Reasoning Parser#
SGLang supports parsing reasoning content out from “normal” content for reasoning models such as DeepSeek R1.
Supported Models & Parsers#
Model |
Reasoning tags |
Parser |
Notes |
|---|---|---|---|
|
|
Supports all variants (R1, R1-0528, R1-Distill) |
|
|
|
Including
DeepSeek‑V3.2.
Supports |
|
|
|
Supports |
|
|
|
Always generates thinking content |
|
|
|
Uses special thinking delimiters. Also requires
|
|
|
|
N/A |
Model-Specific Behaviors#
DeepSeek-R1 Family:
DeepSeek-R1: No
<think>start tag, jumps directly to thinking contentDeepSeek-R1-0528: Generates both
<think>start and</think>end tagsBoth are handled by the same
deepseek-r1parser
DeepSeek-V3 Family:
DeepSeek-V3.1/V3.2: Hybrid model supporting both thinking and non-thinking modes, use the
deepseek-v3parser andthinkingparameter (NOTE: notenable_thinking)
Qwen3 Family:
Standard Qwen3 (e.g., Qwen3-2507): Use
qwen3parser, supportsenable_thinkingin chat templatesQwen3-Thinking (e.g., Qwen3-235B-A22B-Thinking-2507): Use
qwen3orqwen3-thinkingparser, always thinks
Kimi K2:
Kimi K2 Thinking: Uses special
◁think▷and◁/think▷tags. For agentic tool use, also specify--tool-call-parser kimi_k2.
GPT OSS:
GPT OSS: Uses special
<|channel|>analysis<|message|>and<|end|>tags
Usage#
Launching the Server#
Specify the --reasoning-parser option.
[1]:
import requests
from openai import OpenAI
from sglang.test.doc_patch import launch_server_cmd
from sglang.utils import wait_for_server, print_highlight, terminate_process
server_process, port = launch_server_cmd(
"python3 -m sglang.launch_server --model-path deepseek-ai/DeepSeek-R1-Distill-Qwen-7B --host 0.0.0.0 --reasoning-parser deepseek-r1 --log-level warning"
)
wait_for_server(f"http://localhost:{port}", process=server_process)
/actions-runner/_work/sglang/sglang/python/sglang/launch_server.py:54: UserWarning: 'python -m sglang.launch_server' is still supported, but 'sglang serve' is the recommended entrypoint.
Example: sglang serve --model-path <model> [options]
warnings.warn(
Multi-thread loading shards: 100% Completed | 2/2 [00:02<00:00, 1.42s/it]
Compiling num tokens (num_tokens=4): 100%|██████████| 58/58 [00:07<00:00, 7.56it/s]
Capturing num tokens (num_tokens=4 avail_mem=36.94 GB): 100%|██████████| 58/58 [00:03<00:00, 14.60it/s]
/usr/local/lib/python3.10/dist-packages/fastapi/routing.py:120: FastAPIDeprecationWarning: ORJSONResponse is deprecated, FastAPI now serializes data directly to JSON bytes via Pydantic when a return type or response model is set, which is faster and doesn't need a custom response class. Read more in the FastAPI docs: https://fastapi.tiangolo.com/advanced/custom-response/#orjson-or-response-model and https://fastapi.tiangolo.com/tutorial/response-model/
response = await f(request)
NOTE: Typically, the server runs in a separate terminal.
In this notebook, we run the server and notebook code together, so their outputs are combined.
To improve clarity, the server logs are displayed in the original black color, while the notebook outputs are highlighted in blue.
To reduce the log length, we set the log level to warning for the server, the default log level is info.
We are running those notebooks in a CI environment, so the throughput is not representative of the actual performance.
Note that --reasoning-parser defines the parser used to interpret responses.
OpenAI Compatible API#
Using the OpenAI compatible API, the contract follows the DeepSeek API design established with the release of DeepSeek-R1:
reasoning_content: The content of the CoT.content: The content of the final answer.
