Add StableLM3B model

[ita9naiwa]

Resolves #1426

Add new model StableLM

[WoosukKwon]

Hi @ita9naiwa, thanks for submitting the PR! It looks good overall, but needs some minor stylistic changes. Please check out my comments.

[WoosukKwon]

Do we need this class?

[ita9naiwa]

No, I can remove this class and pass arguments directly.

[ita9naiwa]

models like llama in vllm (models/llama.py) there are config classes.

# (llama.py)
from transformers import LlamaConfig
# (yi.py)
from vllm.transformers_utils.configs.yi import YiConfig

StableLM does not have official transformers implementations so I guess it's better to have StableLMEpochConfig to carry configs.

[ita9naiwa]

because passing all arguments for the model will make code really verbose

I'm just asking and I'll follow your opinion

[simon-mo]

I think the discussion point here is that StableLMEpochConfig is never initialized. When vLLM is reading the config, it is read into PretrainedConfig, which you can also access the model specific attributes using config.vocab_size. This class currently only serve as documentation purpose.

[ita9naiwa]

@simon-mo Thank you notifying that. I'll remove StableLMEpochConfig class.

[ita9naiwa]

I removed StableLMEpochConfig class.

[ita9naiwa]

I checked using models/test_models.py the model is consistent with huggingface implementation

[simon-mo]

confirming examples

curl http://localhost:8000/v1/completions \
    -H "Content-Type: application/json" \
    -d '{
        "model": "stabilityai/stablelm-3b-4e1t",
        "prompt": "San Francisco is a",
        "max_tokens": 7,
        "temperature": 0
    }'
{"id":"cmpl-a65e9c6d27d6479eb299c60dd4d3baa0","object":"text_completion","created":66942,"model":"stabilityai/stablelm-3b-4e1t","choices":[{"index":0,"text":" city that is known for its beautiful","logprobs":null,"finish_reason":"length"}],"usage":{"prompt_tokens":4,"total_tokens":11,"completion_tokens":7}}