[Paddle-TRT]fix bug in fill_constant_batch_size_like op (#46334)

* fix beta bug in fill_constant_batch_size_like

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

@@ -44,6 +44,10 @@ class FillConstantBatchSizeLikeOpConverter : public OpConverter {
         PADDLE_GET_CONST(std::string, op_desc.GetAttr("str_value"));
     std::vector<int32_t> shape =
         PADDLE_GET_CONST(std::vector<int32_t>, op_desc.GetAttr("shape"));
+    if (str_value == "") {
+      float value = PADDLE_GET_CONST(float, op_desc.GetAttr("value"));
+      str_value = std::to_string(value);
+    }
     float value = std::stof(str_value);
 
     auto* input_shape_tensor = Shape(input);
@@ -65,7 +69,7 @@ class FillConstantBatchSizeLikeOpConverter : public OpConverter {
     auto layer = TRT_ENGINE_ADD_LAYER(
         engine_, Fill, nvinfer1::Dims{}, nvinfer1::FillOperation::kLINSPACE);
     std::vector<float> value_vec(1, value);
-    std::vector<float> beta_vec(3, 0.);
+    std::vector<float> beta_vec(shape.size(), 0.);
     layer->setAlpha(value);
     layer->setBeta(0.f);
     layer->setInput(0, *out_shape_tensor);

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

@@ -87,7 +87,9 @@ class TrtConvertSliceTest(TrtLayerAutoScanTest):
                                     "input_dim_idx":
                                     0,
                                     "str_value":
-                                    "0.0",
+                                    "",
+                                    "value":
+                                    0.0,
                                     "shape": [K * num_layers, -1, hidden_size],
                                     "output_dim_idx":
                                     1,