vLLM实战部署embedding、reranker、senseVoice、qwen2.5、qwen3模型

概述

一个开源的支持高并发的高性能大模型推理引擎。在这篇博客有简单提到过。

学习资料：官方文档，官方中文文档，中文文档。

modelscope

通过vLLM（或其他平台、框架）部署模型前，需要先下载模型。国内一般使用魔搭社区，模型库里搜索想要下载的模型。输入qwen3会看到很多个候选项，主要是：

• 参数量不一致；
• 量化版本不一致；
• 提交者不一致。

总之，找到模型下载路径（/Qwen/Qwen3-30B-A3B）是需要花点时间的。

下载模型有多种方式：

• 通过浏览器下载：打开魔搭页面，默认下载到C:\Users\johnny\Downloads目录下。模型文件通常非常大，如Qwen3-30B-A3B-FP8，分割为7个文件共30+G。Chrome下载效果如下，不便之处在于文件名可辨识度非常低，最后极有可能不知道下载下来的是哪个模型的子文件。而且这么做，也不太专业。
  
• 通过命令行下载：在有Python和pip的环境下，pip install modelscope安装modelscope。使用modelscope命令行下载，如下载reranker模型：modelscope download --model BAAI/bge-reranker-v2-m3 --local_dir /home/models/BAAI/bge-reranker-v2-m3。
• 通过客户端下载：vim ms_download.py，填入如下内容

from modelscope import snapshot_download
import os

os.environ['HTTP_PROXY'] = 'http://192.168.4.123:7890'
os.environ['HTTPS_PROXY'] = 'http://192.168.4.123:7890'

model_dir = snapshot_download('JunHowie/Qwen3-30B-A3B-GPTQ-Int8')

embedding

使用BAAI提供的模型，docker启动命令（一行）：

docker run --name=embedding --volume /home/models:/home/models --network=host --workdir=/ --restart=unless-stopped --runtime=nvidia --detach=true docker.1ms.run/dustynv/vllm:0.7.4-r36.4.0-cu128-24.04  python3 -m vllm.entrypoints.openai.api_server --served-model-name BAAI/bge-m3 --task embedding --enable-prefix-caching   --model /home/models/BAAI/bge-m3 --host 0.0.0.0 --port 8001 --trust-remote-code

docker logs -f <container_id>查看日志：

实际上，上述截图还省略掉不少有价值的日志：

• embedding模型只花费2.38s就成功加载权重值，只使用1G内存（显存），这也是被称为小模型的原因，下面的reranker和语音模型也是如此；
• 暴露的端点，即Endpoint，或routes，有很多，不一一列举；
• 端口是8001；本地访问使用http://0.0.0.0:8001/或http://localhost:8001，同一个局域网访问需要使用固定IP形式：http://192.168.4.134:8001/openapi.json；
• 192.168.4.123:63143是什么？

curl请求截图：


postman请求接口如下：

一行命令的可视化效果可能不太好，docker启动命令（多行）：

docker run -it \
--name=embedding \
--volume /home/models:/home/models \
--network=host \
--workdir=/ \
--restart=unless-stopped \
--runtime=nvidia \
--detach=true \
dustynv/vllm:0.7.4-r36.4.0-cu128-24.04 \
python3 -m vllm.entrypoints.openai.api_server \
--served-model-name BAAI/bge-m3 \
--task embedding \
--enable-prefix-caching \
--model /home/models/BAAI/bge-m3 \
--host 0.0.0.0 \
--port 8001 \
--trust-remote-code

