vLLM
vLLM 是一个快速且易于使用的 LLM 推理和服务库,提供
- 最先进的服务吞吐量
- 使用 PagedAttention 有效管理注意力键和值内存
- 持续批量处理传入请求
- 优化的 CUDA 内核
本笔记本介绍了如何将 LLM 与 langchain 和 vLLM 一起使用。
要使用,您应该安装 vllm
python 包。
%pip install --upgrade --quiet vllm -q
from langchain_community.llms import VLLM
llm = VLLM(
model="mosaicml/mpt-7b",
trust_remote_code=True, # mandatory for hf models
max_new_tokens=128,
top_k=10,
top_p=0.95,
temperature=0.8,
)
print(llm.invoke("What is the capital of France ?"))
API 参考:VLLM
INFO 08-06 11:37:33 llm_engine.py:70] Initializing an LLM engine with config: model='mosaicml/mpt-7b', tokenizer='mosaicml/mpt-7b', tokenizer_mode=auto, trust_remote_code=True, dtype=torch.bfloat16, use_dummy_weights=False, download_dir=None, use_np_weights=False, tensor_parallel_size=1, seed=0)
INFO 08-06 11:37:41 llm_engine.py:196] # GPU blocks: 861, # CPU blocks: 512
``````output
Processed prompts: 100%|██████████| 1/1 [00:00<00:00, 2.00it/s]
``````output
What is the capital of France ? The capital of France is Paris.
将模型集成到 LLMChain 中
from langchain.chains import LLMChain
from langchain_core.prompts import PromptTemplate
template = """Question: {question}
Answer: Let's think step by step."""
prompt = PromptTemplate.from_template(template)
llm_chain = LLMChain(prompt=prompt, llm=llm)
question = "Who was the US president in the year the first Pokemon game was released?"
print(llm_chain.invoke(question))
API 参考:LLMChain | PromptTemplate
Processed prompts: 100%|██████████| 1/1 [00:01<00:00, 1.34s/it]
``````output
1. The first Pokemon game was released in 1996.
2. The president was Bill Clinton.
3. Clinton was president from 1993 to 2001.
4. The answer is Clinton.
分布式推理
vLLM 支持分布式张量并行推理和服务。
要使用 LLM 类运行多 GPU 推理,请将 tensor_parallel_size
参数设置为您想要使用的 GPU 数量。例如,要在 4 个 GPU 上运行推理
from langchain_community.llms import VLLM
llm = VLLM(
model="mosaicml/mpt-30b",
tensor_parallel_size=4,
trust_remote_code=True, # mandatory for hf models
)
llm.invoke("What is the future of AI?")
API 参考:VLLM
量化
vLLM 支持 awq
量化。要启用它,请将 quantization
传递给 vllm_kwargs
。
llm_q = VLLM(
model="TheBloke/Llama-2-7b-Chat-AWQ",
trust_remote_code=True,
max_new_tokens=512,
vllm_kwargs={"quantization": "awq"},
)
OpenAI 兼容服务器
vLLM 可以部署为模仿 OpenAI API 协议的服务器。这使得 vLLM 可以用作使用 OpenAI API 的应用程序的直接替代品。
此服务器可以使用与 OpenAI API 相同的格式进行查询。
OpenAI 兼容完成
from langchain_community.llms import VLLMOpenAI
llm = VLLMOpenAI(
openai_api_key="EMPTY",
openai_api_base="https://#:8000/v1",
model_name="tiiuae/falcon-7b",
model_kwargs={"stop": ["."]},
)
print(llm.invoke("Rome is"))
API 参考:VLLMOpenAI
a city that is filled with history, ancient buildings, and art around every corner
LoRA 适配器
LoRA 适配器可以与任何实现 SupportsLoRA
的 vLLM 模型一起使用。
from langchain_community.llms import VLLM
from vllm.lora.request import LoRARequest
llm = VLLM(
model="meta-llama/Llama-3.2-3B-Instruct",
max_new_tokens=300,
top_k=1,
top_p=0.90,
temperature=0.1,
vllm_kwargs={
"gpu_memory_utilization": 0.5,
"enable_lora": True,
"max_model_len": 350,
},
)
LoRA_ADAPTER_PATH = "path/to/adapter"
lora_adapter = LoRARequest("lora_adapter", 1, LoRA_ADAPTER_PATH)
print(
llm.invoke("What are some popular Korean street foods?", lora_request=lora_adapter)
)
API 参考:VLLM