diff --git a/internal/maa/framework.go b/internal/maa/framework.go
index 4b8b4e8..9c52f2b 100644
--- a/internal/maa/framework.go
+++ b/internal/maa/framework.go
@@ -39,6 +39,42 @@ type MaaCustomRecognitionCallback func(context uintptr, taskId int64, currentTas
 
 type MaaCustomActionCallback func(context uintptr, taskId int64, currentTaskName, customActionName, customActionParam *byte, recoId int64, box uintptr, transArg uintptr) uint64
 
+type MaaInferenceDevice int32
+
+const (
+	MaaInferenceDevice_CPU  MaaInferenceDevice = -2
+	MaaInferenceDevice_Auto MaaInferenceDevice = -1
+	MaaInferenceDevice_0    MaaInferenceDevice = 0
+	MaaInferenceDevice_1    MaaInferenceDevice = 1
+	// and more gpu id or flag...
+)
+
+type MaaInferenceExecutionProvider int32
+
+const (
+
+	// I don't recommend setting up MaaResOption_InferenceDevice in this case,
+	// because you don't know which EP will be used on different user devices.
+	MaaInferenceExecutionProvider_Auto = 0
+
+	// MaaResOption_InferenceDevice will not work.
+	MaaInferenceExecutionProvider_CPU = 1
+
+	// MaaResOption_InferenceDevice will be used to set adapter id,
+	// It's from Win32 API `EnumAdapters1`.
+	MaaInferenceExecutionProvider_DirectML = 2
+
+	// MaaResOption_InferenceDevice will be used to set coreml_flag,
+	// Reference to
+	// https://github.com/microsoft/onnxruntime/blob/main/include/onnxruntime/core/providers/coreml/coreml_provider_factory.h
+	// But you need to pay attention to the onnxruntime version we use, the latest flag may not be supported.
+	MaaInferenceExecutionProvider_CoreML = 3
+
+	// MaaResOption_InferenceDevice will be used to set NVIDIA GPU ID
+	// TODO!
+	MaaInferenceExecutionProvider_CUDA = 4
+)
+
 type MaaResOption int32
 
 const (
@@ -49,7 +85,14 @@ const (
 	///
 	/// value: MaaInferenceDevice, eg: 0; val_size: sizeof(MaaInferenceDevice)
 	/// default value is MaaInferenceDevice_Auto
-	MaaResOption_InterfaceDevice MaaResOption = 1
+	MaaResOption_InferenceDevice MaaResOption = 1
+
+	/// Use the specified inference execution provider
+	/// Please set this option before loading the model.
+	///
+	/// value: MaaInferenceExecutionProvider, eg: 0; val_size: sizeof(MaaInferenceExecutionProvider)
+	/// default value is MaaInferenceExecutionProvider_Auto
+	MaaResOption_InferenceExecutionProvider MaaResOption = 2
 )
 
 var (
diff --git a/resource.go b/resource.go
index 66c3ac2..1e49b02 100644
--- a/resource.go
+++ b/resource.go
@@ -70,25 +70,55 @@ func (r *Resource) setOption(key maa.MaaResOption, value unsafe.Pointer, valSize
 	)
 }
 
-type InterfaceDevice int32
-
-// InterfaceDevice
-const (
-	InterfaceDeviceCPU  InterfaceDevice = -2
-	InterfaceDeviceAuto InterfaceDevice = -1
-	InterfaceDeviceGPU0 InterfaceDevice = 0
-	InterfaceDeviceGPU1 InterfaceDevice = 1
-	// and more gpu id...
-)
-
-func (r *Resource) SetInterfaceDevice(device InterfaceDevice) bool {
+func (r *Resource) setInferenceDevice(device maa.MaaInferenceDevice) bool {
 	return r.setOption(
-		maa.MaaResOption_InterfaceDevice,
+		maa.MaaResOption_InferenceDevice,
 		unsafe.Pointer(&device),
 		unsafe.Sizeof(device),
 	)
 }
 
+func (r *Resource) setInferenceExecutionProvider(ep maa.MaaInferenceExecutionProvider) bool {
+	return r.setOption(
+		maa.MaaResOption_InferenceExecutionProvider,
+		unsafe.Pointer(&ep),
+		unsafe.Sizeof(ep),
+	)
+}
+
+func (r *Resource) setInference(ep maa.MaaInferenceExecutionProvider, deviceID maa.MaaInferenceDevice) bool {
+	return r.setInferenceExecutionProvider(ep) && r.setInferenceDevice(deviceID)
+}
+
+// UseCPU
+func (r *Resource) UseCPU() bool {
+	return r.setInference(maa.MaaInferenceExecutionProvider_CPU, maa.MaaInferenceDevice_CPU)
+}
+
+type InterenceDevice = maa.MaaInferenceDevice
+
+const (
+	InterenceDeviceAuto int32 = -1
+	InferenceDevice0    int32 = 0
+	InferenceDevice1    int32 = 1
+	// and more gpu id or flag...
+)
+
+// UseDirectml
+func (r *Resource) UseDirectml(deviceID InterenceDevice) bool {
+	return r.setInference(maa.MaaInferenceExecutionProvider_DirectML, deviceID)
+}
+
+// UseCoreml
+func (r *Resource) UseCoreml(coremlFlag InterenceDevice) bool {
+	return r.setInference(maa.MaaInferenceExecutionProvider_CoreML, coremlFlag)
+}
+
+// UseAutoExecutionProvider
+func (r *Resource) UseAutoExecutionProvider() bool {
+	return r.setInference(maa.MaaInferenceExecutionProvider_Auto, maa.MaaInferenceDevice_Auto)
+}
+
 // RegisterCustomRecognition registers a custom recognition to the resource.
 func (r *Resource) RegisterCustomRecognition(name string, recognition CustomRecognition) bool {
 	id := registerCustomRecognition(recognition)