fix layernormTrt meanVar alloc bug (#45255)

* fix layernormTrt meanVar alloc bug

paddle/fluid/inference/tensorrt/convert/layer_norm_op.cc

@@ -56,12 +56,14 @@ class LayerNormOpConverter : public OpConverter {
 
     nvinfer1::ILayer* layernorm_layer = nullptr;
     if (engine_->with_dynamic_shape()) {
-      int input_num = 1;
-      for (int i = begin_norm_axis; i < X->getDimensions().nbDims; i++) {
-        input_num *= X->getDimensions().d[i];
+      int statis_num = 1;
+      // For dynamic shape,
+      // the batch num will be taken into account in plugin runtime.
+      for (int i = 1; i < begin_norm_axis; i++) {
+        statis_num *= X->getDimensions().d[i];
       }
-      std::vector<int64_t> mean_shape{input_num};
-      std::vector<int64_t> variance_shape{input_num};
+      std::vector<int64_t> mean_shape{statis_num};
+      std::vector<int64_t> variance_shape{statis_num};
       plugin::LayerNormPluginDynamic* plugin =
           new plugin::LayerNormPluginDynamic(
               static_cast<const float*>(bias_weight.get().values),
@@ -74,12 +76,12 @@ class LayerNormOpConverter : public OpConverter {
               variance_shape);
       layernorm_layer = engine_->AddDynamicPlugin(&X, 1, plugin);
     } else {
-      int input_num = 1;
-      for (int i = begin_norm_axis - 1; i < X->getDimensions().nbDims; i++) {
-        input_num *= X->getDimensions().d[i];
+      int statis_num = 1;
+      for (int i = 0; i < begin_norm_axis; i++) {
+        statis_num *= X->getDimensions().d[i];
       }
-      std::vector<int64_t> mean_shape{input_num};
-      std::vector<int64_t> variance_shape{input_num};
+      std::vector<int64_t> mean_shape{statis_num};
+      std::vector<int64_t> variance_shape{statis_num};
       plugin::LayerNormPlugin* plugin = new plugin::LayerNormPlugin(
           static_cast<const float*>(bias_weight.get().values),
           bias_weight.get().count,

paddle/fluid/inference/tensorrt/plugin/layer_norm_op_plugin.cu

@@ -175,6 +175,11 @@ int LayerNormPluginDynamic::enqueue(
   for (int i = 0; i < input_dims.nbDims; i++) {
     input_shape.push_back(input_dims.d[i]);
   }
+  // in dynamic shape
+  // the batch num should be involved in mean/variance shape
+  mean_shape_[0] *= input_dims.d[0];
+  variance_shape_[0] *= input_dims.d[0];
+
   const auto input_ddim = phi::make_ddim(input_shape);
   auto matrix_dim = phi::flatten_to_2d(input_ddim, begin_norm_axis);
   int feature_size = static_cast<int>(matrix_dim[1]);

python/paddle/fluid/tests/unittests/ir/inference/test_trt_convert_layer_norm.py

@@ -140,5 +140,123 @@ class TrtConvertLayerNormTest(TrtLayerAutoScanTest):
         self.run_test()
 