测试

使用postman简单测试，POST请求，requestBody如下：

{
    "model":"BAAI/bge-m3",
    "input":"智能家居系统如何帮助提高家庭能源效率？"
}

执行效果如下：

使用更加专业的py脚本来执行测试：

import requests
import concurrent.futures

BASE_URL = "http://192.168.4.134:30011"
MODEL = "BAAI/bge-m3"

def create_request_body():
	return {
		"model": MODEL,
		"input": "智能家居系统如何帮助提高家庭能源效率？"
	}

def make_request(request_body):
	headers = {
		"Content-Type": "application/json"
	}
	response = requests.post(f"{BASE_URL}/v1/embeddings", json=request_body, headers=headers, verify=False)
	return response.json()

def parallel_requests(num_requests):
	request_body = create_request_body()
	with concurrent.futures.ThreadPoolExecutor(max_workers=num_requests) as executor:
		futures = [executor.submit(make_request, request_body) for _ in range(num_requests)]
		results = [future.result() for future in concurrent.futures.as_completed(futures)]
	return results

if __name__ == "__main__":
	num_requests = 50  # Example: Set the number of parallel requests
	responses = parallel_requests(num_requests)
	for i, response in enumerate(responses):
		print(f"Response {i+1}: {response}")

k8s

上面是使用docker方式启动模型，--restart=unless-stopped参数可以实现自动重启，但是重启（包括探针）机制不如k8s。

将上面的启动脚本扔给GPT，再经过调整，不难得到一个符合k8s或k3s的yaml文件：

# embedding-deployment.yaml
apiVersion: apps/v1
kind: Deployment
metadata:
  name: vllm-embedding
spec:
  replicas: 1
  selector:
    matchLabels:
      app: vllm-embedding
  template:
    metadata:
      labels:
        app: vllm-embedding
    spec:
      nodeSelector:
        kubernetes.io/hostname: k3s-master-2
      runtimeClassName: nvidia  # 需提前配置GPU运行时
      containers:
        - name: vllm-embedding
          image: dustynv/vllm:0.7.4-r36.4.0-cu128-24.04
          command: ["python3"]
          args:
            - "-m"
            - "vllm.entrypoints.openai.api_server"
            - "--served-model-name"
            - "BAAI/bge-m3"
            - "--task"
            - "embedding"  # 关键参数：指定任务类型
            - "--enable-prefix-caching"
            - "--model"
            - "/home/models/BAAI/bge-m3"
            - "--host"
            - "0.0.0.0"
            - "--port"
            - "8011"
            - "--trust-remote-code"
          volumeMounts:
            - name: model-storage
              mountPath: /home/models
          startupProbe:
            httpGet:
              path: /health
              port: 8011
            initialDelaySeconds: 120
            periodSeconds: 10
            failureThreshold: 3
          livenessProbe:
            httpGet:
              path: /health
              port: 8011
            initialDelaySeconds: 90
            periodSeconds: 10
            timeoutSeconds: 5
            failureThreshold: 3
          readinessProbe:
            httpGet:
              path: /health
              port: 8011
            initialDelaySeconds: 90
            periodSeconds: 5
            timeoutSeconds: 3
            failureThreshold: 1
      volumes:
        - name: model-storage
          hostPath:
            path: /home/models  # 主机模型存储路径
---
# embedding-service.yaml
apiVersion: v1
kind: Service
metadata:
  name: vllm-embedding-service
spec:
  selector:
    app: vllm-embedding
  ports:
    - protocol: TCP
      port: 8011      # 集群内部访问端口
      targetPort: 8011  # 容器端口
      nodePort: 30011
  type: NodePort      # 外部访问方式

reranker

与embedding经常一起出现的就是重排序，即reranker小模型。魔搭社区地址，一般用BAAI/bge-reranker-v2-m3

直接给出docker启动模型命令：

docker run -it --name=rerank --volume /home/models:/home/models --network=host --workdir=/ --restart=unless-stopped --runtime=nvidia --detach=true dustynv/vllm:0.7.4-r36.4.0-cu128-24.04  python3 -m vllm.entrypoints.openai.api_server --served-model-name BAAI/bge-reranker-v2-m3 --enable-prefix-caching --model /home/models/BAAI/bge-reranker-v2-m3 --host 0.0.0.0 --port 8002  --trust-remote-code

