Currently pad token by default in [HuggingFace Tokenizer is `None`](https://github.com/huggingface/transformers/blob/main/src/transformers/models/llama/tokenization_llama.py#L110). We add the padding token as a special token to the tokenizer, which in this case requires to resize the token_embeddings as shown below:
Padding would be required for batch inference. In this this [example](../inference/inference.py), batch size = 1 so essentially padding is not required. However,We added the code pointer as an example in case of batch inference.
**Chat completion**
The inference folder also includes a chat completion example, that adds built-in safety features in fine-tuned models to the prompt tokens. To run the example: