Contrastive Search

[KexinFeng]

Description

This PR succeeds PR #2509. The model tracing is shown therein.

It implements contrastive search algo based on torchscript gpt2 model. The onnx model support waits for the issue huggingface/optimum#972 to be solved.

Benchmarked with huggingface transformers' output.

Ref.
https://huggingface.co/blog/introducing-csearch

Demo output

In the demo TestLMSearch.java, we feed in batch sequence input, using right padding with the space token ' ' (id = 220).

["DeepMind Company is",  
 "Memories follow me left and right. I can"]

Output (topk = 3, maxLength = 50):

'DeepMind Company is      \xa0\na small startup that aims to build a better understanding 
of neural networks and how they work. We are currently developing a simple way to create
 a new type of machine learning model that can be used'

"Memories follow me left and right. I can't remember what happened last night, but I know 
that it was very sad. I don't know how long I've been here, but I'm sure there's something
 wrong with me. I'm"

The output successfully avoids the repetitive token output, as expected in Ref. https://huggingface.co/blog/introducing-csearch.

Model tracing

The onnx model gpt2.onnx is loaded from https://huggingface.co/docs/optimum/main/en/exporters/onnx/usage_guides/export_a_model#exporting-a-model-using-past-keysvalues-in-the-decoder.
See also https://github.com/huggingface/optimum/releases.

The gpt2.pt is traced with the following scripts: https://gist.github.com/KexinFeng/4876c6bfb27f40abffe4d5a92c02acff

[KexinFeng]

Merged in #2637