| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """An example shows how to generate chat completions from reasoning models |
| 4 | like DeepSeekR1. |
| 5 | |
| 6 | To run this example, you need to start the vLLM server |
| 7 | with the reasoning parser: |
| 8 | |
| 9 | ```bash |
| 10 | vllm serve deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B \ |
| 11 | --reasoning-parser deepseek_r1 |
| 12 | ``` |
| 13 | |
| 14 | This example demonstrates how to generate chat completions from reasoning models |
| 15 | using the OpenAI Python client library. |
| 16 | """ |
| 17 | |
| 18 | from openai import OpenAI |
| 19 | |
| 20 | # Modify OpenAI's API key and API base to use vLLM's API server. |
| 21 | openai_api_key = "EMPTY" |
| 22 | openai_api_base = "http://localhost:8000/v1" |
| 23 | |
| 24 | |
| 25 | def main(): |
| 26 | client = OpenAI( |
| 27 | api_key=openai_api_key, |
| 28 | base_url=openai_api_base, |
| 29 | ) |
| 30 | |
| 31 | models = client.models.list() |
| 32 | model = models.data[0].id |
| 33 | |
| 34 | # Round 1 |
| 35 | messages = [{"role": "user", "content": "9.11 and 9.8, which is greater?"}] |
| 36 | # ruff: noqa: E501 |
| 37 | # For granite, add: `extra_body={"chat_template_kwargs": {"thinking": True}}` |
| 38 | response = client.chat.completions.create(model=model, messages=messages) |
| 39 | |
| 40 | reasoning = response.choices[0].message.reasoning |
| 41 | content = response.choices[0].message.content |
| 42 | |
| 43 | print("reasoning for Round 1:", reasoning) |
| 44 | print("content for Round 1:", content) |
| 45 | |
| 46 | # Round 2 |
| 47 | messages.append({"role": "assistant", "content": content}) |
| 48 | messages.append( |
| 49 | { |
| 50 | "role": "user", |
| 51 | "content": "How many Rs are there in the word 'strawberry'?", |
| 52 | } |
| 53 | ) |
| 54 | response = client.chat.completions.create(model=model, messages=messages) |
| 55 | |
| 56 | reasoning = response.choices[0].message.reasoning |
| 57 | content = response.choices[0].message.content |
| 58 | |
| 59 | print("reasoning for Round 2:", reasoning) |
| 60 | print("content for Round 2:", content) |
| 61 | |
| 62 | |
| 63 | if __name__ == "__main__": |
| 64 | main() |