vllm 本地大模型加速推理

背景

使用 vllm 可以加速模型的推理过程；vllm 加速的代码也很少，只需要调用包即可实现，没有太大的学习成本，而且好处很多，可以极大提升模型的推理速度；

简介

使用 modelscope 的 chatglm3-6B，调用 vllm 加速推理；

我的显卡显存为 24G ；
chatglm3-6B，如果不用vllm，我的显存不够，必须使用 half 才能放进显存 ；
使用 vllm 后，vllm 加载的大模型模型权重占用空间会小一点；不使用 half，恰好能放进我的显存空间；

实操

By default, vLLM downloads model from HuggingFace. If you would like to use models from ModelScope in the following examples, please set the environment variable:

export VLLM_USE_MODELSCOPE=True

from vllm import LLM, SamplingParams

sampling_params = SamplingParams(temperature=0.8, top_p=0.95, max_tokens=2048)

SamplingParams?

Init signature:
SamplingParams(
    n: int = 1,
    best_of: Optional[int] = None,
    presence_penalty: float = 0.0,
    frequency_penalty: float = 0.0,
    repetition_penalty: float = 1.0,
    temperature: float = 1.0,
    top_p: float = 1.0,
    top_k: int = -1,
    min_p: float = 0.0,
    use_beam_search: bool = False,
    length_penalty: float = 1.0,
    early_stopping: Union[bool, str] = False,
    stop: Union[List[str], str, NoneType] = None,
    stop_token_ids: Optional[List[int]] = None,
    include_stop_str_in_output: bool = False,
    ignore_eos: bool = False,
    max_tokens: Optional[int] = 16,
    logprobs: Optional[int] = None,
    prompt_logprobs: Optional[int] = None,
    skip_special_tokens: bool = True,
    spaces_between_special_tokens: bool = True,
    logits_processors: Optional[List[Callable[[List[int], torch.Tensor], torch.Tensor]]] = None,
) -> None
Docstring:     
Sampling parameters for text generation.  

利用SamplingParams的参数，控制大模型的输出结果；
使用大模型针对如下 prompt 做推理：

prompts = [
    "Hello, my name is",
    "The president of the United States is",
    "The capital of France is",
    "The future of AI is",
    "今天天气真好，咱们出去",
    "明天就要开学了，我的作业还没写完，",
    "请你介绍一下你自己。AI："
]
sampling_params = SamplingParams(temperature=0.8, top_p=0.95, max_tokens=2048)

model_path = 'ZhipuAI/chatglm3-6b'
llm = LLM(
    model=model_path,
    trust_remote_code=True,
    tokenizer=model_path,
    tokenizer_mode='slow',
    tensor_parallel_size=1
)

outputs = llm.generate(prompts, sampling_params)

推理耗时7秒，速度快了很多；

查看推理结果

# Print the outputs.
for output in outputs:
    prompt = output.prompt
    generated_text = output.outputs[0].text
    print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")

代码公开

本文代码发布在 github：https://github.com/JieShenAI/csdn/blob/main/24/04/vllm/vllm_demo.ipynb