[Bug]: deepseek reasoning_content is always None , because deepseek change the chat template

[mofanke]

Your current environment

https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B/commit/3865e12a1eb7cbd641ab3f9dfc28c588c6b0c1e9

vllm version V0.7.2

input : hello
output like

Alright, the user just said "hello." That's a friendly greeting. I should respond in a warm and welcoming manner. Maybe say something like, "Hello! How can I assist you today?" That way, I'm open to whatever they need help with.
</think>

Hello! How can I assist you today?

just has '</think>' no '<think>' , so reasoning_content is None

🐛 Describe the bug

vllm V0.7.2
model
https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B

Before submitting a new issue...

• Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.

[goufbest]

same just has '</think>' and no '<think>'. And I do not kown why.
do you think it is caused by [DeepSeek-R1-Distill-Qwen-32B] the model itself ？ have you tried another model like DeepSeek-R1-Distill-Qwen-7B？it works or the same?
btw ,before and after the chat_template changed，the model's output both do not contain '<think>'.
it is so confused.

[linpan]

see https://docs.vllm.ai/en/latest/features/reasoning_outputs.html

[linpan]

same just has '' and no ''. And I do not kown why. do you think it is caused by [DeepSeek-R1-Distill-Qwen-32B] the model itself ？ have you tried another model like DeepSeek-R1-Distill-Qwen-7B？it works or the same? btw ,before and after the chat_template changed，the model's output both do not contain ''. it is so confused.

see: https://docs.vllm.ai/en/latest/features/reasoning_outputs.html

[gaocegege]

I will file a PR to fix it today.

[Sqlver]

I have same bug on DeepSeek-R1-Distill-Qwen-14B,
mycli is
python3 -m vllm.entrypoints.openai.api_server --model /app/llm/deepseek-ai/DeepSeek-R1-Distill-Qwen-14B --served-model-name DeepSeek-R1-Distill-Qwen-14B --max-model-len=15392 --port 8001 --tensor-parallel-size 2 --enable-reasoning --reasoning-parser deepseek_r1

use https://docs.vllm.ai/en/latest/features/reasoning_outputs.html the example py

[gaocegege]

DeepSeek updated the chat template yesterday: link.

They added the <think> token to the chat template, indicating that the first section represents the thinking process and won't output <think> in the final output.

I had a quick fix #13025

[gaocegege]

The latest version should be fully compatible with the newly downloaded R1 series model.

[linpan]

I have same bug on DeepSeek-R1-Distill-Qwen-14B, mycli is python3 -m vllm.entrypoints.openai.api_server --model /app/llm/deepseek-ai/DeepSeek-R1-Distill-Qwen-14B --served-model-name DeepSeek-R1-Distill-Qwen-14B --max-model-len=15392 --port 8001 --tensor-parallel-size 2 --enable-reasoning --reasoning-parser deepseek_r1

use https://docs.vllm.ai/en/latest/features/reasoning_outputs.html the example py

没有问题。 自己处理。 docker vllm ,还不支持。