Unable to use "instruct" mode when using the API to send prompts to an LLM

[aaron13100]

Describe the bug

When using the API to interact with the text-generation-webui, I am unable to get the LLM to operate in "instruct" mode. The responses I receive are consistent with "chat" or "chat-instruct" mode, even when I explicitly try to set the mode to "instruct". This prevents me from using the LLM in a way that directly follows instructions without engaging in conversation.

Is there an existing issue for this?

• I have searched the existing issues

Reproduction

1. Set up text-generation-webui with the Meta-Llama-3.1-8B-Instruct-Q8_0.gguf model.
2. Use the API to send the following sequence of prompts:
   a. "translate everything I say to french."
   b. "I like cake."
3. Observe that the response to the second prompt is conversational rather than a direct French translation.

Expected behavior: In "instruct" mode, the response to "I like cake." should be "J'aime le gateau."
Actual behavior: The LLM responds with conversational text about liking cake, indicating it's in "chat" or "chat-instruct" mode.

Screenshot

No response

Logs

no errors

System Info

MacOS 14.6.1 (23G93)
Apple M3 Max