| from opencompass.models import HuggingFaceCausalLM | |
| _meta_template = dict( | |
| round=[ | |
| dict(role='HUMAN', begin='### Human: \n', end='\n\n'), | |
| dict(role='BOT', begin='### Assistant: ', end='</s>', generate=True), | |
| ], | |
| ) | |
| models = [ | |
| dict( | |
| type=HuggingFaceCausalLM, | |
| abbr='nanbeige-16b-chat-hf', | |
| path="Nanbeige/Nanbeige-16B-Chat", | |
| tokenizer_path='Nanbeige/Nanbeige-16B-Chat', | |
| model_kwargs=dict( | |
| device_map='auto', | |
| trust_remote_code=True, | |
| torch_dtype='auto', | |
| ), | |
| tokenizer_kwargs=dict( | |
| padding_side='left', | |
| truncation_side='left', | |
| trust_remote_code=True, | |
| use_fast=False, | |
| ), | |
| meta_template=_meta_template, | |
| batch_padding=False, | |
| max_out_len=100, | |
| max_seq_len=4096, | |
| batch_size=8, | |
| run_cfg=dict(num_gpus=2, num_procs=1), | |
| end_str='</s>', | |
| ) | |
| ] | |