Can not convert 'tf.AddV2' with dynamic shape from TF dialect to mhlo dialect in MLIR

[mofheka]

System information

• Have I written custom code (as opposed to using a stock example script provided in TensorFlow): No
• OS Platform and Distribution (e.g., Linux Ubuntu 16.04): Linux Ubuntu 18.04
• TensorFlow installed from (source or binary): source
• TensorFlow version (use command below): libtensorflow_framework.so.2.5.0
• Python version: Python 3.7.0
• Bazel version (if compiling from source): bazel-3.7.2-linux-x86_64
• GCC/Compiler version (if compiling from source): 10.1.0
• CUDA/cuDNN version: 11.0
• GPU model and memory: TITAN V 12066MiB

Describe the current behavior

command:

$MODEL_DIR/$model_mhlo_dialect_name $build_dir/tf-opt --tf-to-hlo-pipeline $MODEL_DIR/$model_tf_dialect_name -o $MODEL_DIR/$model_mhlo_dialect_name

then throw an error：

error: 'shape.cstr_broadcastable' op operand #1 must be shape or extent tensor, but got 'tensor<1xi32>' %239 = "tf.AddV2"(%223, %238) {device = ""} : (tensor<?xi32>, tensor<?xi32>) -> tensor<?xi32>