[BFCL] Add empower functions models and the supporting handler

[dyang415]

Thanks for the great benchmark @HuanzhiMao!

Create this PR to add the empower functions v1.1 model series models:

• Register empower functions model series
• Add empower_handler.py in support of running the newly added models
• Update vllm to 0.5.5 in support of Llama 3.1 based models

Evaluation result:

Rank	Overall Acc	Model	AST Summary	Exec Summary	Irrelevance Detection	Relevance Detection
1	86.92%	Empower-Fucntions-Large-v1.1 (FC)	82.55%	92.57%	78.51%	90.24%
2	83.89%	Empower-Fucntions-Small-v1.1 (FC)	77.45%	89.43%	82.69%	85.37%

[dyang415]

Hi @HuanzhiMao can you please take a look when you get a chance? Appreciate it!

[HuanzhiMao]

Apologize for the long delay. Will definitely take a look after the Monday ICLR deadline.

[dyang415]

Thanks @HuanzhiMao , rebased and updated to support BFCL v3.

[HuanzhiMao]

Hi @dyang415,

Thanks for the PR.
Your model's chat template on huggingface (link) is the exact same as the one that meta-llama/Meta-Llama-3-70B-Instruct has (link). Therefore, the empower series model should just use the LlamaHandler (code link); there is no need for a separate handler.

[dyang415]

Hi @HuanzhiMao , though the chat template is the same, the way encoding function/response is very different compared with the llama function handler, so we still need the place to put the logic.

[HuanzhiMao]

empower functions v1.1 model series

Where can I find the mention of the encoding stuff in your huggingface or other documentation?

[dyang415]

It's in the handler: https://github.com/ShishirPatil/gorilla/pull/630/files#diff-ace43050c359c63d5751e03cbac92a1657b72e872e779913c471acba4dfbf3e6R21-R100, also other use cases we provided a repo which includes the logic https://github.com/empower-ai/empower-functions