[NeuralChat] fix wrong output of multi-round prediction

[Spycsh]

Type of Change

bug fix

Description

A simple example will reproduce this issue, when the second output contains template information and no correct output. I think we should by default clear the previous prompt template unless we know exactly our model can work on multi-round speech.

from intel_extension_for_transformers.neural_chat import build_chatbot, PipelineConfig
config = PipelineConfig()
chatbot = build_chatbot(config)
out = chatbot.predict("Tell me about Intel Xeon Scalable Processors.")
print(out)
out = chatbot.predict("Tell me about Intel")
print(out)

This is because the new query is simply appended to the previous template and the default model cannot deal with it.

Expected Behavior & Potential Risk

bug fix

How has this PR been tested?

example

Dependency Change?

None