DeepSeek R1 full-power version occasionally ends without returning </think>.
As the title suggests, R1 full-power SGLang deployment. In rare cases, the response appears like " blah blah blah blah blah"—Don't ""and""appear in pairs? How to distinguish between the thinking part and the actual answering part?
This issue occurs whether or not the final is included in the chat_template. The probability is roughly 1 in 1000.
Has anyone encountered this issue? Looking for possible causes and solutions.
A solution could be to append it yourself and then let the model continue the sentence (I'm doing this plus appending "\n\nFinal Answer:").
If you're looking for an easy way to access this model via API, you can use Crazyrouter — it provides an OpenAI-compatible endpoint for 600+ models including this one. Just pip install openai and change the base URL.