[2]:
# Initialize OpenAI-like client
client = OpenAI(api_key="None", base_url=f"http://0.0.0.0:{port}/v1")
model_name = client.models.list().data[0].id
messages = [
{
"role": "user",
"content": "What is 1+3?",
}
]
Non-Streaming Request#
[3]:
response_non_stream = client.chat.completions.create(
model=model_name,
messages=messages,
temperature=0.6,
top_p=0.95,
stream=False, # Non-streaming
extra_body={"separate_reasoning": True},
)
print_highlight("==== Reasoning ====")
print_highlight(response_non_stream.choices[0].message.reasoning_content)
print_highlight("==== Text ====")
print_highlight(response_non_stream.choices[0].message.content)
Next, I add the two numbers together: 1 plus 3 equals 4.
Therefore, the final answer is 4.
**Solution:**
We are asked to find the sum of 1 and 3.
\[
1 + 3 = 4
\]
Therefore, the final answer is \(\boxed{4}\).
Streaming Request#
[4]:
response_stream = client.chat.completions.create(
model=model_name,
messages=messages,
temperature=0.6,
top_p=0.95,
stream=True, # Non-streaming
extra_body={"separate_reasoning": True},
)
reasoning_content = ""
content = ""
for chunk in response_stream:
if chunk.choices[0].delta.content:
content += chunk.choices[0].delta.content
if chunk.choices[0].delta.reasoning_content:
reasoning_content += chunk.choices[0].delta.reasoning_content
print_highlight("==== Reasoning ====")
print_highlight(reasoning_content)
print_highlight("==== Text ====")
print_highlight(content)
Next, I add the two numbers together: 1 plus 3 equals 4.
Therefore, the final answer is 4.
**Solution:**
We need to find the sum of 1 and 3.
\[
1 + 3 = 4
\]
Therefore, the final answer is \(\boxed{4}\).
Optionally, you can buffer the reasoning content to the last reasoning chunk (or the first chunk after the reasoning content).
[5]:
response_stream = client.chat.completions.create(
model=model_name,
messages=messages,
temperature=0.6,
top_p=0.95,
stream=True, # Non-streaming
extra_body={"separate_reasoning": True, "stream_reasoning": False},
)
reasoning_content = ""
content = ""
for chunk in response_stream:
if chunk.choices[0].delta.content:
content += chunk.choices[0].delta.content
if chunk.choices[0].delta.reasoning_content:
reasoning_content += chunk.choices[0].delta.reasoning_content
print_highlight("==== Reasoning ====")
print_highlight(reasoning_content)
print_highlight("==== Text ====")
print_highlight(content)
I will add the two numbers together.
The result of adding 1 and 3 is 4.
**Solution:**
We need to find the sum of 1 and 3.
\[
1 + 3 = 4
\]
**Answer:** \boxed{4}
The reasoning separation is enable by default when specify . To disable it, set the ``separate_reasoning`` option to ``False`` in request.
[6]:
response_non_stream = client.chat.completions.create(
model=model_name,
messages=messages,
temperature=0.6,
top_p=0.95,
stream=False, # Non-streaming
extra_body={"separate_reasoning": False},
)
print_highlight("==== Original Output ====")
print_highlight(response_non_stream.choices[0].message.content)
Next, I perform the addition of these two numbers to find the sum.
Finally, I conclude that the result of 1 plus 3 is 4.
To solve the addition of 1 and 3, follow these simple steps:
1. **Identify the numbers to add:**
\[
1 \quad \text{and} \quad 3
\]
2. **Perform the addition:**
\[
1 + 3 = 4
\]
3. **Present the final answer:**
\[
\boxed{4}
\]
SGLang Native API#
[7]:
from transformers import AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained("deepseek-ai/DeepSeek-R1-Distill-Qwen-7B")
input = tokenizer.apply_chat_template(
messages, tokenize=False, add_generation_prompt=True, return_dict=False
)
gen_url = f"http://localhost:{port}/generate"
gen_data = {
"text": input,
"sampling_params": {
"skip_special_tokens": False,
"max_new_tokens": 1024,
"temperature": 0.6,
"top_p": 0.95,
},
}
gen_response = requests.post(gen_url, json=gen_data).json()["text"]
print_highlight("==== Original Output ====")
print_highlight(gen_response)
parse_url = f"http://localhost:{port}/separate_reasoning"
separate_reasoning_data = {
"text": gen_response,
"reasoning_parser": "deepseek-r1",
}
separate_reasoning_response_json = requests.post(
parse_url, json=separate_reasoning_data
).json()
print_highlight("==== Reasoning ====")
print_highlight(separate_reasoning_response_json["reasoning_text"])
print_highlight("==== Text ====")
print_highlight(separate_reasoning_response_json["text"])
I start by identifying the two numbers involved in the addition.