 
+class TrtConvertLayerNormTest_2(TrtLayerAutoScanTest):
+
+    def is_program_valid(self, program_config: ProgramConfig) -> bool:
+        inputs = program_config.inputs
+        weights = program_config.weights
+        attrs = [
+            program_config.ops[i].attrs for i in range(len(program_config.ops))
+        ]
+
+        if attrs[0]['epsilon'] < 0 or attrs[0]['epsilon'] > 0.001:
+            return False
+        if attrs[0]['begin_norm_axis'] <= 0 or attrs[0]['begin_norm_axis'] >= (
+                len(inputs['input_data'].shape) - 1):
+            return False
+
+        return True
+
+    def sample_program_configs(self):
+
+        def generate_input1(attrs: List[Dict[str, Any]], shape_input):
+            return np.ones(shape_input).astype(np.float32)
+
+        def generate_input2(attrs: List[Dict[str, Any]], shape_input):
+            begin = attrs[0]["begin_norm_axis"]
+            sum = 1
+            for x in range(begin, len(shape_input)):
+                sum *= shape_input[x]
+            return np.ones([sum]).astype(np.float32)
+
+        for epsilon in [0.0005, -1, 1]:
+            for begin_norm_axis in [1, 0, -1, 2, 3]:
+                dics = [{
+                    "epsilon": epsilon,
+                    "begin_norm_axis": begin_norm_axis
+                }, {}]
+
+                ops_config = [{
+                    "op_type": "layer_norm",
+                    "op_inputs": {
+                        "X": ["input_data"],
+                        "Scale": ["scale_data"],
+                        "Bias": ["bias_data"]
+                    },
+                    "op_outputs": {
+                        "Y": ["y_data"],
+                        "Mean": ["saved_mean_data"],
+                        "Variance": ["saved_variance_data"]
+                    },
+                    "op_attrs": dics[0]
+                }]
+                ops = self.generate_op_config(ops_config)
+                shape_input = [2, 64, 3, 3]
+                program_config = ProgramConfig(
+                    ops=ops,
+                    weights={
+                        "bias_data":
+                        TensorConfig(data_gen=partial(generate_input2, dics,
+                                                      shape_input)),
+                        "scale_data":
+                        TensorConfig(data_gen=partial(generate_input2, dics,
+                                                      shape_input))
+                    },
+                    inputs={
+                        "input_data":
+                        TensorConfig(data_gen=partial(generate_input1, dics,
+                                                      shape_input))
+                    },
+                    outputs=["y_data"])
+
+                yield program_config
+
+    def sample_predictor_configs(
+            self, program_config) -> (paddle_infer.Config, List[int], float):
+
+        def generate_dynamic_shape(attrs):
+            self.dynamic_shape.min_input_shape = {"input_data": [1, 64, 3, 3]}
+            self.dynamic_shape.max_input_shape = {"input_data": [4, 64, 3, 3]}
+            self.dynamic_shape.opt_input_shape = {"input_data": [2, 64, 3, 3]}
+
+        def clear_dynamic_shape():
+            self.dynamic_shape.min_input_shape = {}
+            self.dynamic_shape.max_input_shape = {}
+            self.dynamic_shape.opt_input_shape = {}
+
+        def generate_trt_nodes_num(attrs, dynamic_shape):
+            inputs = program_config.inputs
+            #if not dynamic_shape:
+            #    if attrs[0]["begin_norm_axis"] >= len(inputs["input_data"].shape) - 1:
+            #        print ("iiiiiii")
+            #        return 0, 3
+            return 1, 2
+
+        attrs = [
+            program_config.ops[i].attrs for i in range(len(program_config.ops))
+        ]
+
+        # for static_shape
+        clear_dynamic_shape()
+        self.trt_param.precision = paddle_infer.PrecisionType.Float32
+        yield self.create_inference_config(), generate_trt_nodes_num(
+            attrs, False), 1e-5
+        self.trt_param.precision = paddle_infer.PrecisionType.Half
+        yield self.create_inference_config(), generate_trt_nodes_num(
+            attrs, False), 1e-2
+
+        # for dynamic_shape
+        generate_dynamic_shape(attrs)
+        self.trt_param.precision = paddle_infer.PrecisionType.Float32
+        yield self.create_inference_config(), generate_trt_nodes_num(
+            attrs, True), 1e-5
+        self.trt_param.precision = paddle_infer.PrecisionType.Half
+        yield self.create_inference_config(), generate_trt_nodes_num(
+            attrs, True), 1e-2
+
+    def test(self):
+        self.run_test()
+
+
 if __name__ == "__main__":
     unittest.main()