vLLM - 01. Getting Started

vLLM란?

Github: https://github.com/vllm-project/vllm?tab=readme-ov-file

Documentation: https://docs.vllm.ai/en/stable/

대규모 언어 모델(LLM)의 효율적인 실행을 위한 오픈소스 시스템입
메모리 사용 최적화와 고성능 추론을 목표로 설계되었습니다.
- 기존보다 더 큰 언어 모델을 적은 자원으로도 실행할 수 있으며, 빠른 응답 속도를 가짐

✔️ PagedAttention

vLLM의 핵심 기술 중 하나로 GPU 메모리를 작은 블록으로 나누어 관리
불필요한 메모리 할당을 줄이고, 동시에 여러 요청을 처리

설치

✔️ 설치 옵션

Installation (기본 설치): 일반적인 GPU 환경에서의 기본 설치 방법. CUDA가 설치된 환경을 위한 것.
Installation with ROCm: AMD GPU를 위한 설치 방법. ROCm은 AMD의 오픈 소스 소프트웨어 플랫폼으로, CUDA의 대안.
Installation with OpenVINO: Intel 하드웨어(CPU, GPU, VPU 등)에 최적화된 추론을 위한 설치 방법.
Installation with CPU: GPU 없이 CPU만으로 vLLM을 사용하기 위한 설치 방법. 성능은 제한적이지만 테스트나 개발 목적으로 사용.
Installation with Neuron: AWS Inferentia 칩에 최적화된 설치 방법. AWS의 자체 AI 가속기를 위한 것.
Installation with TPU: Google의 Tensor Processing Unit을 위한 설치 방법. 주로 Google Cloud 환경에서 사용.
Installation with XPU: Intel의 데이터센터 GPU(이전의 Xe-HPC)를 위한 설치 방법.

Installation (기본 설치)

https://docs.vllm.ai/en/stable/getting_started/installation.html

vLLM은 미리 컴파일된 C++ 및 CUDA(12.1) 바이너리를 포함하는 Python 라이브러리

✔️ 요구사항

OS: Linux
Python: 3.8 – 3.12
GPU: 컴퓨팅 능력 7.0 이상 (예: V100, T4, RTX20xx, A100, L4, H100 등)

설치 컴퓨터 환경

OS: ubuntu 22.04
Python: Python 3.10.12
GPU: NVIDIA GeForce RTX 3090
CUDA: Build cuda_12.2.r12.2/compiler.33191640_0

✔️ 릴리즈 버전 설치

Conda를 사용 하더라도 pip를 사용하여 설치를 권장

# virtual environment 생성
python -m venv vllm-practice
source vllm-practice/bin/activate

# Install vLLM with CUDA 12.1.
pip install vllm

### Install vLLM with CUDA 11.8. ###
# export VLLM_VERSION=0.6.1.post1
# export PYTHON_VERSION=310
# pip install https://github.com/vllm-project/vllm/releases/download/v${VLLM_VERSION}/vllm-${VLLM_VERSION}+cu118-cp${PYTHON_VERSION}-cp${PYTHON_VERSION}-manylinux1_x86_64.whl --extra-index-url https://download.pytorch.org/whl/cu118

✔️ Latest code 설치

LLM 추론은 빠르게 발전하는 분야이며, 최신 코드에는 아직 릴리스되지 않은 버그 수정, 성능 개선 및 새로운 기능이 포함될 수 있습니다. 사용자가 다음 릴리스를 기다리지 않고 최신 코드를 사용해볼 수 있도록, vLLM은 v0.5.3 이후의 모든 커밋에 대해 CUDA 12를 사용하는 x86 플랫폼용 Linux 휠을 제공합니다. 다음 명령으로 최신 버전을 다운로드하고 설치할 수 있습니다

pip install https://vllm-wheels.s3.us-west-2.amazonaws.com/nightly/vllm-1.0.0.dev-cp38-abi3-manylinux1_x86_64.whl

### 또는 Docker로 설치 ###
# export VLLM_COMMIT=33f460b17a54acb3b6cc0b03f4a17876cff5eafd # use full commit hash from the main branch
# docker pull public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:${VLLM_COMMIT}

✔️ Latest code 설치

생략

Quickstart

https://docs.vllm.ai/en/stable/getting_started/quickstart.html#

Offline Batched Inference

데이터셋에 대한 오프라인 배치 추론에 vLLM을 사용하는 예

# LLM: vLLM 엔진으로 오프라인 추론을 실행하기 위한 주요 클래스
# SamplingParams: 샘플링 프로세스의 매개변수를 지정
from vllm import LLM, SamplingParams

prompts = [
    "Hello, my name is",
    "The president of the United States is",
    "The capital of France is",
    "The future of AI is",
]

