I’m troubleshooting a locally hosted DeepSeek-derived instruction model. Sometimes it answers my question and then continues with a new user message that I never wrote. In other cases, role markers appear in the output.

Before blaming the model’s instruction following, I want to check whether my inference setup is formatting the conversation incorrectly. The exact checkpoint and its tokenizer files seem relevant here; I don’t want to assume every DeepSeek-derived model uses the same format.

What should I inspect first?

1. Whether the server uses the chat template shipped with the checkpoint.
2. Whether my client adds role markers that the server also inserts.
3. Whether the rendered prompt ends at the expected assistant-generation boundary.
4. Whether the runtime recognizes the checkpoint’s end-of-turn tokens.

Is that a sensible order for chat template debugging, or would you start elsewhere? I’m especially interested in how to distinguish an incorrectly formatted prompt from a stopping issue: both seem capable of producing an answer followed by an invented conversation.

I can share the checkpoint identifier, runtime version, and a sanitized rendered prompt. Which of those is most useful for diagnosing this without posting an entire server log?