litellm - 💡(How to fix) Fix [Bug]: Azure gpt-4o rejects `max_tokens` — needs translation to `max_completion_tokens` like o-series/gpt-5 [1 comments, 2 participants]

What happened?

Azure OpenAI has deprecated the max_tokens parameter for gpt-4o models. Requests that include max_tokens now fail with:

AzureException BadRequestError - Unsupported parameter: 'max_tokens' is not supported with this model. Use 'max_completion_tokens' instead.

LiteLLM already handles this translation for o-series models (in o_series_transformation.py) and gpt-5 models (in gpt_5_transformation.py), where max_tokens is automatically mapped to max_completion_tokens:

if "max_tokens" in non_default_params:
    optional_params["max_completion_tokens"] = non_default_params.pop("max_tokens")

However, this translation is missing for the gpt-4o family in gpt_transformation.py. When a client (e.g. Open WebUI) sends max_tokens, LiteLLM passes it through unchanged to Azure, which rejects it.

Steps to Reproduce

1. Configure an Azure gpt-4o model in LiteLLM proxy:

model_list:
  - model_name: azure-se/gpt-4o
    litellm_params:
      model: azure/gpt-4o
      api_base: https://your-endpoint.openai.azure.com/
      api_key: os.environ/AZURE_API_KEY

1. Send a request with max_tokens set (e.g. from Open WebUI which sends this parameter by default):

curl -X POST http://localhost:4000/chat/completions \
  -H "Authorization: Bearer sk-..." \
  -H "Content-Type: application/json" \
  -d '{
    "model": "azure-se/gpt-4o",
    "messages": [{"role": "user", "content": "Hello"}],
    "max_tokens": 4096
  }'

1. Request fails with:

litellm.BadRequestError: AzureException BadRequestError - Unsupported parameter: 'max_tokens' is not supported with this model. Use 'max_completion_tokens' instead.

Expected Behavior

LiteLLM should automatically translate max_tokens → max_completion_tokens for Azure gpt-4o models, the same way it already does for o-series and gpt-5.

Suggested Fix

Add the same max_tokens → max_completion_tokens mapping to the gpt-4o code path (or to the Azure-specific transformation for models that no longer accept max_tokens).

Relevant log output

litellm.BadRequestError: AzureException BadRequestError - Unsupported parameter: 'max_tokens' is not supported with this model. Use 'max_completion_tokens' instead.
No fallback model group found for original model_group=azure-se/gpt-4o. Fallbacks=[].

What part of LiteLLM is this about?

Proxy, LLM Translation

What LiteLLM version are you on?

main-stable (latest as of 2026-03-30)

Twitter / LinkedIn details

No response