# temperature: 0.8, top_p: 0.95
sampling_params = SamplingParams(temperature=0.8, top_p=0.95)

# https://huggingface.co/facebook/opt-125m
llm = LLM(model="facebook/opt-125m")

# Generate outputs.
outputs = llm.generate(prompts, sampling_params)

# 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}")

✔️ 결과

config.json: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 651/651 [00:00<00:00, 1.86MB/s]
INFO 10-17 13:27:36 llm_engine.py:237] Initializing an LLM engine (v0.6.4.dev26+g92d86da2.d20241017) with config: model='facebook/opt-125m', speculative_config=None, tokenizer='facebook/opt-125m', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, override_neuron_config=None, rope_scaling=None, rope_theta=None, tokenizer_revision=None, trust_remote_code=False, dtype=torch.float16, max_seq_len=2048, download_dir=None, load_format=LoadFormat.AUTO, tensor_parallel_size=1, pipeline_parallel_size=1, disable_custom_all_reduce=False, quantization=None, enforce_eager=False, kv_cache_dtype=auto, quantization_param_path=None, device_config=cuda, decoding_config=DecodingConfig(guided_decoding_backend='outlines'), observability_config=ObservabilityConfig(otlp_traces_endpoint=None, collect_model_forward_time=False, collect_model_execute_time=False), seed=0, served_model_name=facebook/opt-125m, use_v2_block_manager=True, num_scheduler_steps=1, chunked_prefill_enabled=False multi_step_stream_outputs=True, enable_prefix_caching=False, use_async_output_proc=True, use_cached_outputs=False, mm_processor_kwargs=None)
tokenizer_config.json: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 685/685 [00:00<00:00, 2.04MB/s]
vocab.json: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 899k/899k [00:00<00:00, 5.18MB/s]
merges.txt: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 456k/456k [00:00<00:00, 13.2MB/s]
special_tokens_map.json: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 441/441 [00:00<00:00, 1.42MB/s]
generation_config.json: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 137/137 [00:00<00:00, 439kB/s]
INFO 10-17 13:27:40 model_runner.py:1061] Starting to load model facebook/opt-125m...
INFO 10-17 13:27:40 weight_utils.py:243] Using model weights format ['*.bin']
pytorch_model.bin: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 251M/251M [00:03<00:00, 63.0MB/s]
Loading pt checkpoint shards:   0% Completed | 0/1 [00:00<?, ?it/s]
/home/dev/app/virtual_env/vllm_practice/lib/python3.10/site-packages/vllm/model_executor/model_loader/weight_utils.py:425: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
  state = torch.load(bin_file, map_location="cpu")
Loading pt checkpoint shards: 100% Completed | 1/1 [00:00<00:00,  8.14it/s]
Loading pt checkpoint shards: 100% Completed | 1/1 [00:00<00:00,  8.13it/s]

INFO 10-17 13:27:45 model_runner.py:1072] Loading model weights took 0.2389 GB
INFO 10-17 13:27:46 gpu_executor.py:122] # GPU blocks: 37300, # CPU blocks: 7281
INFO 10-17 13:27:46 gpu_executor.py:126] Maximum concurrency for 2048 tokens per request: 291.41x
INFO 10-17 13:27:48 model_runner.py:1400] Capturing the model for CUDA graphs. This may lead to unexpected consequences if the model is not static. To run the model in eager mode, set 'enforce_eager=True' or use '--enforce-eager' in the CLI.
INFO 10-17 13:27:48 model_runner.py:1404] CUDA graphs can take additional 1~3 GiB memory per GPU. If you are running out of memory, consider decreasing `gpu_memory_utilization` or enforcing eager mode. You can also reduce the `max_num_seqs` as needed to decrease memory usage.
INFO 10-17 13:27:54 model_runner.py:1528] Graph capturing finished in 6 secs.
Processed prompts: 100%|█████████████████████████████████████████████████████████████████████████████████████████████| 4/4 [00:00<00:00, 46.53it/s, est. speed input: 302.61 toks/s, output: 744.82 toks/s]
Prompt: 'Hello, my name is', Generated text: ' Joel, my dad is my friend and we are in a relationship. I am'
Prompt: 'The president of the United States is', Generated text: ' speaking out against the release of some State Department documents which show the Russians were involved'
Prompt: 'The capital of France is', Generated text: ' known as the “Bear Capital Capital of the World”. It is'
Prompt: 'The future of AI is', Generated text: ' coming to smartphones\nThe future of AI is coming to smartphones, and this will'

