fix two bugs related to the bert model

src/onnx/onnx.cpp

@@ -490,6 +490,16 @@ struct onnx_parser
         {
             axis = parse_value(attributes.at("axis")).at<int>();
         }
+
+        if(!args[0]->get_shape().standard())
+        {
+            args[0] = prog.add_instruction(op::contiguous{}, args[0]);
+        }
+        if(!args[1]->get_shape().standard())
+        {
+            args[1] = prog.add_instruction(op::contiguous{}, args[1]);
+        }
+
         op::gather op{axis};
         return prog.add_instruction(op, std::move(args));
     }

src/quantization.cpp

@@ -53,7 +53,7 @@ void quantize(program& prog, const std::vector<std::string>& ins_names)
                 // if the input is a convert operator, uses its input
                 // as its current input
                 instruction_ref input_fp16{};
-                if(input->name() == "convert")
+                if(input->name() == "convert" and input->inputs().front()->get_shape().type() == shape::half_type)
                 {
                     input_fp16 = input->inputs().front();
                 }