Best effort support for all Hugging Face transformers models

[dwyatte]

With https://github.com/huggingface/text-generation-inference adopting a less friendly license, this seems like a good opportunity to add best effort support for all Hugging Face transformers models that generate text e.g., via AutoModelForCausalLM and AutoModelForSeq2SeqLM. This would allow them to take advantage of vLLM's other serving features while specific models can retain optimized implementations or gain them as they are implemented

• https://github.com/huggingface/text-generation-inference/blob/ecf6dc3a5a31c1b0e1ed48988ddf2416b5e35660/server/text_generation_server/models/causal_lm.py#L451
• https://github.com/huggingface/text-generation-inference/blob/ecf6dc3a5a31c1b0e1ed48988ddf2416b5e35660/server/text_generation_server/models/seq2seq_lm.py#L501

[liu-godzilla]

any updates here?

[mgoin]

Considering there are 35 distinct model architectures defined as of now, I think we are achieving "best effort". Not to mention several models are implemented by aliasing within the llama architecture. Please open an issue for any specific models you would like to see supported!