Bug: Empty response in interactive mode

[batmanonline]

when running with the -t 12 -i -r "### Human:" flags llama returns control
the cpu activity goes to 0 and the user sends a new input
however, llama now continues responding to the previous input (or returns no response) completely ignoring the new input...

from here llama completely breaks the chat, it even generates the prompt "### Human:" and starts completing the questions written by a human

Note: tested this many times with vicuna and gpt4all
Note: this might be related to issues #990 #941