OpenCompass/configs/models/qwen/vllm_qwen1_5_72b_chat.py

from opencompass.models import VLLM


_meta_template = dict(
    round=[
        dict(role='HUMAN', begin='<|im_start|>user\n', end='<|im_end|>\n'),
        dict(role='BOT', begin='<|im_start|>assistant\n', end='<|im_end|>\n', generate=True),
    ],
)

models = [
    dict(
        type=VLLM,
        abbr='qwen1.5-72b-chat-vllm',
        path='Qwen/Qwen1.5-72B-Chat',
        model_kwargs=dict(tensor_parallel_size=4),
        meta_template=_meta_template,
        max_out_len=100,
        max_seq_len=2048,
        batch_size=32,
        generation_kwargs=dict(temperature=0),
        stop_words=['<|im_end|>'],
        run_cfg=dict(num_gpus=4, num_procs=1),
    )
]