Deploy DeepSeek R1 without <think> output

[pescn]

Issue with Deploying DeepSeek R1 using SGlang

Problem Description

When deploying DeepSeek R1 with SGlang, an issue occurs where the token is not properly outputted. However, the initial response does contain thinking-related content, and the token is correctly generated.

Deployment Script

python3 -m sglang.launch_server \
  --model deepseek-ai/DeepSeek-R1 \
  --tp 8 \
  --trust-remote-code \
  --mem-fraction-static 0.95 \
  --port 30000 \
  --disable-cuda-graph \
  --served-model-name deepseek-reasoner \
  --max-running-requests 32 \
  --context-length 131072

Deployment Environment

Python: 3.12.9
GPU 0,1,2,3,4,5,6,7: NVIDIA H20
PyTorch: 2.5.1+cu124
sglang: 0.4.2.post3
flashinfer: 0.2.0.post2+cu124torch2.5

[pescn]

oh, thers is something i found!

Addition Infomation

python3 -m sglang.launch_server \
  --model deepseek-ai/DeepSeek-R1 \
  --tp 8 \
  --trust-remote-code \
  --mem-fraction-static 0.95 \
  --port 30000 \
  --disable-cuda-graph \
  --served-model-name deepseek-reasoner \
  --max-running-requests 32 \
  --context-length 131072 \
  --revision f7361cd9ff99396dbf6bd644ad846015e59ed4fc

It works good, can output token, but not everytime will output .

So maybe the issue can be fixed by output /n before generate other tokens?