Bi-gram Repetation Penalty for the TGI configuration

[mertege]

Feature request

Add a new configuration parameter, "bigram_repetition_penalty", to the Text Generation Inference module. This parameter will introduce a mechanism that penalizes repeated bigrams in generated text, similar to how "repetition_penalty" operates for individual tokens, but extend this logic to two-token sequences (bigrams).
Integrate the "bigram_repetition_penalty" as an optional parameter within the text generation configuration. This would function by reducing the probability of selecting any token that, combined with the last generated token, forms a bigram already appearing in the generated sequence.

Motivation

Repetition in the generated text is a common issue in large language models, where the output may become redundant or repetitive, diminishing the quality of responses. While the existing "repetition_penalty" addresses this concern for individual tokens, it can decrease the usage of "stop words" and affect the response's nature. However, adding a "bigram_repetition_penalty" would provide finer control over text generation by preventing the model from generating the exact two-token phrases repeatedly.

Your contribution

I am open to collaborating with the community to implement this feature.