[BFCL] Evaluation with Correct Precision Settings for Locally-Hosted Models

[HuanzhiMao]

The following models are intended to be evaluated using bfloat16 precision instead of float16 according to their model card on HuggingFace. We should change the default precision setting for their model handlers. This means they cannot be evaluated using v100 GPUs.

• deepseek-ai/deepseek-coder-6.7b-instruct
• google/gemma-7b-it
• meetkai/functionary-small-v2.2-FC
• meetkai/functionary-medium-v2.2-FC
• meetkai/functionary-small-v2.4-FC
• meetkai/functionary-medium-v2.4-FC
• NousResearch/Hermes-2-Pro-Llama-3-70B
• NousResearch/Hermes-2-Pro-Mistral-7B
• NousResearch/Hermes-2-Theta-Llama-3-8B
• NousResearch/Hermes-2-Theta-Llama-3-70B
• meta-llama/Meta-Llama-3-8B-Instruct
• meta-llama/Meta-Llama-3-70B-Instruct
• ibm-granite/granite-20b-functioncalling
• THUDM/glm-4-9b-chat

[HuanzhiMao]

Starting with #572: We will have all the locally-hosted models' responses generated with bloat16 using A6000 GPU.