Not sure if it's a bug or a configuration error...

[0wwafa]

With the same model (Mistral Instruct v03)
If I use llama.cpp (any version of the last 3 weeks) and the same prompt,
if I say "hello, how are you?" it answers "I am fine, and you?" (or similar)

If I do it using gradio and llama_cpp_python
I get
"I am an assistant.....blah blah.. I don't have feelings... blah blah"

I even took all the parameters from llama.cpp and set the same parameters in llama_cpp_python and the gradio interface.
I really don't understand what might be the difference.