Tensorflow Model permanently becomes corrupt when input embeddings exceed max_position_embeddings

I use Tensorflow C++ API.

I have a Tensorflow model.

I give some inputs to this model.

There is a parameter called max_position_embeddings

This parameter determines maximum acceptable input dimensions

When I give a very long input for inference I get the exception:

{{function_node __inference__inference_10663}} {{function_node __inference__inference_10663}} indices[0,2048] = 2049 is not in [0, 2049)

[[{{node decoder/position_embeddings/Gather}}]] [[StatefulPartitionedCall/StatefulPartitionedCall]]

So far, everything is normal. However, after catching this exception, I try a second inference with a small size input and again I get the very same error:

{{function_node __inference__inference_10663}} {{function_node __inference__inference_10663}} indices[0,2048] = 2049 is not in [0, 2049)

I have debugged the code and in the 2nd inference attempt with good input, just before inference and during initializing tensors to feed to inference, I got the exception during tensor initialization:

Model Mdl = ... // this is the model gets corrupted in the previous inference call

// Define the tensors

Tensor input_ids{ Mdl,serving_default_input_ids }; //this is where I get exception.

What I wonder is that why Tensorflow graphs and models are stateful and once I get an exception for max_position_embeddings , I can no longer do an inference with this model.

Is there a solution to fix this issue?

Thanks.