Autogen llm_config error

[doanaktar]

Hi, when my version is pyautogen==0.2.16 Autogen gives this error:

ValueError: Please either set llm_config to False, or specify a non-empty 'model' either in 'llm_config' or in each config of 'config_list'.

But when i switch to the pyautogen==0.2.2 version. It does not gives any error.

This is my llm_config:

{'config_list': 
        [{
        'base_url': 'http://0.0.0.0:8053',
        'api_key': 'NULL'
    }],
    'cache_seed': 25}

And litellm runs on that port without any issues.

[qingyun-wu]

There are some updates in the required format of config_list. Could you set llm_config to False, or specify a non-empty 'model' either in 'llm_config' or in each config of 'config_list' as suggested in the error message if you are using 0.2.16? Thank you!

[jackgerrits]

If I am interpreting this correctly it seems that litellm does not need a model arg to work? So perhaps requiring this is too heavy handed?

[sonichi]

@jackgerrits good point. I'm OK with removing the check in ConversableAgent. @gunnarku what do you think?

[ekzhu]

Looping in @olgavrou

[gunnarku]

The dilemma is between breaking the backward compatibility and more robust validation for preconditions (and I admit that I am biased towards more formality ;-)).

If we remove the checking for the presence of the model attribute, we'll regress to the state described in #1522 (the root cause for #1654).

If it's the only known case of overly aggressive validation, then modifying the configuration by the user is an obvious fix.

Issuing a warning message is a middle-ground solution.

However, @@jackgerrits and @sonichi, this is your show - do what's right for AutoGen customers ;-)

[sonichi]

I tried to find a solution to accommodate both requirements in #1946

[olgavrou]

Looping in @olgavrou

custom models also don't necessarily need the model to be set