LLM models via OpenAI API#

This demo shows how to deploy LLM models in the OpenVINO Model Server using continuous batching and paged attention algorithms. Text generation use case is exposed via OpenAI API chat/completions and completions endpoints. That makes it easy to use and efficient especially on on Intel® Xeon® processors and ARC GPUs.

Note: This demo was tested on 4th - 6th generation Intel® Xeon® Scalable Processors, and Intel® Core Ultra Series on Ubuntu24 and Windows11.

Prerequisites#

Model Server deployment: Installed Docker Engine or OVMS binary package according to the baremetal deployment guide

(Optional) Client: Git and Python for using OpenAI client package and vLLM benchmark app

Server Deployment#

Container on Linux and CPU target device

Running this command starts the container with CPU only target device:

mkdir -p ${HOME}/models
docker run -it -p 8000:8000 --rm -e MOE_USE_MICRO_GEMM_PREFILL=0 --user $(id -u):$(id -g) -v ${HOME}/models:/models/:rw openvino/model_server:weekly --model_repository_path /models --source_model OpenVINO/Qwen3-30B-A3B-Instruct-2507-int4-ov --task text_generation --target_device CPU --tool_parser hermes3 --rest_port 8000 --model_name Qwen3-30B-A3B-Instruct-2507-int4-ov

Note: In case you want to use GPU target device, add extra docker parameters --device /dev/dri --group-add=$(stat -c "%g" /dev/dri/render* | head -n 1) to docker run command. The parameter --target_device should be also updated to GPU.

Binary package on Windows 11 with GPU target device

After ovms is installed according to steps from baremetal deployment guide, run the following command:

set MOE_USE_MICRO_GEMM_PREFILL=0
ovms.exe --model_repository_path c:\models --source_model OpenVINO/Qwen3-30B-A3B-Instruct-2507-int4-ov --task text_generation --target_device GPU --tool_parser hermes3 --rest_port 8000 --model_name Qwen3-30B-A3B-Instruct-2507-int4-ov

Readiness Check#

Wait for the model to load. You can check the status with a simple command:

curl http://localhost:8000/v3/models

{
  "object": "list",
  "data": [
    {
      "id": "Qwen3-30B-A3B-Instruct-2507-int4-ov",
      "object": "model",
      "created": 1772928358,
      "owned_by": "OVMS"
    }
  ]
}

Request Generation#

Model exposes both chat/completions and completions endpoints with and without stream capabilities. Chat endpoint is expected to be used for scenarios where conversation context should be pasted by the client and the model prompt is created by the server based on the jinja model template. Completion endpoint should be used to pass the prompt directly by the client and for models without the jinja template. Here is demonstrated model Qwen/Qwen3-30B-A3B-Instruct-2507 in int4 precision. It has chat capability so chat/completions endpoint will be employed:

Unary calls to chat/completions endpoint using cURL#

Linux

curl http://localhost:8000/v3/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "Qwen3-30B-A3B-Instruct-2507-int4-ov",
    "max_completion_tokens":500,
    "stream":false,
    "messages": [
      {
        "role": "system",
        "content": "You are a helpful assistant."
      },
      {
        "role": "user",
        "content": "If 1=3 2=3 3=5 4=4 5=4 Then, 6=?"
      }
    ]
  }'| jq .

Windows

Windows Powershell

(Invoke-WebRequest -Uri "http://localhost:8000/v3/chat/completions" `
 -Method POST `
 -Headers @{ "Content-Type" = "application/json" } `
 -Body '{"model": "Qwen3-30B-A3B-Instruct-2507-int4-ov", "max_completion_tokens": 500, "temperature": 0, "stream": false, "messages": [{"role": "system", "content": "You are a helpful assistant."}, {"role": "user", "content": "If 1=3 2=3 3=5 4=4 5=4 Then, 6=?"}]}').Content

Windows Command Prompt

curl -s http://localhost:8000/v3/chat/completions -H "Content-Type: application/json" -d "{\"model\": \"Qwen3-30B-A3B-Instruct-2507-int4-ov\", \"max_completion_tokens\": 500, \"temperature\": 0, \"stream\": false, \"messages\": [{\"role\": \"system\", \"content\": \"You are a helpful assistant.\"}, {\"role\": \"user\", \"content\": \"If 1=3 2=3 3=5 4=4 5=4 Then, 6=?\"}]}"

OpenAI Python package#

The endpoints chat/completions and completions are compatible with OpenAI client so it can be easily used to generate code also in streaming mode:

Install the client library:

pip3 install openai

Chat completions with streaming

from openai import OpenAI

client = OpenAI(
  base_url="http://localhost:8000/v3",
  api_key="unused"
)

stream = client.chat.completions.create(
    model="Qwen3-30B-A3B-Instruct-2507-int4-ov",
    messages=[{"role": "user", "content": "If 1=3 2=3 3=5 4=4 5=4 Then, 6=?"}],
    temperature=0,
    stream=True,
)
for chunk in stream:
    if chunk.choices[0].delta.content is not None:
        print(chunk.choices[0].delta.content, end="", flush=True)

Output:

We are given a pattern:

- 1 = 3  
- 2 = 3  
- 3 = 5  
- 4 = 4  
- 5 = 4  
- 6 = ?

We need to find the value for 6.

Let’s look at the pattern. The numbers on the left are integers, and the values on the right seem to represent something about the number itself.

Let’s consider: **the number of letters in the English word for the number.**

Check:

- **1** → "one" → 3 letters → matches 3 ✅  
- **2** → "two" → 3 letters → matches 3 ✅  
- **3** → "three" → 5 letters → matches 5 ✅  
- **4** → "four" → 4 letters → matches 4 ✅  
- **5** → "five" → 4 letters → matches 4 ✅  
- **6** → "six" → 3 letters → so **6 = 3**

### ✅ Answer: **3**

Chat completions with unary response

pip3 install openai

from openai import OpenAI

client = OpenAI(
  base_url="http://localhost:8000/v3",
  api_key="unused"
)

response = client.chat.completions.create(
    model="Qwen3-30B-A3B-Instruct-2507-int4-ov",
    messages=[{"role": "user", "content": "If 1=3 2=3 3=5 4=4 5=4 Then, 6=?"}],
    temperature=0,
    stream=False,
)
print(response.choices[0].message.content)

Output:

We are given a pattern:

- 1 = 3  
- 2 = 3  
- 3 = 5  
- 4 = 4  
- 5 = 4  
- 6 = ?

We need to find the value for 6.

Let’s look at the pattern. The numbers on the left are integers, and the values on the right seem to represent something about the **number of letters** when the number is written out in English.

Let’s check:

- **1** = "one" → 3 letters → matches 3 ✅  
- **2** = "two" → 3 letters → matches 3 ✅  
- **3** = "three" → 5 letters → matches 5 ✅  
- **4** = "four" → 4 letters → matches 4 ✅  
- **5** = "five" → 4 letters → matches 4 ✅  
- **6** = "six" → 3 letters → so 6 = **3**

### ✅ Answer: **3**

So, **6 = 3**.