AutoGen hangs on querying LMStudio, not able to kill or start other workflow

[stevenbaert]

Describe the issue

No response

Steps to reproduce

1. asked uncensored model running on LMStudio a standard censored (by other llms) question "how to break into a car?"
2. Output uncensored model (see that in LMStudio) was there but AutoGenStudio wanted to create a script for it and failed/hung.
3. There is no kill option or way to start other workflow once one is started

Screenshots and logs

No response

Additional Information

No response

[victordibia]

@sonichi .. is there a way to programatically interrupt an "in-progress" agent chat (ie a run of initiate_chat)?
Any pointers here will be helpful in crafting a solution that is lifted up into autogenstudio

[sonichi]

One plausible approach is to use the async method a_initiate_chat() and in a_check_termination_and_human_input() check an event's status which can be set outside.

[victordibia]

@sonichi, is there anyone else looking at this who can share insights? Maybe @ekzhu ?

It would be good if you can share more details on the async method you are referring to above or pointers to examples. Concretely, what is a code example where a chat can be initiated and then also terminated midway execution.

In the mean time, a developer experience that seems reasonable to me includes:

• each time an agent conversation is initiated initiate_chat, we get a process_id or some type of handle
• to end that conversation we provide some method like autogen.terminate_chate(process_id) which terminates that process
• we also provide some allow some type of termination_callback (that returns the current state of the agents) that developers can register to allow their apps degrade gracefully. E.g. my user makes a request to agents, in the middle the user terminates the request, this should fire my termination call back to ensure something sensible is returned to the user.

It's 100% possible I am thinking of this the wrong way, so happy to get thoughts.

[victordibia]

@jackgerrits , this is the issue I mentioned on the ability to terminate an agent run.