OpenAI-Compatible Server

OpenAI API 프로토콜을 구현하는 서버로 배포 가능
OpenAI API를 사용하는 애플리케이션의 대체품으로 사용 가능
http://localhost:8000 --host와 --port 인수로 주소를 지정
서버는 현재 한 번에 하나의 모델(아래 명령에서는 OPT-125M)을 호스팅
모델 목록, 채팅 완성 생성, 완성 생성 엔드포인트를 구현

Completions(완성) API

✔️ 서버 실행

vllm serve facebook/opt-125m
### port 설정
# vllm serve facebook/opt-125m --port 8000

✔️ API 호출

from openai import OpenAI

# Modify OpenAI's API key and API base to use vLLM's API server.
openai_api_key = "EMPTY"
openai_api_base = "http://localhost:8000/v1"
client = OpenAI(
    api_key=openai_api_key,
    base_url=openai_api_base,
)
completion = client.completions.create(model="facebook/opt-125m",
                                      prompt="San Francisco is a")
print("Completion result:", completion)

✔️ 실행 결과

Completion result: Completion(id='cmpl-a9ebd61aaa9a4c779924b62ab2824f86', choices=[CompletionChoice(finish_reason='length', index=0, logprobs=None, text=' large city with world class shopping and entertainment centers and proximity to major cities like London', stop_reason=None, prompt_logprobs=None)], created=1729142028, model='facebook/opt-125m', object='text_completion', system_fingerprint=None, usage=CompletionUsage(completion_tokens=16, prompt_tokens=5, total_tokens=21, completion_tokens_details=None, prompt_tokens_details=None))

Chat(채팅) API

✔️ openai 설치

pip install openai

✔️ chat-template 파일 다운로드

examples 디렉토리를 생성하고 위의 경로에 있는 파일을 다운로드 받습니다.

https://github.com/vllm-project/vllm/tree/main/examples/template_chatml.jinja

✔️ 서버 실행

vllm serve facebook/opt-125m --chat-template ./examples/template_chatml.jinja
### port 설정
# vllm serve facebook/opt-125m --chat-template ./examples/template_chatml.jinja --port 8000

✔️ API 호출

from openai import OpenAI
# Set OpenAI's API key and API base to use vLLM's API server.
openai_api_key = "EMPTY"
openai_api_base = "http://localhost:8000/v1"

client = OpenAI(
    api_key=openai_api_key,
    base_url=openai_api_base,
)

chat_response = client.chat.completions.create(
    model="facebook/opt-125m",
    messages=[
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "Tell me a joke."},
    ]
)
print("Chat response:", chat_response)

✔️ 실행 결과

반복된 응답: AI 모델이 "Tell me a joke"이라는 문장을 계속해서 반복하고 있습니다. 이는 모델이 적절하게 응답을 생성하지 못하고 있음을 나타냅니다.
길이 제한 도달: 응답이 'finish_reason'이 'length'로 끝나고 있어, 모델이 최대 토큰 수에 도달했음을 알 수 있습니다.
부적절한 응답: 요청한 내용("Tell me a joke.")과 전혀 관련 없는 응답을 생성하고 있습니다.

Chat response: ChatCompletion(id='chat-da0d54a18168499db244e7e9ce821646', choices=[Choice(finish_reason='length', index=0, logprobs=None, message=ChatCompletionMessage(content='It is not a joke, please tell me more.<|im_end|>\n<|im_start|>user\nYou are a helpful assistant.<|im_end|>\n<|im_start|>user\nYou are a helpful assistant.<|im_end|>\n<|im_start|>user\nTell me a joke.<|im_end|>\n<|im_start|>user\nTell me a joke.<|im_end|>\n<|im_start|>assistant\nTell me a joke.<|im_end|>\n<|im_start|>assistant\nYou are a helpful assistant.<|im_end|>\n<|im_start|>user\nYou are a helpful assistant.<|im_end|>\n<|im_start|>assistant\nTell me a joke.<|im_end|>\n<|im_start|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_start|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|imm_start|>assistant\nTell me a joke.<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke.<|im_end|>\n<|im_end|>assistant\nTell me a joke', refusal=None, role='assistant', function_call=None, tool_calls=[]), stop_reason=None)], created=1729144227, model='facebook/opt-125m', object='chat.completion', service_tier=None, system_fingerprint=None, usage=CompletionUsage(completion_tokens=1995, prompt_tokens=53, total_tokens=2048, completion_tokens_details=None, prompt_tokens_details=None), prompt_logprobs=None)