转化为k3s或k8s的yaml文件也很简单，将上面的embedding模型的yaml文件稍加改动即可。

语音

魔搭社区地址为iic/SenseVoiceSmall。启动SenseVoiceSmall模型的镜像不是vLLM，需要自己构建Docker镜像。如果转换为k3s yaml文件，则还需要将Docker镜像转化成crictl镜像。具体参考docker、ctr、crictl命令简介与使用。

qwen2.5

部署的模型为DeepSeek-R1-Distill-Qwen-32B-AWQ，魔搭社区地址为Valdemardi/DeepSeek-R1-Distill-Qwen-32B-AWQ。

Docker启动模型的命令为：

docker run -d --name=qwen-32b --hostname=ubuntu --workdir=/ --restart unless-stopped --runtime nvidia --detach=true -v /home/models:/models --network host dustynv/vllm:0.7.4-r36.4.0-cu128-24.04 python -m vllm.entrypoints.openai.api_server --served-model-name deepseek-r1-distill-qwen-32b-awq --model "/models/deepseek-r1-32b" --gpu-memory-utilization 0.6 --host 0.0.0.0 --port 8000 --tensor-parallel-size 1 --max-model-len 32768 --max-num-seqs 8

Docker命令转换成k3s的yaml文件不难：

# deepseek-deployment.yaml
apiVersion: apps/v1
kind: Deployment
metadata:
  name: deepseek-r1-worker
spec:
  replicas: 1
  selector:
    matchLabels:
      app: deepseek-r1-worker
  template:
    metadata:
      labels:
        app: deepseek-r1-worker
        # 负载均衡
        service: deepseek-r1-distill
    spec:
      nodeSelector:
        kubernetes.io/hostname: k3s-worker-3
      runtimeClassName: nvidia  # 依赖NVIDIA设备插件
      containers:
        - name: vllm
          image: dustynv/vllm:0.7.4-r36.4.0-cu128-24.04
          command: ["python"]
          args:
            - "-m"
            - "vllm.entrypoints.openai.api_server"
            - "--served-model-name"
            - "deepseek-r1-distill-qwen-32b-awq"
            - "--model"
            - "/home/models/deepseek-r1-32b"
            - "--gpu-memory-utilization"
            - "0.85" # 调度到32G从节点
            - "--max-model-len"
            - "8192"
            - "--host"
            - "0.0.0.0"
            - "--port"
            - "8010"
          volumeMounts:
            - name: model-storage
              mountPath: /home/models
          startupProbe:
            httpGet:
              path: /health
              port: 8010
            initialDelaySeconds: 200
            periodSeconds: 10
            failureThreshold: 3
          livenessProbe:
            httpGet:
              path: /health
              port: 8010
            initialDelaySeconds: 150
            periodSeconds: 10
            timeoutSeconds: 5
            failureThreshold: 3
          readinessProbe:
            httpGet:
              path: /health
              port: 8010
            initialDelaySeconds: 150
            periodSeconds: 10
            timeoutSeconds: 3
            failureThreshold: 1
      volumes:
        - name: model-storage
          hostPath:
            path: /home/models

在使用k3s，需要注意的是：主节点已经crictl pull下载vLLM镜像，通过modelscope download下载过模型文件，在主节点执行k apply -f qwen2.5-worker.yaml文件，希望通过k8s调度到从节点部署；则从节点，也需要对应的镜像文件和crictl镜像。此时通过scp传输应该比下载来得快。

qwen3

Qwen3-30B-A3B

上面部署embedding、reranker、qwen2模型使用的vLLM镜像版本都是dustynv/vllm:0.7.4-r36.4.0-cu128-24.04，继续使用此版本部署qwen3模型，比如没有经过量化的Qwen3-30B-A3B，启动日志里有个报错：