Next, I perform the addition operation by combining these numbers.
Finally, I calculate the total to arrive at the answer.
Sure! Let's solve the problem step by step.
**Question:** What is \(1 + 3\)?
**Solution:**
1. **Identify the numbers to add:**
\[
1 \quad \text{and} \quad 3
\]
2. **Perform the addition:**
\[
1 + 3 = 4
\]
**Answer:** \(\boxed{4}\)
/usr/local/lib/python3.10/dist-packages/fastapi/routing.py:328: FastAPIDeprecationWarning: ORJSONResponse is deprecated, FastAPI now serializes data directly to JSON bytes via Pydantic when a return type or response model is set, which is faster and doesn't need a custom response class. Read more in the FastAPI docs: https://fastapi.tiangolo.com/advanced/custom-response/#orjson-or-response-model and https://fastapi.tiangolo.com/tutorial/response-model/
return await dependant.call(**values)
I start by identifying the two numbers involved in the addition.
Next, I perform the addition operation by combining these numbers.
Finally, I calculate the total to arrive at the answer.
Sure! Let's solve the problem step by step.
**Question:** What is \(1 + 3\)?
**Solution:**
1. **Identify the numbers to add:**
\[
1 \quad \text{and} \quad 3
\]
2. **Perform the addition:**
\[
1 + 3 = 4
\]
**Answer:** \(\boxed{4}\)
[8]:
terminate_process(server_process)
Offline Engine API#
[9]:
import sglang as sgl
from sglang.srt.parser.reasoning_parser import ReasoningParser
from sglang.utils import print_highlight
llm = sgl.Engine(model_path="deepseek-ai/DeepSeek-R1-Distill-Qwen-7B")
tokenizer = AutoTokenizer.from_pretrained("deepseek-ai/DeepSeek-R1-Distill-Qwen-7B")
input = tokenizer.apply_chat_template(
messages, tokenize=False, add_generation_prompt=True, return_dict=False
)
sampling_params = {
"max_new_tokens": 1024,
"skip_special_tokens": False,
"temperature": 0.6,
"top_p": 0.95,
}
result = llm.generate(prompt=input, sampling_params=sampling_params)
generated_text = result["text"] # Assume there is only one prompt
print_highlight("==== Original Output ====")
print_highlight(generated_text)
parser = ReasoningParser("deepseek-r1")
reasoning_text, text = parser.parse_non_stream(generated_text)
print_highlight("==== Reasoning ====")
print_highlight(reasoning_text)
print_highlight("==== Text ====")
print_highlight(text)
Multi-thread loading shards: 100% Completed | 2/2 [00:02<00:00, 1.32s/it]
Compiling num tokens (num_tokens=4): 100%|██████████| 58/58 [00:07<00:00, 7.53it/s]
Capturing num tokens (num_tokens=4 avail_mem=45.73 GB): 100%|██████████| 58/58 [00:03<00:00, 14.67it/s]
First, I recognize that 1 is a single unit, and 3 represents three units.
By combining these two quantities, I get a total of 4 units.
Therefore, the sum of 1 and 3 is 4.
Sure! Let's solve the addition step by step.
**Problem:**
What is \(1 + 3\)?
**Solution:**
1. Start with the first number:
\(1\)
2. Add the second number:
\(1 + 3\)
3. Combine the quantities:
\(1 + 3 = 4\)
**Answer:**
\(\boxed{4}\)
First, I recognize that 1 is a single unit, and 3 represents three units.
By combining these two quantities, I get a total of 4 units.
Therefore, the sum of 1 and 3 is 4.
Sure! Let's solve the addition step by step.
**Problem:**
What is \(1 + 3\)?
**Solution:**
1. Start with the first number:
\(1\)
2. Add the second number:
\(1 + 3\)
3. Combine the quantities:
\(1 + 3 = 4\)
**Answer:**
\(\boxed{4}\)
[10]:
llm.shutdown()
Supporting New Reasoning Model Schemas#
For future reasoning models, you can implement the reasoning parser as a subclass of BaseReasoningFormatDetector in python/sglang/srt/reasoning_parser.py and specify the reasoning parser for new reasoning model schemas accordingly.