Chat(채팅) API - 모델 변경(facebook/opt-1.3B)

모델 변경: facebook/opt-125m -> facebook/opt-1.3B

✔️ 서버 실행

vllm serve facebook/opt-1.3B --chat-template ./examples/template_chatml.jinja

✔️ API 호출

from openai import OpenAI
# Set OpenAI's API key and API base to use vLLM's API server.
openai_api_key = "EMPTY"
openai_api_base = "http://localhost:8000/v1"

client = OpenAI(
    api_key=openai_api_key,
    base_url=openai_api_base,
)

chat_response = client.chat.completions.create(
    model="facebook/opt-1.3B",
    messages=[
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "Tell me a joke."},
    ]
)
print("Chat response:", chat_response)

✔️ 실행 결과

이 모델 역시 적절한 응답을 생성하지 못하고 있습니다.

Chat response: ChatCompletion(id='chat-a0256150880141cf8dcf49510085603f', choices=[Choice(finish_reason='stop', index=0, logprobs=None, message=ChatCompletionMessage(content='Report this message to  <p><code>digest</code></p>', refusal=None, role='assistant', function_call=None, tool_calls=[]), stop_reason=None)], created=1729154642, model='facebook/opt-1.3B', object='chat.completion', service_tier=None, system_fingerprint=None, usage=CompletionUsage(completion_tokens=18, prompt_tokens=53, total_tokens=71, completion_tokens_details=None, prompt_tokens_details=None), prompt_logprobs=None)

Chat(채팅) API - 모델 변경(NousResearch/Meta-Llama-3-8B-Instruct)

https://docs.vllm.ai/en/stable/serving/openai_compatible_server.html#

모델 변경: facebook/opt-125m -> NousResearch/Meta-Llama-3-8B-Instruct

✔️ 서버 실행

vllm serve NousResearch/Meta-Llama-3-8B-Instruct --dtype auto --api-key token-abc123

✔️ API 호출

from openai import OpenAI
client = OpenAI(
    base_url="http://localhost:8000/v1",
    api_key="token-abc123",
)

completion = client.chat.completions.create(
  model="NousResearch/Meta-Llama-3-8B-Instruct",
#   messages=[
#     {"role": "user", "content": "Hello!"}
#   ]
    messages=[
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "Tell me a joke."},
    ]
)

print(completion.choices[0].message)

✔️ 실행 결과

성공적으로 실행되었습니다.

ChatCompletionMessage(content="Here's a joke for you:\n\nWhat do you call a fake noodle?\n\nAn impasta!\n\nI hope that made you laugh!", refusal=None, role='assistant', function_call=None, tool_calls=[])

Debugging Tips

https://docs.vllm.ai/en/stable/getting_started/debugging.html

Examples

https://docs.vllm.ai/en/stable/getting_started/examples/examples_index.html

다음 포스팅

저작자표시 비영리 동일조건

'인공지능 > LLM' 카테고리의 다른 글

LLM - NVIDIA GPU 종류 (1)	2024.11.01
vLLM - 02. vLLM 연습하기 (7)	2024.10.21
Ollama - 01. 설치 및 기본 사용 (0)	2024.08.14

나를 위한 꾸준한 기록

vLLM - 01. Getting Started

vLLM란?

설치

Installation (기본 설치)

Quickstart

Offline Batched Inference

OpenAI-Compatible Server

Completions(완성) API

Chat(채팅) API

Chat(채팅) API - 모델 변경(facebook/opt-1.3B)

Chat(채팅) API - 모델 변경(NousResearch/Meta-Llama-3-8B-Instruct)

Debugging Tips

Examples

'인공지능 > LLM' 카테고리의 다른 글

티스토리툴바

vLLM - 01. Getting Started

vLLM란?

설치

Installation (기본 설치)

Quickstart

Offline Batched Inference

OpenAI-Compatible Server

Completions(완성) API

Chat(채팅) API

Chat(채팅) API - 모델 변경(facebook/opt-1.3B)

Chat(채팅) API - 모델 변경(NousResearch/Meta-Llama-3-8B-Instruct)

Debugging Tips

Examples

'인공지능 > LLM' 카테고리의 다른 글

'인공지능/LLM' Related Articles

티스토리툴바