升级使用的vLLM版本到dustynv/vllm:0.8.6-r36.4-cu128-24.04，即可成功部署。启动日志如下：

Qwen3-30B-A3B-GPTQ-Int4

如上截图所示，一个64G显存一体的Jetson Orin GPU卡，部署好Qwen3-30B-A3B模型，已经用去56G显存。那这张卡基本上就不能再部署其他服务或模型。

因此需要考虑部署量化版的Qwen3，在modelscope下载多个不同版本的模型，以及反复试错，解决问题后，成功部署Qwen3-30B-A3B-GPTQ-Int4。

给出yaml文件如下：

# deepseek-deployment.yaml
apiVersion: apps/v1
kind: Deployment
metadata:
  name: deepseek-r1-master
spec:
  replicas: 1
  selector:
    matchLabels:
      app: deepseek-r1-master
  template:
    metadata:
      labels:
        app: deepseek-r1-master
        # 负载均衡配置
        service: deepseek-r1-distill
    spec:
      nodeSelector:
        kubernetes.io/hostname: k3s-master-2
      runtimeClassName: nvidia  # 依赖NVIDIA设备插件
      containers:
        - name: vllm
          image: dustynv/vllm:0.8.6-r36.4-cu128-24.04
          command: ["python"]
          args:
            - "-m"
            - "vllm.entrypoints.openai.api_server"
            - "--served-model-name"
            - "Qwen3-30B-A3B-GPTQ-Int4"
            - "--model"
            - "/home/models/Qwen/Qwen3-30B-A3B-GPTQ-Int4"
            - "--gpu-memory-utilization"
            - "0.6" # 调度到64G k3s主节点
			# - "0.7" # 调度到32G k3s从节点
            - "--max-model-len"
            - "8192"
            - "--enable-reasoning"
            - "--reasoning-parser"
            - "deepseek_r1"
            - "--host"
            - "0.0.0.0"
            - "--port"
            - "8010"
          volumeMounts:
            - name: model-storage
              mountPath: /home/models
            - name: local-py
              mountPath: /opt/venv/lib/python3.12/site-packages/vllm/model_executor/layers/quantization/gptq_marlin.py
          env:
            - name: VLLM_USE_V1
              value: "0"
          startupProbe:
            httpGet:
              path: /health
              port: 8010
            initialDelaySeconds: 500
            periodSeconds: 10
            failureThreshold: 3
          livenessProbe:
            httpGet:
              path: /health
              port: 8010
            initialDelaySeconds: 150
            periodSeconds: 10
            timeoutSeconds: 5
            failureThreshold: 3
          readinessProbe:
            httpGet:
              path: /health
              port: 8010
            initialDelaySeconds: 150
            periodSeconds: 10
            timeoutSeconds: 3
            failureThreshold: 1
      volumes:
        - name: model-storage
          hostPath:
            path: /home/models
        - name: local-py
          hostPath:
            path: /home/models/gptq_marlin.py
---
# deepseek-service.yaml
apiVersion: v1
kind: Service
metadata:
  name: deepseek-service
spec:
  selector:
    service: deepseek-r1-distill
  ports:
    - protocol: TCP
      port: 8010      # 集群内部访问端口
      targetPort: 8010  # 容器端口
      nodePort: 30010
  type: NodePort      # 外部访问方式或LoadBalancer

