Correct way to access the logit_bias feature

[NullMagic2]

It seems the OpenAI API supports the attribute .logit_bias, which allows us to encourage or discorage / ban certain words: https://help.openai.com/en/articles/5247780-using-logit-bias-to-alter-token-probability-with-the-openai-api and the lmstudio API also supports it:
https://lmstudio.ai/docs/app/api/endpoints/openai

However, when I try to use it, it seems to have no effect.
For instance, the word "time" for Qwen3-14b is token no. 1678:

token = a.tokenize("time")
print(token)
[1678]

But if I try to pass logit_bias to Qwen 3, it ignores logit_bias completely:

self.model.act(
chat_object,
tool_list, # Pass the actual tools list
config={
"temp": 0.2,
"maxTokens": MAX_HISTORY_LENGTH,
"logit_bias": {"1678":-100},
},
on_prediction_fragment=self._stream_response,
)

[USER (10:08 AM)]: Complete this sentence: "Once upon a..."

[AI (10:08 AM)]: Once upon a time, there lived a curious little girl who loved exploring the enchanted forest behind her cottage.

Is this the right way to access logit_bias? Or is it a specific model issue?

[ncoghlan]

Unfortunately, this is a bug in the Python SDK (there are a handful of config fields currently missing from the config JSON Schema exported from the Typescript SDK, so the settings end up being ignored entirely at runtime rather than being passed to the server or raising an exception).

(However, a static typechecker will pick up that an unknown key is being passed in)