
-    auto dt = tensor_proto->data_type();
-    if (dt == ONNX_NAMESPACE::TensorProto_DataType_BFLOAT16) {
-      auto raw_data = tensor_proto->has_raw_data() ? reinterpret_cast<std::uint16_t*>(tensor_proto->mutable_raw_data()->data()) : nullptr;
-      if (raw_data) {
-        tensor_proto->set_data_type(ONNX_NAMESPACE::TensorProto_DataType_FLOAT16);
-        std::int64_t size = 1;
-        for (int i = 0; i < tensor_proto->dims_size(); ++i)
-          size *= tensor_proto->dims()[i];
-        for (std::int64_t i = 0; i < size; ++i) {
-          raw_data[i] = onnxruntime::MLFloat16(onnxruntime::BFloat16::FromBits(raw_data[i])).val;
-        }
-      }
-    }
-  }
-}
-
-Status Transform(const GraphViewer& src_graph_viewer,
-                 const logging::Logger& logger,
-                 /*out*/ std::unique_ptr<onnxruntime::Model>& model) {
-  auto status = qdq_scales_fix::copy_model(src_graph_viewer, logger, model);
-  auto g = qdq_scales_fix::generate_graph_from_onnx(model->MainGraph());
-
-  replace_bf16_with_fp16(g);
-  return status;
-}
-}  // namespace bfloat16_fix
 }  // namespace openvino_ep
 }  // namespace onnxruntime

onnxruntime/core/providers/openvino/qdq_transformations/qdq_scales_fix.h

@@ -15,10 +15,5 @@ Status Transform(const GraphViewer& src_graph,
                  const logging::Logger& logger,
                  /*out*/ std::unique_ptr<onnxruntime::Model>& model);
 }
-namespace bfloat16_fix {
-Status Transform(const GraphViewer& src_graph,
-                 const logging::Logger& logger,
-                 /*out*/ std::unique_ptr<onnxruntime::Model>& model);
-}
 }  // namespace openvino_ep
 }  // namespace onnxruntime