其中一个非常关键的配置是本地gptq_marlin.py文件挂载，替换vLLM默认的gptq_marlin.py文件。

其他

日志

vLLM部署Qwen3-30B-A3B-GPTQ-Int4模型的启动日志：

INFO 05-30 11:26:31 [config.py:730] This model supports multiple tasks: {'classify', 'embed', 'generate', 'score', 'reward'}. Defaulting to 'generate'.
INFO 05-30 11:26:33 [gptq_marlin.py:143] The model is convertible to gptq_marlin during runtime. Using gptq_marlin kernel.
INFO 05-30 11:26:33 [api_server.py:246] Started engine process with PID 79
/opt/venv/lib/python3.12/site-packages/transformers/utils/hub.py:105: FutureWarning: Using `TRANSFORMERS_CACHE` is deprecated and will be removed in v5 of Transformers. Use `HF_HOME` instead.
  warnings.warn(
INFO 05-30 11:26:38 [__init__.py:239] Automatically detected platform cuda.
INFO 05-30 11:26:42 [llm_engine.py:240] Initializing a V0 LLM engine (v0.8.6) with config: model='/home/models/Qwen/Qwen3-30B-A3B', speculative_config=None, tokenizer='/home/models/Qwen/Qwen3-30B-A3B', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, override_neuron_config=None, tokenizer_revision=None, trust_remote_code=False, dtype=torch.float16, max_seq_len=8192, download_dir=None, load_format=LoadFormat.AUTO, tensor_parallel_size=1, pipeline_parallel_size=1, disable_custom_all_reduce=False, quantization=gptq_marlin, enforce_eager=False, kv_cache_dtype=auto,  device_config=cuda, decoding_config=DecodingConfig(guided_decoding_backend='xgrammar', reasoning_backend='deepseek_r1'), observability_config=ObservabilityConfig(show_hidden_metrics=False, otlp_traces_endpoint=None, collect_model_forward_time=False, collect_model_execute_time=False), seed=None, served_model_name=Qwen3-30B-A3B, num_scheduler_steps=1, multi_step_stream_outputs=True, enable_prefix_caching=None, chunked_prefill_enabled=False, use_async_output_proc=True, pooler_config=None, compilation_config={"splitting_ops":[],"compile_sizes":[],"cudagraph_capture_sizes":[256,248,240,232,224,216,208,200,192,184,176,168,160,152,144,136,128,120,112,104,96,88,80,72,64,56,48,40,32,24,16,8,4,2,1],"max_capture_size":256}, use_cached_outputs=True, 
INFO 05-30 11:26:43 [cuda.py:292] Using Flash Attention backend.
[W530 11:26:44.180101328 ProcessGroupNCCL.cpp:959] Warning: TORCH_NCCL_AVOID_RECORD_STREAMS is the default now, this environment variable is thus deprecated. (function operator())
[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0
INFO 05-30 11:26:44 [parallel_state.py:1004] rank 0 in world size 1 is assigned as DP rank 0, PP rank 0, TP rank 0
INFO 05-30 11:26:44 [model_runner.py:1108] Starting to load model /home/models/Qwen/Qwen3-30B-A3B...
WARNING 05-30 11:26:44 [utils.py:168] The model class Qwen3MoeForCausalLM has not defined `packed_modules_mapping`, this may lead to incorrect mapping of quantized or ignored modules
INFO 05-30 11:26:44 [gptq_marlin.py:238] Using MarlinLinearKernel for GPTQMarlinLinearMethod
Loading safetensors checkpoint shards:   0% Completed | 0/1 [00:00<?, ?it/s]
Loading safetensors checkpoint shards: 100% Completed | 1/1 [00:28<00:00, 28.55s/it]
Loading safetensors checkpoint shards: 100% Completed | 1/1 [00:28<00:00, 28.55s/it]

