[Bug]: Bert tokenizer is tokenizing some tokens as `UNK`

[maxdebayser]

Your current environment

The output of `python collect_env.py`

Your output of `python collect_env.py` here

Model Input Dumps

No response

🐛 Describe the bug

With some Bert and Roberta models like sentence-transformers/all-MiniLM-L12-v2 I found that the output is not similar to the one generated by sentence-transformers. If I place the following prints in _normalize_prompt_text_to_input() in serving_engine.py

        print(f"{input_ids=}")

I get [101, 100, 3007, 1997, 100, 2003, 100, 1012, 102] for the sentence "The capital of France is Paris.". 100 is the UNK token. When I run with sentence-transformers, I get [ 101, 1996, 3007, 1997, 2605, 2003, 3000, 1012, 102] . This problem happens both with --tokenizer-mode auto and --tokenizer-mode slow.

cc: @DarkLight1337

Before submitting a new issue...

• Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.

[maxdebayser]

In the model's sentence_bert_config.json do_lower_case is false

{
  "max_seq_length": 128,
  "do_lower_case": false
}

but in the tokenizer_config.json it's true:

{"do_lower_case": true, ....

When we added support for sentence_bert_config.json we assumed this configuration was meant to override the tokenizer configuration, but that's probably wrong:

vllm/vllm/transformers_utils/tokenizer_group/__init__.py

Line 29 in 0d8451c

if (model_config.encoder_config is not None

Disabling this argument make the tokenizer convert the text to lower case correctly. I think we need to do the same thing as the sentence-transformer library and call lower() before passing the input to the tokenizer: https://github.com/UKPLab/sentence-transformers/blob/cfb883ce34ad301aa9444bca7462734565456e81/sentence_transformers/models/Transformer.py#L500

cc: @flaviabeo