INFO 05-30 11:27:15 [loader.py:458] Loading weights took 28.98 seconds
INFO 05-30 11:27:20 [model_runner.py:1140] Model loading took 15.6069 GiB and 36.310342 seconds
WARNING 05-30 11:27:21 [fused_moe.py:668] Using default MoE config. Performance might be sub-optimal! Config file not found at /opt/venv/lib/python3.12/site-packages/vllm/model_executor/layers/fused_moe/configs/E=128,N=4096,device_name=Orin_(nvgpu).json
INFO 05-30 11:27:25 [worker.py:287] Memory profiling takes 4.08 seconds
INFO 05-30 11:27:25 [worker.py:287] the current vLLM instance can use total_gpu_memory (61.37GiB) x gpu_memory_utilization (0.60) = 36.82GiB
INFO 05-30 11:27:25 [worker.py:287] model weights take 15.61GiB; non_torch_memory takes -4.03GiB; PyTorch activation peak memory takes 1.42GiB; the rest of the memory reserved for KV Cache is 23.83GiB.
INFO 05-30 11:27:25 [executor_base.py:112] # cuda blocks: 16267, # CPU blocks: 2730
INFO 05-30 11:27:25 [executor_base.py:117] Maximum concurrency for 8192 tokens per request: 31.77x
INFO 05-30 11:27:32 [model_runner.py:1450] Capturing cudagraphs for decoding. 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. If out-of-memory error occurs during cudagraph capture, consider decreasing `gpu_memory_utilization` or switching to eager mode. You can also reduce the `max_num_seqs` as needed to decrease memory usage.
Capturing CUDA graph shapes: 100%|██████████| 35/35 [00:38<00:00,  1.11s/it]
INFO 05-30 11:28:11 [model_runner.py:1592] Graph capturing finished in 39 secs, took 0.66 GiB
INFO 05-30 11:28:11 [llm_engine.py:437] init engine (profile, create kv cache, warmup model) took 50.42 seconds
WARNING 05-30 11:28:11 [config.py:1252] Default sampling parameters have been overridden by the model's Hugging Face generation config recommended from the model creator. If this is not intended, please relaunch vLLM instance with `--generation-config vllm`.
INFO 05-30 11:28:11 [serving_chat.py:118] Using default chat sampling params from model: {'temperature': 0.6, 'top_k': 20, 'top_p': 0.95}
INFO 05-30 11:28:11 [serving_completion.py:61] Using default completion sampling params from model: {'temperature': 0.6, 'top_k': 20, 'top_p': 0.95}

显存利用

其中有一段值得深扒：

INFO 05-30 11:27:25 [worker.py:287] the current vLLM instance can use total_gpu_memory (61.37GiB) x gpu_memory_utilization (0.60) = 36.82GiB
INFO 05-30 11:27:25 [worker.py:287] model weights take 15.61GiB; non_torch_memory takes -4.03GiB; PyTorch activation peak memory takes 1.42GiB; the rest of the memory reserved for KV Cache is 23.83GiB.

注意到non_torch_memory使用的内存是负数。

GPT解读：
在vLLM的资源分配逻辑中，内存使用被分为几个部分：

内存类型	含义
model weights	模型权重占用的显存（通常为FP16或INT8
non_torch_memory	非PyTorch分配器管理的显存，如CUDA内存池、自定义内存等
PyTorch activation peak memory	PyTorch激活函数使用的峰值内存（启用PyTorch backend）
KV Cache	用于缓存attention key/value的内存

这些加起来应该小于等于总的可用显存 × 利用率：
total_gpu_memory * gpu_memory_utilization = model_weights + non_torch_memory + pytorch_activation + kv_cache

出现负数的可能原因：

1. 数值溢出或精度误差：这些值可能是从GPU显存总量中减去其他已知项推导出来的；
2. 估算方式不够严谨，或某些值过大，可能导致结果为负数。

原因：

• 模型太大：模型本身权重就很大，如Qwen3-30B；
• 配置不合理：设置较高的max_num_seqs或max_seq_len，会导致KV Cache需求激增；加上一些激活内存开销，整体显存需求超过预留的60%，从而导致non_torch_memory被迫为负。

但是！！！

• 显存大小相同的两张Jetson Orin卡，都是64G；
• 都是Qwen3-30B-A3B-GPTQ-Int4模型；
• --gpu-memory-utilization参数都是0.6。

一张卡是正数。另一张卡出现上述负数，第一次部署模型时，内存异常，触发pod重启，此时free -h查看剩余显存不够300M；

pod重启日志里该参数还是负数，等待pod启动成功，free -h查看剩余显存有1~2G。

杀掉进程，重新部署模型，还是负数。

作为TODO有待后续跟进的2个问题：

• 为啥是负数？
• 这个负数的绝对值大小有什么学问，也就是为啥第一次会触发pod异常重启？